diff --git a/.github/gpu_ci/madmatrix_checks.py b/.github/gpu_ci/madmatrix_checks.py
index 797d9f78fc..fc11077bb4 100644
--- a/.github/gpu_ci/madmatrix_checks.py
+++ b/.github/gpu_ci/madmatrix_checks.py
@@ -26,6 +26,13 @@
    synchronise with the default stream) gives the same |M|^2 and helicities as on the
    default stream: every memset, kernel and copy of the matrix element must go to the
    stream umami is given, the helicity filtering of the first call included.
+4. helicity_index: ``g g > t t~ z`` (massless and massive vectors, a longitudinal Z,
+   massive fermions) evaluated for one given helicity per event (UMAMI_IN_HELICITY_INDEX,
+   --hel-index), on both backends: the |M|^2 of the ncomb helicities, each given to every
+   event, add up to the helicity-summed |M|^2, the GPU ones equal the CPU ones helicity by
+   helicity, and with a random helicity per event (the threads of a warp, or the lanes of
+   a SIMD vector, on different helicities in lock-step) every event gets the |M|^2 of its
+   helicity and returns it as the chosen helicity.
 
 Any CPU backend but scalar (simd_128, ...) works too, to try the script without a GPU;
 check 3 is then skipped. Writes summary.txt (key=value lines) in the current directory;
@@ -46,7 +53,7 @@
 
 GPU_BACKENDS = ('cuda', 'hip')
 CPU_BACKEND = 'scalar'
-CHECKS = ('mssm_gpu_vs_cpu', 'zero_me', 'nonblocking_stream')
+CHECKS = ('mssm_gpu_vs_cpu', 'zero_me', 'nonblocking_stream', 'helicity_index')
 
 results = {}
 
@@ -66,7 +73,7 @@ def run(cmd, cwd=None, log=None):
 
 
 def write(path, typecode, values):
-    """A raw binary file of check_sa.exe: 'd' double, 'I' 32-bit unsigned int."""
+    """A raw binary file of check_sa.exe: 'd' double, 'I'/'i' 32-bit unsigned/signed int."""
     with open(path, 'wb') as f:
         array.array(typecode, values).tofile(f)
     return os.path.abspath(path)
@@ -267,6 +274,70 @@ def check_nonblocking_stream(backend, repo, ncalls=4):
           '%d calls of %d events identical to the default stream' % (ncalls, nevt // ncalls))
 
 
+# ---------------------------------------------------------------------------
+# 4. one given helicity per event
+# ---------------------------------------------------------------------------
+
+def check_helicity_index(backend, repo, nevt=512):
+    section('g g > t t~ z, one given helicity per event (%s, %s)' % (backend, CPU_BACKEND))
+    name = 'ggttxz_hel'
+    (proc_dir,) = output(repo, name, 'sm', 'g g > t t~ z', (CPU_BACKEND, backend))
+    data = open(os.path.join(proc_dir, 'ProcessData.h')).read()
+    ncomb = int(re.search(r'constexpr int ncomb = (\d+);', data).group(1))
+    batch = [1, nevt, 1]
+    momenta = dump(name, 'momenta')
+    check_sa(proc_dir, CPU_BACKEND, ['--energy', 2000, '--dump-momenta', momenta] + batch,
+             name + '_rambo.log')
+    rng = random.Random(17)
+    mixed = [rng.randrange(ncomb) for _ in range(nevt)]
+    problems, details = [], []
+    perhel = {}
+    for run_backend in (CPU_BACKEND, backend):
+        prefix = '%s_%s' % (name, run_backend)
+        check_sa(proc_dir, run_backend, ['--momenta', momenta, '--dump-me', dump(prefix, 'me')] + batch,
+                 prefix + '.log')
+        summed = read(dump(prefix, 'me'), 'd')
+        perhel[run_backend] = []
+        for ihel in range(ncomb):
+            hprefix = '%s_h%d' % (prefix, ihel)
+            hels = write(hprefix + '_in.bin', 'i', [ihel] * nevt)
+            check_sa(proc_dir, run_backend, ['--momenta', momenta, '--hel-index', hels,
+                                             '--dump-me', dump(hprefix, 'me'),
+                                             '--dump-hel', dump(hprefix, 'hel')] + batch,
+                     hprefix + '.log')
+            perhel[run_backend].append(read(dump(hprefix, 'me'), 'd'))
+            if set(read(dump(hprefix, 'hel'), 'i')) != {ihel}:
+                problems.append('%s: helicity %d not returned as the chosen helicity' % (run_backend, ihel))
+        total = [sum(perhel[run_backend][ihel][ievt] for ihel in range(ncomb)) for ievt in range(nevt)]
+        rels = [abs(a - b) / abs(b) if b != 0 else float('inf') for a, b in zip(total, summed)]
+        bad = [ievt for ievt, rel in enumerate(rels) if not rel < 1e-10]
+        if bad or len(summed) != nevt:
+            problems.append('%s: the sum over the helicities differs from the helicity sum in %d of %d events'
+                            % (run_backend, len(bad), nevt))
+        # a random helicity per event, the same events
+        mprefix = '%s_mixed' % prefix
+        hels = write(mprefix + '_in.bin', 'i', mixed)
+        check_sa(proc_dir, run_backend, ['--momenta', momenta, '--hel-index', hels,
+                                         '--dump-me', dump(mprefix, 'me'),
+                                         '--dump-hel', dump(mprefix, 'hel')] + batch,
+                 mprefix + '.log')
+        me, hel = read(dump(mprefix, 'me'), 'd'), read(dump(mprefix, 'hel'), 'i')
+        expected = [perhel[run_backend][ihel][ievt] for ievt, ihel in enumerate(mixed)]
+        bad = [ievt for ievt, (a, b) in enumerate(zip(me, expected)) if not abs(a - b) <= 1e-12 * abs(b)]
+        if bad or list(hel) != mixed:
+            problems.append('%s: a random helicity per event: %d |M|^2 differ from that helicity\'s, '
+                            '%d helicities not returned' % (run_backend, len(bad),
+                                                            sum(1 for a, b in zip(hel, mixed) if a != b)))
+        details.append('%s: max rel diff of the helicity sum %.1e' % (run_backend, max(rels, default=0.)))
+        print('  ' + details[-1])
+    bad = sum(1 for ihel in range(ncomb) for a, b in zip(perhel[backend][ihel], perhel[CPU_BACKEND][ihel])
+              if not abs(a - b) <= 1e-9 * max(abs(b), 1e-300))
+    if bad:
+        problems.append('%d (helicity, event) |M|^2 differ between %s and %s' % (bad, backend, CPU_BACKEND))
+    check('helicity_index', not problems, '; '.join(problems[:3]) or
+          '%d helicities x %d events; %s' % (ncomb, nevt, '; '.join(details)))
+
+
 def main():
     parser = argparse.ArgumentParser(description=__doc__.split('\n')[0])
     parser.add_argument('--repo', required=True)
diff --git a/.github/workflows/acceptancetest_mg7.yml b/.github/workflows/acceptancetest_mg7.yml
index 2cc8e69569..baa3fdb1a1 100644
--- a/.github/workflows/acceptancetest_mg7.yml
+++ b/.github/workflows/acceptancetest_mg7.yml
@@ -564,6 +564,39 @@ jobs:
             ./tests/test_manager.py test_systematics_mg7 -pA -t0 -l INFO
 
 
+  acceptancetest_mg7_madtrex:
+    needs: build_madspace
+    # MadtRex (madtrex/, 'install madtrex'): built on top of the cached madspace
+    # install, then its tests, which all skip without an installed madtrex (and
+    # a skipped test fails the step, so a broken install cannot pass quietly):
+    # test_madtrex_standalone (the standalone driver/bin/reweight, the card
+    # readers, in-place weights) and test_reweight_backends (classic module and
+    # MadtRex reweight the same events and agree event by event; needs f2py).
+    runs-on: ubuntu-24.04
+    if: github.event_name == 'push' || github.event.pull_request.head.repo.fork == true
+    steps:
+      - uses: actions/checkout@v4
+      - uses: ./.github/actions/checkout_mg5
+      - uses: ./.github/actions/install_madspace
+      - uses: ./.github/actions/restore_heptools
+      - uses: ./.github/actions/restore-pip-cache
+      - name: install madtrex
+        run: |
+            cd $GITHUB_WORKSPACE
+            python3 madtrex/install.py --yes -j 4
+      - name: test_madtrex_standalone
+        if: ${{ !cancelled() }}
+        run: |
+            cd $GITHUB_WORKSPACE
+            ./tests/test_manager.py test_madtrex_standalone -t0
+      - name: test_reweight_backends
+        if: ${{ !cancelled() }}
+        run: |
+            cd $GITHUB_WORKSPACE
+            export PATH="$HOME/.cache/HEPtools/bin:$PATH"
+            env -u LHAPDF_DATA_PATH ./tests/test_manager.py test_reweight_backends -pA -t0 -l INFO
+
+
   acceptancetest_mg7_uncovered:
     # mg7 acceptance tests that no workflow selected (audit of 2026-10-24).
     # test_mg7_custom_functions and test_mg7_interference compare to runs with
diff --git a/.github/workflows/unittest.yml b/.github/workflows/unittest.yml
index 6797dffac5..9fd1bb3584 100644
--- a/.github/workflows/unittest.yml
+++ b/.github/workflows/unittest.yml
@@ -324,6 +324,17 @@ jobs:
             cd $GITHUB_WORKSPACE
             ./tests/test_manager.py test_lhe_ordering test_reweight_ordering test_boost_by_a_momentum_at_rest test_nested_identical_children -t0
 
+      # ---- unittest_reweight_backend ----
+      # the reweighting backend choice (classic module or MadtRex) and the
+      # MadtRex reweight-card readers: pure Python, no madspace/madtrex. The
+      # tests needing an installed madtrex run in acceptancetest_mg7.yml
+      # (acceptancetest_mg7_madtrex).
+      - name: "unittest_reweight_backend: reweighting backend choice and MadtRex cards"
+        if: ${{ !cancelled() }}
+        run: |
+            cd $GITHUB_WORKSPACE
+            ./tests/test_manager.py test_reweight_backend NpyToLheNamedWeightsTest HelicityReweightingCardTest ClassicVocabularyTest GridpackSourceBuildTest -t0
+
       # Leave /tmp empty for the next job on this self-hosted machine.
       - if: always()
         uses: ./.github/actions/clean_tmp
diff --git a/MadSpin/interface_madspin.py b/MadSpin/interface_madspin.py
index 8cc58875c8..008627c9eb 100755
--- a/MadSpin/interface_madspin.py
+++ b/MadSpin/interface_madspin.py
@@ -12064,7 +12064,7 @@ def _density_libtag(self, tree, prod_or_decay):
         that build's tag."""
         tag = madspin.density_libtag(self._ms_run_id,
                                      'p' if prod_or_decay == 'prod' else 'd')
-        if os.path.exists(pjoin(tree, 'lib', 'libmadmatrix_common%s_scalar.so' % tag)):
+        if os.path.exists(madmatrix_density.common_library_path(tree, tag)):
             return tag
         try:
             return madmatrix_density.find_libtag(tree)
diff --git a/UpdateNotes.txt b/UpdateNotes.txt
index 016fe9c68d..d3a63020cc 100644
--- a/UpdateNotes.txt
+++ b/UpdateNotes.txt
@@ -36,6 +36,42 @@ ANNOUNCEMENT:
        NNPDF40_nlo_as_01180_qed (lhaid 335900), at NLO also for a model with a
        massive b (NNPDF4.0 has no 4-flavour QED set). The NNPDF4.0 defaults
        have no photon. A PDF set by the user is kept.
+   ZW: MG7 outputs have two reweighting backends: MadtRex ('install madtrex';
+       matrix-element and PDF/scale reweighting on lhe or lhe_npy events) and the
+       classic reweighting module (f2py). Both read Cards/reweight_card.dat with the
+       classic vocabulary ('change' for options, 'set' for parameters). The backend
+       comes from 'generate_events --reweight-backend=classic|madtrex|auto', the
+       reweight switch of the launch question (now OFF/ON/classic/madtrex/density,
+       ON being auto), or the card's new 'change backend' line; auto prefers
+       MadtRex when it can run the card. The choice is logged and recorded in
+       Events/<run>/reweighting.json, and a choice that cannot run stops the launch
+       before the generation. Reweighting runs once, after the systematics;
+       '--reweight=CARD' now just switches it on with CARD as the reweight card.
+       MadtRex cards: 'change helicity', 'change fptype', 'change mode LO|pdf',
+       'change verbose/max_batch/...' replace the former 'set' spellings (still
+       accepted, with a warning), 'set MT 175' (a parameter by its param_card
+       name) works, and directives only the classic module knows are an error
+       instead of being ignored. MG7 outputs get their own default reweight card
+       documenting both backends. madevent and aMC@NLO outputs are unchanged.
+       Both backends add their weights to the run's own event file
+       (Events/<run>/events.lhe.gz, or events.npy for an lhe_npy run reweighted by
+       MadtRex, as rwgt_<name> columns that npy_to_lhe carries over), after the
+       systematics weights; MadtRex no longer writes a separate *_reweighted file
+       (bin/reweight --output still can). MadtRex reads lhe_npy events with
+       systematics weights (they were refused as "Invalid header for event file").
+   ZW: Without f2py_compiler set, the f2py installed with the running Python (e.g.
+       in a virtual environment) is used before f2py3/f2py on the PATH, which may
+       build modules for another Python version (reweighting then failed with "file
+       too short" or a missing all_matrix2py module).
+   OM: MadtRex follow-ups: 'install madtrex' builds with clang/macOS; madtrex refuses a
+       madspace other than the one it was built against (with the command to rebuild it;
+       MadGraph then reweights with the classic module) and its installer a too old
+       madspace; a failing reweighting no longer stops the run, and 'reweight RUN' (after
+       'launch PROC -i') reweights the events of an earlier run; MadtRex reweight cards
+       accept '!' comments and multi-index parameters ('set NMIX 1 2 0.5'); the baseline
+       PDF set and precision of a MadtRex reweighting are those of the run that made the
+       events (its LHE header), not of the current run_card; a gridpack made with
+       include_madtrex_source builds madtrex from its sources in bin/reweight.
    OM: new command 'install cuttools_iregi' recompiling CutTools and IREGI (vendor/)
        with the current fortran compiler. The error raised when CutTools/IREGI were
        compiled with another compiler (e.g. after a gfortran update) now points to it.
diff --git a/aloha/template_files/madmatrix/helas.h b/aloha/template_files/madmatrix/helas.h
index a243ce0d96..f33d986f1d 100644
--- a/aloha/template_files/madmatrix/helas.h
+++ b/aloha/template_files/madmatrix/helas.h
@@ -26,7 +26,7 @@
   __host__ __device__ INLINE void
   ixxxxx( const fptype_momenta momenta[], // input: momenta
           const fptype_amp fmass,     // input: fermion mass
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavour
           ALOHAOBJ & fi,          // output: aloha objects
@@ -41,7 +41,7 @@
   __host__ __device__ INLINE void
   ipzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavor index
           ALOHAOBJ & fi,          // output: wavefunctions
@@ -56,7 +56,7 @@
   __host__ __device__ INLINE void
   imzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavor index
           ALOHAOBJ & fi,          // output: wavefunctions
@@ -71,7 +71,7 @@
   __host__ __device__ INLINE void
   ixzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavor index
           ALOHAOBJ & fi,          // output: wavefunctions
@@ -85,7 +85,7 @@
   __host__ __device__ INLINE void
   vxxxxx( const fptype_momenta momenta[], // input: momenta
           const fptype_amp vmass,     // input: vector boson mass
-          const int nhel,         // input: -1, 0 (only if vmass!=0) or +1 (helicity of vector boson)
+          const fptype_amp_sv nhel, // input: -1, 0 (only if vmass!=0) or +1 (helicity of vector boson, per SIMD lane/GPU thread)
           const int nsv,          // input: +1 (final) or -1 (initial)
           const int flv,          // input: flavor index
           ALOHAOBJ & vc,          // output: wavefunctions
@@ -113,7 +113,7 @@
   __host__ __device__ INLINE void
   oxxxxx( const fptype_momenta momenta[], // input: momenta
           const fptype_amp fmass,     // input: fermion mass
-          const int nhel,         // input: -1, 0 (only if vmass!=0) or +1 (helicity of vector boson)
+          const fptype_amp_sv nhel, // input: -1, 0 (only if vmass!=0) or +1 (helicity of vector boson, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavor index
           ALOHAOBJ & fo,          // output: wavefunctions
@@ -128,7 +128,7 @@
   __host__ __device__ INLINE void
   opzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavor index
           ALOHAOBJ & fo,          // output: wavefunctions
@@ -143,7 +143,7 @@
   __host__ __device__ INLINE void
   omzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavor index
           ALOHAOBJ & fo,          // output: wavefunctions
@@ -157,7 +157,7 @@
   __host__ __device__ INLINE void
   oxzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavor index
           ALOHAOBJ & fo,          // output: wavefunctions
@@ -171,7 +171,7 @@
   __host__ __device__ void
   ixxxxx( const fptype_momenta momenta[], // input: momenta
           const fptype_amp fmass,     // input: fermion mass
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavour
           ALOHAOBJ & fi,          // output: wavefunctions
@@ -196,7 +196,12 @@
     fi.pvec[2] = -pvec2_ * static_cast<fptype_momenta>(nsf);
     fi.pvec[3] = -pvec3_ * static_cast<fptype_momenta>(nsf);
     fi.flv_index = flv;
-    const int nh = nhel * nsf;
+    const fptype_amp_sv nh = nhel * static_cast<fptype_amp>( nsf );
+    // BRANCHLESS HELICITY: the helicity is not a branch condition but enters as the truth values
+    // tp (nh == +1) and tm (nh == -1), 1 or 0, which multiply the two alternatives. Different
+    // SIMD lanes or GPU threads can then carry different helicities and still run in lock-step.
+    const fptype_amp_sv tp = ( 1 + nh ) * (fptype_amp)0.5;
+    const fptype_amp_sv tm = 1 - tp;
     if( fmass != 0. )
     {
 #ifndef MGONGPU_CPPSIMD
@@ -205,11 +210,9 @@
       volatile fptype_amp_sv p2 = pvec1 * pvec1 + pvec2 * pvec2 + pvec3 * pvec3; // volatile fixes #736
       const fptype_amp_sv pp = fpmin( pvec0, fpsqrt( p2 ) );
 #endif
-      // In C++ ixxxxx, use a single ip/im numbering that is valid both for pp==0 and pp>0, which have two numbering schemes in Fortran ixxxxx:
-      // for pp==0, Fortran sqm(0:1) has indexes 0,1 as in C++; but for Fortran pp>0, omega(2) has indexes 1,2 and not 0,1
-      // NB: this is only possible in ixxxx, but in oxxxxx two different numbering schemes must be used
-      const int ip = ( 1 + nh ) / 2; // NB: same as in Fortran pp==0, differs from Fortran pp>0, which is (3+nh)/2 because omega(2) has indexes 1,2
-      const int im = ( 1 - nh ) / 2; // NB: same as in Fortran pp==0, differs from Fortran pp>0, which is (3-nh)/2 because omega(2) has indexes 1,2
+      // The Fortran ip/im indices ( 1 + nh ) / 2 and ( 1 - nh ) / 2 (for pp > 0, ( 3 +- nh ) / 2 as omega(2) has
+      // indexes 1,2) are the truth values tp and tm: an array element x[ip] is tm * x[0] + tp * x[1], and
+      // ip * x[ip] is just tp * x[1] (and likewise for im)
 #ifndef MGONGPU_CPPSIMD
       if( pp == 0. )
       {
@@ -217,10 +220,10 @@
         fptype_amp sqm[2] = { fpsqrt( std::abs( fmass ) ), 0. }; // possibility of negative fermion masses
         //sqm[1] = ( fmass < 0. ? -abs( sqm[0] ) : abs( sqm[0] ) ); // AV: why abs here?
         sqm[1] = ( fmass < 0. ? -sqm[0] : sqm[0] ); // AV: removed an abs here
-        w[0] = cxmake( ip * sqm[ip], 0 );
-        w[1] = cxmake( im * nsf * sqm[ip], 0 );
-        w[2] = cxmake( ip * nsf * sqm[im], 0 );
-        w[3] = cxmake( im * sqm[im], 0 );
+        w[0] = cxmake( tp * sqm[1], 0 );
+        w[1] = cxmake( tm * nsf * sqm[0], 0 );
+        w[2] = cxmake( tp * nsf * sqm[0], 0 );
+        w[3] = cxmake( tm * sqm[1], 0 );
       }
       else
       {
@@ -228,30 +231,32 @@
                                fptype_amp( 1 + nsf - ( 1 - nsf ) * nh ) * (fptype_amp)0.5 };
         fptype_amp omega[2] = { fpsqrt( pvec0 + pp ), 0. };
         omega[1] = fmass / omega[0];
-        const fptype_amp sfomega[2] = { sf[0] * omega[ip], sf[1] * omega[im] };
+        const fptype_amp sfomega[2] = { sf[0] * ( tm * omega[0] + tp * omega[1] ), sf[1] * ( tp * omega[0] + tm * omega[1] ) };
         const fptype_amp pp3 = fpmax( fpternary( fpsignbit( pvec3 ), ( pvec1 * pvec1 + pvec2 * pvec2 ) / ( pp - pvec3 ), pp + pvec3 ), 0. );
         const cxtype_amp chi[2] = { cxmake( fpsqrt( pp3 * (fptype_amp)0.5 / pp ), 0. ),
                                 ( pp3 == 0. ? cxmake( -nh, 0. ) : cxmake( nh * pvec1, pvec2 ) / fpsqrt( (fptype_amp)2. * pp * pp3 ) ) };
-        w[0] = sfomega[0] * chi[im];
-        w[1] = sfomega[0] * chi[ip];
-        w[2] = sfomega[1] * chi[im];
-        w[3] = sfomega[1] * chi[ip];
+        const cxtype_amp chi_im = tp * chi[0] + tm * chi[1]; // chi[im]
+        const cxtype_amp chi_ip = tm * chi[0] + tp * chi[1]; // chi[ip]
+        w[0] = sfomega[0] * chi_im;
+        w[1] = sfomega[0] * chi_ip;
+        w[2] = sfomega[1] * chi_im;
+        w[3] = sfomega[1] * chi_ip;
       }
 #else
       // Branch A: pp == 0.
       // NB: Do not use "abs" for floats! It returns an integer with no build warning! Use std::abs!
       fptype_amp sqm[2] = { fpsqrt( std::abs( fmass ) ), 0 }; // possibility of negative fermion masses (NB: SCALAR!)
       sqm[1] = ( fmass < 0 ? -sqm[0] : sqm[0] );          // AV: removed an abs here (as above)
-      const cxtype_amp fiA_2 = ip * sqm[ip];                  // scalar cxtype: real part initialised from fptype, imag part = 0
-      const cxtype_amp fiA_3 = im * nsf * sqm[ip];            // scalar cxtype: real part initialised from fptype, imag part = 0
-      const cxtype_amp fiA_4 = ip * nsf * sqm[im];            // scalar cxtype: real part initialised from fptype, imag part = 0
-      const cxtype_amp fiA_5 = im * sqm[im];                  // scalar cxtype: real part initialised from fptype, imag part = 0
+      const cxtype_amp_v fiA_2 = cxmake( tp * sqm[1], 0 );
+      const cxtype_amp_v fiA_3 = cxmake( tm * (fptype_amp)nsf * sqm[0], 0 );
+      const cxtype_amp_v fiA_4 = cxmake( tp * (fptype_amp)nsf * sqm[0], 0 );
+      const cxtype_amp_v fiA_5 = cxmake( tm * sqm[1], 0 );
       // Branch B: pp != 0.
-      const fptype_amp sf[2] = { fptype_amp( 1 + nsf + ( 1 - nsf ) * nh ) * (fptype_amp)0.5,
-                             fptype_amp( 1 + nsf - ( 1 - nsf ) * nh ) * (fptype_amp)0.5 };
+      const fptype_amp_v sf[2] = { ( (fptype_amp)( 1 + nsf ) + (fptype_amp)( 1 - nsf ) * nh ) * (fptype_amp)0.5,
+                                   ( (fptype_amp)( 1 + nsf ) - (fptype_amp)( 1 - nsf ) * nh ) * (fptype_amp)0.5 };
       fptype_v omega[2] = { fpsqrt( pvec0 + pp ), 0 };
       omega[1] = fmass / omega[0];
-      const fptype_amp_v sfomega[2] = { sf[0] * omega[ip], sf[1] * omega[im] };
+      const fptype_amp_v sfomega[2] = { sf[0] * ( tm * omega[0] + tp * omega[1] ), sf[1] * ( tp * omega[0] + tm * omega[1] ) };
       const fptype_amp_v pp3 = fpmax( fpternary( fpsignbit( pvec3 ), ( pvec1 * pvec1 + pvec2 * pvec2 ) / ( pp - pvec3 ), pp + pvec3 ), 0 );
       volatile fptype_amp_v ppDENOM = fpternary( pp != 0, pp, 1. );    // hack: ppDENOM[ieppV]=1 if pp[ieppV]==0
       volatile fptype_amp_v pp3DENOM = fpternary( pp3 != 0, pp3, 1. ); // hack: pp3DENOM[ieppV]=1 if pp3[ieppV]==0
@@ -259,11 +264,13 @@
       const cxtype_amp_v chi[2] = { cxmake( fpsqrt( chi0r2 ), 0 ),     // hack: dummy[ieppV] is not used if pp[ieppV]==0
                                 cxternary( ( pp3 == 0. ),
                                            cxmake( -nh, 0 ),
-                                           cxmake( (fptype_amp)nh * pvec1, pvec2 ) / fpsqrt( (fptype_amp)2. * ppDENOM * pp3DENOM ) ) }; // hack: dummy[ieppV] is not used if pp[ieppV]==0
-      const cxtype_amp_v fiB_2 = sfomega[0] * chi[im];
-      const cxtype_amp_v fiB_3 = sfomega[0] * chi[ip];
-      const cxtype_amp_v fiB_4 = sfomega[1] * chi[im];
-      const cxtype_amp_v fiB_5 = sfomega[1] * chi[ip];
+                                           cxmake( nh * pvec1, pvec2 ) / fpsqrt( (fptype_amp)2. * ppDENOM * pp3DENOM ) ) }; // hack: dummy[ieppV] is not used if pp[ieppV]==0
+      const cxtype_amp_v chi_im = tp * chi[0] + tm * chi[1]; // chi[im]
+      const cxtype_amp_v chi_ip = tm * chi[0] + tp * chi[1]; // chi[ip]
+      const cxtype_amp_v fiB_2 = sfomega[0] * chi_im;
+      const cxtype_amp_v fiB_3 = sfomega[0] * chi_ip;
+      const cxtype_amp_v fiB_4 = sfomega[1] * chi_im;
+      const cxtype_amp_v fiB_5 = sfomega[1] * chi_ip;
       // Choose between the results from branch A and branch B
       const bool_v mask = ( pp == 0. );
       w[0] = cxternary( mask, fiA_2, fiB_2 );
@@ -289,8 +296,8 @@
       volatile fptype_amp_sv sqp0p3DENOM = fpternary( sqp0p3 != 0, (fptype_sv)sqp0p3, 1. ); // hack: dummy sqp0p3DENOM[ieppV]=1 if sqp0p3[ieppV]==0
       cxtype_sv chi[2] = { cxmake( (fptype_v)sqp0p3, 0. ),
                            cxternary( sqp0p3 == 0,
-                                      cxmake( -(fptype_amp)nhel * fpsqrt( (fptype_amp)2. * pvec0 ), 0. ),
-                                      cxmake( (fptype_amp)nh * pvec1, pvec2 ) / (const fptype_v)sqp0p3DENOM ) }; // hack: dummy[ieppV] is not used if sqp0p3[ieppV]==0
+                                      cxmake( -nhel * fpsqrt( (fptype_amp)2. * pvec0 ), 0. ),
+                                      cxmake( nh * pvec1, pvec2 ) / (const fptype_v)sqp0p3DENOM ) }; // hack: dummy[ieppV] is not used if sqp0p3[ieppV]==0
 #else
       // pvec0+pvec3 is a cancelling difference of two ~|p| numbers for a
       // backward-moving massless fermion, and everything below divides by it.
@@ -303,22 +310,12 @@
                                           fptype_sv{ 0 },
                                           fpsqrt( p0p3stable ) * (fptype_amp)nsf );
       const cxtype_amp_sv chi[2] = { cxmake( sqp0p3, 0. ),
-                                 ( sqp0p3 == 0. ? cxmake( -(fptype_amp)nhel * fpsqrt( (fptype_amp)2. * pvec0 ), 0. ) : cxmake( (fptype_amp)nh * pvec1, pvec2 ) / sqp0p3 ) };
+                                 ( sqp0p3 == 0. ? cxmake( -nhel * fpsqrt( (fptype_amp)2. * pvec0 ), 0. ) : cxmake( nh * pvec1, pvec2 ) / sqp0p3 ) };
 #endif
-      if( nh == 1 )
-      {
-        w[0] = cxzero_sv();
-        w[1] = cxzero_sv();
-        w[2] = chi[0];
-        w[3] = chi[1];
-      }
-      else
-      {
-        w[0] = chi[1];
-        w[1] = chi[0];
-        w[2] = cxzero_sv();
-        w[3] = cxzero_sv();
-      }
+      w[0] = tm * chi[1];
+      w[1] = tm * chi[0];
+      w[2] = tp * chi[0];
+      w[3] = tp * chi[1];
     }
     mgDebug( 1, __FUNCTION__ );
     return;
@@ -332,7 +329,7 @@
   __host__ __device__ void
   ipzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavour
           ALOHAOBJ & fi,          // output: wavefunctions
@@ -347,20 +344,17 @@
     fi.pvec[2] = fptype_momenta_sv{ 0 };
     fi.pvec[3] = -pvec3_ * static_cast<fptype_momenta>(nsf);
     fi.flv_index = flv;
-    const int nh = nhel * nsf;
+    const fptype_amp_sv nh = nhel * static_cast<fptype_amp>( nsf );
+    // BRANCHLESS HELICITY: the helicity is not a branch condition but enters as the truth values
+    // tp (nh == +1) and tm (nh == -1), 1 or 0, which multiply the two alternatives. Different
+    // SIMD lanes or GPU threads can then carry different helicities and still run in lock-step.
+    const fptype_amp_sv tp = ( 1 + nh ) * (fptype_amp)0.5;
+    const fptype_amp_sv tm = 1 - tp;
     const cxtype_amp_sv sqp0p3 = cxmake( fpsqrt( (fptype_amp)2. * pvec3 ) * (fptype_amp)nsf, 0. );
-    w[0] = cxmake( fpamp_of_mom( fi.pvec[1] ), fpamp_of_mom( fi.pvec[2] ) );
-    if( nh == 1 )
-    {
-      w[1] = cxmake( fpamp_of_mom( fi.pvec[1] ), fpamp_of_mom( fi.pvec[2] ) );
-      w[2] = sqp0p3;
-    }
-    else
-    {
-      w[1] = sqp0p3;
-      w[2] = cxmake( fpamp_of_mom( fi.pvec[1] ), fpamp_of_mom( fi.pvec[2] ) );
-    }
-    w[3] = cxmake( fpamp_of_mom( fi.pvec[1] ), fpamp_of_mom( fi.pvec[2] ) );
+    w[0] = cxzero_sv();
+    w[1] = tm * sqp0p3;
+    w[2] = tp * sqp0p3;
+    w[3] = cxzero_sv();
     mgDebug( 1, __FUNCTION__ );
     return;
   }
@@ -373,7 +367,7 @@
   __host__ __device__ void
   imzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavour
           ALOHAOBJ & fi,          // output: wavefunctions
@@ -388,20 +382,17 @@
     fi.pvec[2] = fptype_momenta_sv{ 0 };
     fi.pvec[3] = -pvec3_ * static_cast<fptype_momenta>(nsf);
     fi.flv_index = flv;
-    const int nh = nhel * nsf;
-    const cxtype_amp_sv chi = cxmake( -(fptype_amp)nhel * fpsqrt( -(fptype_amp)2. * pvec3 ), 0. );
+    const fptype_amp_sv nh = nhel * static_cast<fptype_amp>( nsf );
+    // BRANCHLESS HELICITY: the helicity is not a branch condition but enters as the truth values
+    // tp (nh == +1) and tm (nh == -1), 1 or 0, which multiply the two alternatives. Different
+    // SIMD lanes or GPU threads can then carry different helicities and still run in lock-step.
+    const fptype_amp_sv tp = ( 1 + nh ) * (fptype_amp)0.5;
+    const fptype_amp_sv tm = 1 - tp;
+    const cxtype_amp_sv chi = cxmake( -nhel * fpsqrt( -(fptype_amp)2. * pvec3 ), 0. );
+    w[0] = tm * chi;
     w[1] = cxzero_sv();
     w[2] = cxzero_sv();
-    if( nh == 1 )
-    {
-      w[0] = cxzero_sv();
-      w[3] = chi;
-    }
-    else
-    {
-      w[0] = chi;
-      w[3] = cxzero_sv();
-    }
+    w[3] = tp * chi;
     mgDebug( 1, __FUNCTION__ );
     return;
   }
@@ -414,7 +405,7 @@
   __host__ __device__ void
   ixzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavour
           ALOHAOBJ & fi,          // output: wavefunctions
@@ -436,7 +427,12 @@
     fi.pvec[3] = -pvec3_ * static_cast<fptype_momenta>(nsf);
     fi.flv_index = flv;
 
-    const int nh = nhel * nsf;
+    const fptype_amp_sv nh = nhel * static_cast<fptype_amp>( nsf );
+    // BRANCHLESS HELICITY: the helicity is not a branch condition but enters as the truth values
+    // tp (nh == +1) and tm (nh == -1), 1 or 0, which multiply the two alternatives. Different
+    // SIMD lanes or GPU threads can then carry different helicities and still run in lock-step.
+    const fptype_amp_sv tp = ( 1 + nh ) * (fptype_amp)0.5;
+    const fptype_amp_sv tm = 1 - tp;
     //const float sqp0p3 = sqrtf( pvec0 + pvec3 ) * nsf; // AV: why force a float here?
     // pvec0+pvec3 is a cancelling difference of two ~|p| numbers for a
     // backward-moving massless fermion, and everything below divides by it.
@@ -448,21 +444,11 @@
                                             pvec0 + pvec3 );
     const fptype_amp_sv sqp0p3 = fpsqrt( p0p3stable ) * static_cast<fptype_amp>(nsf);
     const cxtype_amp_sv chi0 = cxmake( sqp0p3, 0. );
-    const cxtype_amp_sv chi1 = cxmake( (fptype_amp)nh * pvec1 / sqp0p3, pvec2 / sqp0p3 );
-    if( nh == 1 )
-    {
-      w[0] = cxzero_sv();
-      w[1] = cxzero_sv();
-      w[2] = chi0;
-      w[3] = chi1;
-    }
-    else
-    {
-      w[0] = chi1;
-      w[1] = chi0;
-      w[2] = cxzero_sv();
-      w[3] = cxzero_sv();
-    }
+    const cxtype_amp_sv chi1 = cxmake( nh * pvec1 / sqp0p3, pvec2 / sqp0p3 );
+    w[0] = tm * chi1;
+    w[1] = tm * chi0;
+    w[2] = tp * chi0;
+    w[3] = tp * chi1;
     mgDebug( 1, __FUNCTION__ );
     return;
   }
@@ -474,7 +460,7 @@
   __host__ __device__ void
   vxxxxx( const fptype_momenta momenta[], // input: momenta
           const fptype_amp vmass,     // input: vector boson mass
-          const int nhel,         // input: -1, 0 (only if vmass!=0) or +1 (helicity of vector boson)
+          const fptype_amp_sv nhel, // input: -1, 0 (only if vmass!=0) or +1 (helicity of vector boson, per SIMD lane/GPU thread)
           const int nsv,          // input: +1 (final) or -1 (initial)
           const int flv,          // input: flavour
           ALOHAOBJ & vc,          // output: wavefunctions
@@ -500,11 +486,14 @@
     vc.pvec[3] = pvec3_ * static_cast<fptype_momenta>(nsv);
     vc.flv_index = flv;
     const fptype_amp sqh = fpsqrt( 0.5 ); // AV this is > 0!
-    const fptype_amp hel = nhel;
+    const fptype_amp_sv hel = nhel;
+    // BRANCHLESS HELICITY: |hel| is hel * hel for hel in { -1, 0, +1 }, so that the helicity can
+    // differ from one SIMD lane or GPU thread to the next with no branch and no std::abs
+    const fptype_amp_sv abshel = hel * hel;
     if( vmass != 0. )
     {
-      const int nsvahl = nsv * std::abs( hel );
-      const fptype_amp hel0 = 1. - std::abs( hel );
+      const fptype_amp_sv nsvahl = static_cast<fptype_amp>( nsv ) * abshel;
+      const fptype_amp_sv hel0 = 1 - abshel; // truth value of hel == 0 (longitudinal polarisation)
 #ifndef MGONGPU_CPPSIMD
       const fptype_amp_sv pt2 = ( pvec1 * pvec1 ) + ( pvec2 * pvec2 );
       const fptype_amp_sv pp = fpmin( pvec0, fpsqrt( pt2 + ( pvec3 * pvec3 ) ) );
@@ -545,9 +534,9 @@
       const fptype_amp_sv pt = fpmin( pp, fpsqrt( pt2 ) );
       // Branch A: pp == 0.
       const cxtype_amp vcA_2 = cxmake( 0, 0 );
-      const cxtype_amp vcA_3 = cxmake( -hel * sqh, 0 );
-      const cxtype_amp vcA_4 = cxmake( 0, nsvahl * sqh );
-      const cxtype_amp vcA_5 = cxmake( hel0, 0 );
+      const cxtype_amp_v vcA_3 = cxmake( -hel * sqh, 0 );
+      const cxtype_amp_v vcA_4 = cxmake( 0., nsvahl * sqh );
+      const cxtype_amp_v vcA_5 = cxmake( hel0, 0 );
       // Branch B: pp != 0.
       volatile fptype_amp_v ppDENOM = fpternary( pp != 0, pp, 1. ); // hack: ppDENOM[ieppV]=1 if pp[ieppV]==0
       const fptype_amp_v emp = pvec0 / ( vmass * ppDENOM );         // hack: dummy[ieppV] is not used if pp[ieppV]==0
@@ -556,11 +545,11 @@
       // Branch B1: pp != 0. and pt != 0.
       volatile fptype_amp_v ptDENOM = fpternary( pt != 0, pt, 1. );                                                     // hack: ptDENOM[ieppV]=1 if pt[ieppV]==0
       const fptype_amp_v pzpt = pvec3 / ( ppDENOM * ptDENOM ) * sqh * hel;                                              // hack: dummy[ieppV] is not used if pp[ieppV]==0
-      const cxtype_amp_v vcB1_3 = cxmake( hel0 * pvec1 * emp - pvec1 * pzpt, -fpamp_scalar(nsvahl) * pvec2 / ptDENOM * sqh ); // hack: dummy[ieppV] is not used if pt[ieppV]==0
-      const cxtype_amp_v vcB1_4 = cxmake( hel0 * pvec2 * emp - pvec2 * pzpt, fpamp_scalar(nsvahl) * pvec1 / ptDENOM * sqh );  // hack: dummy[ieppV] is not used if pt[ieppV]==0
+      const cxtype_amp_v vcB1_3 = cxmake( hel0 * pvec1 * emp - pvec1 * pzpt, -nsvahl * pvec2 / ptDENOM * sqh ); // hack: dummy[ieppV] is not used if pt[ieppV]==0
+      const cxtype_amp_v vcB1_4 = cxmake( hel0 * pvec2 * emp - pvec2 * pzpt, nsvahl * pvec1 / ptDENOM * sqh );  // hack: dummy[ieppV] is not used if pt[ieppV]==0
       // Branch B2: pp != 0. and pt == 0.
-      const cxtype_amp vcB2_3 = cxmake( -hel * sqh, 0. );
-      const cxtype_amp_v vcB2_4 = cxmake( 0., fpamp_scalar(nsvahl) * fpternary( ( pvec3 < 0 ), -sqh, sqh ) ); // AV: removed an abs here
+      const cxtype_amp_v vcB2_3 = cxmake( -hel * sqh, 0. );
+      const cxtype_amp_v vcB2_4 = cxmake( 0., nsvahl * fpternary( ( pvec3 < 0 ), -sqh, sqh ) ); // AV: removed an abs here
       // Choose between the results from branch A and branch B (and from branch B1 and branch B2)
       const bool_v mask = ( pp == 0. );
       const bool_v maskB = ( pt != 0. );
@@ -599,11 +588,11 @@
       // Branch A: pt != 0.
       volatile fptype_amp_v ptDENOM = fpternary( pt != 0, pt, 1. );                             // hack: ptDENOM[ieppV]=1 if pt[ieppV]==0
       const fptype_amp_v pzpt = pvec3 / ( pp * ptDENOM ) * sqh * hel;                           // hack: dummy[ieppV] is not used if pt[ieppV]==0
-      const cxtype_amp_v vcA_3 = cxmake( -pvec1 * pzpt, -fpamp_scalar(nsv) * pvec2 / ptDENOM * sqh ); // hack: dummy[ieppV] is not used if pt[ieppV]==0
-      const cxtype_amp_v vcA_4 = cxmake( -pvec2 * pzpt, fpamp_scalar(nsv) * pvec1 / ptDENOM * sqh );  // hack: dummy[ieppV] is not used if pt[ieppV]==0
+      const cxtype_amp_v vcA_3 = cxmake( -pvec1 * pzpt, -(fptype_amp)nsv * pvec2 / ptDENOM * sqh ); // hack: dummy[ieppV] is not used if pt[ieppV]==0
+      const cxtype_amp_v vcA_4 = cxmake( -pvec2 * pzpt, (fptype_amp)nsv * pvec1 / ptDENOM * sqh );  // hack: dummy[ieppV] is not used if pt[ieppV]==0
       // Branch B: pt == 0.
-      const cxtype_amp vcB_3 = cxmake( -(fptype_amp)hel * sqh, 0 );
-      const cxtype_amp_v vcB_4 = cxmake( 0, fpamp_scalar(nsv) * fpternary( ( pvec3 < 0 ), -sqh, sqh ) ); // AV: removed an abs here
+      const cxtype_amp_v vcB_3 = cxmake( -hel * sqh, 0 );
+      const cxtype_amp_v vcB_4 = cxmake( 0, (fptype_amp)nsv * fpternary( ( pvec3 < 0 ), -sqh, sqh ) ); // AV: removed an abs here
       // Choose between the results from branch A and branch B
       const bool_v mask = ( pt != 0. );
       w[1] = cxternary( mask, vcA_3, vcB_3 );
@@ -654,7 +643,7 @@
   __host__ __device__ void
   oxxxxx( const fptype_momenta momenta[], // input: momenta
           const fptype_amp fmass,     // input: fermion mass
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           int flv,                // input: flavour
           ALOHAOBJ & fo,          // output: wavefunctions
@@ -680,7 +669,12 @@
     fo.pvec[2] = pvec2_ * static_cast<fptype_momenta>(nsf);
     fo.pvec[3] = pvec3_ * static_cast<fptype_momenta>(nsf);
     fo.flv_index = flv;
-    const int nh = nhel * nsf;
+    const fptype_amp_sv nh = nhel * static_cast<fptype_amp>( nsf );
+    // BRANCHLESS HELICITY: the helicity is not a branch condition but enters as the truth values
+    // tp (nh == +1) and tm (nh == -1), 1 or 0, which multiply the two alternatives. Different
+    // SIMD lanes or GPU threads can then carry different helicities and still run in lock-step.
+    const fptype_amp_sv tp = ( 1 + nh ) * (fptype_amp)0.5;
+    const fptype_amp_sv tm = 1 - tp;
     if( fmass != 0. )
     {
 #ifndef MGONGPU_CPPSIMD
@@ -691,12 +685,12 @@
         fptype_amp sqm[2] = { fpsqrt( std::abs( fmass ) ), 0. }; // possibility of negative fermion masses
         //sqm[1] = ( fmass < 0. ? -abs( sqm[0] ) : abs( sqm[0] ) ); // AV: why abs here?
         sqm[1] = ( fmass < 0. ? -sqm[0] : sqm[0] ); // AV: removed an abs here
-        const int ip = -( ( 1 - nh ) / 2 ) * nhel;  // NB: Fortran sqm(0:1) also has indexes 0,1 as in C++
-        const int im = ( 1 + nh ) / 2 * nhel;       // NB: Fortran sqm(0:1) also has indexes 0,1 as in C++
-        w[0] = cxmake( im * sqm[std::abs( ip )], 0 );
-        w[1] = cxmake( ip * nsf * sqm[std::abs( ip )], 0 );
-        w[2] = cxmake( im * nsf * sqm[std::abs( im )], 0 );
-        w[3] = cxmake( ip * sqm[std::abs( im )], 0 );
+        // The Fortran ip = -( ( 1 - nh ) / 2 ) * nhel and im = ( 1 + nh ) / 2 * nhel are -tm * nhel and
+        // tp * nhel, so that |ip| = tm and |im| = tp: sqm[|ip|] is sqm[1] wherever ip != 0, and so on
+        w[0] = cxmake( tp * nhel * sqm[0], 0 );
+        w[1] = cxmake( -tm * nhel * nsf * sqm[1], 0 );
+        w[2] = cxmake( tp * nhel * nsf * sqm[1], 0 );
+        w[3] = cxmake( -tm * nhel * sqm[0], 0 );
       }
       else
       {
@@ -704,17 +698,18 @@
                                fptype_amp( 1 + nsf - ( 1 - nsf ) * nh ) * (fptype_amp)0.5 };
         fptype_amp omega[2] = { fpsqrt( pvec0 + pp ), 0. };
         omega[1] = fmass / omega[0];
-        const int ip = ( 1 + nh ) / 2; // NB: Fortran is (3+nh)/2 because omega(2) has indexes 1,2 and not 0,1
-        const int im = ( 1 - nh ) / 2; // NB: Fortran is (3-nh)/2 because omega(2) has indexes 1,2 and not 0,1
-        const fptype_amp sfomeg[2] = { sf[0] * omega[ip], sf[1] * omega[im] };
+        // The Fortran ip/im indices ( 3 +- nh ) / 2 (omega(2) has indexes 1,2) are the truth values tp/tm
+        const fptype_amp sfomeg[2] = { sf[0] * ( tm * omega[0] + tp * omega[1] ), sf[1] * ( tp * omega[0] + tm * omega[1] ) };
         const fptype_amp pp3 = fpmax( fpternary( fpsignbit( pvec3 ), ( pvec1 * pvec1 + pvec2 * pvec2 ) / ( pp - pvec3 ), pp + pvec3 ), 0. );
         const cxtype_amp chi[2] = { cxmake( fpsqrt( pp3 * (fptype_amp)0.5 / pp ), 0. ),
                                 ( ( pp3 == 0. ) ? cxmake( -nh, 0. )
                                                 : cxmake( nh * pvec1, -pvec2 ) / fpsqrt( (fptype_amp)2. * pp * pp3 ) ) };
-        w[0] = sfomeg[1] * chi[im];
-        w[1] = sfomeg[1] * chi[ip];
-        w[2] = sfomeg[0] * chi[im];
-        w[3] = sfomeg[0] * chi[ip];
+        const cxtype_amp chi_im = tp * chi[0] + tm * chi[1]; // chi[im]
+        const cxtype_amp chi_ip = tm * chi[0] + tp * chi[1]; // chi[ip]
+        w[0] = sfomeg[1] * chi_im;
+        w[1] = sfomeg[1] * chi_ip;
+        w[2] = sfomeg[0] * chi_im;
+        w[3] = sfomeg[0] * chi_ip;
       }
 #else
       volatile fptype_amp_sv p2 = pvec1 * pvec1 + pvec2 * pvec2 + pvec3 * pvec3; // volatile fixes #736
@@ -723,20 +718,17 @@
       // NB: Do not use "abs" for floats! It returns an integer with no build warning! Use std::abs!
       fptype_amp sqm[2] = { fpsqrt( std::abs( fmass ) ), 0 }; // possibility of negative fermion masses
       sqm[1] = ( fmass < 0 ? -sqm[0] : sqm[0] );          // AV: removed an abs here (as above)
-      const int ipA = -( ( 1 - nh ) / 2 ) * nhel;
-      const int imA = ( 1 + nh ) / 2 * nhel;
-      const cxtype_amp foA_2 = imA * sqm[std::abs( ipA )];
-      const cxtype_amp foA_3 = ipA * nsf * sqm[std::abs( ipA )];
-      const cxtype_amp foA_4 = imA * nsf * sqm[std::abs( imA )];
-      const cxtype_amp foA_5 = ipA * sqm[std::abs( imA )];
+      // (see the scalar code above for the truth values tp/tm in place of the Fortran ip/im indices)
+      const cxtype_amp_v foA_2 = cxmake( tp * nhel * sqm[0], 0 );
+      const cxtype_amp_v foA_3 = cxmake( -tm * nhel * (fptype_amp)nsf * sqm[1], 0 );
+      const cxtype_amp_v foA_4 = cxmake( tp * nhel * (fptype_amp)nsf * sqm[1], 0 );
+      const cxtype_amp_v foA_5 = cxmake( -tm * nhel * sqm[0], 0 );
       // Branch B: pp != 0.
-      const fptype_amp sf[2] = { fptype_amp( 1 + nsf + ( 1 - nsf ) * nh ) * (fptype_amp)0.5,
-                             fptype_amp( 1 + nsf - ( 1 - nsf ) * nh ) * (fptype_amp)0.5 };
+      const fptype_amp_v sf[2] = { ( (fptype_amp)( 1 + nsf ) + (fptype_amp)( 1 - nsf ) * nh ) * (fptype_amp)0.5,
+                                   ( (fptype_amp)( 1 + nsf ) - (fptype_amp)( 1 - nsf ) * nh ) * (fptype_amp)0.5 };
       fptype_v omega[2] = { fpsqrt( pvec0 + pp ), 0 };
       omega[1] = fmass / omega[0];
-      const int ipB = ( 1 + nh ) / 2;
-      const int imB = ( 1 - nh ) / 2;
-      const fptype_amp_v sfomeg[2] = { sf[0] * omega[ipB], sf[1] * omega[imB] };
+      const fptype_amp_v sfomeg[2] = { sf[0] * ( tm * omega[0] + tp * omega[1] ), sf[1] * ( tp * omega[0] + tm * omega[1] ) };
       const fptype_amp_v pp3 = fpmax( fpternary( fpsignbit( pvec3 ), ( pvec1 * pvec1 + pvec2 * pvec2 ) / ( pp - pvec3 ), pp + pvec3 ), 0. );
       volatile fptype_amp_v ppDENOM = fpternary( pp != 0, pp, 1. );    // hack: ppDENOM[ieppV]=1 if pp[ieppV]==0
       volatile fptype_amp_v pp3DENOM = fpternary( pp3 != 0, pp3, 1. ); // hack: pp3DENOM[ieppV]=1 if pp3[ieppV]==0
@@ -744,11 +736,13 @@
       const cxtype_amp_v chi[2] = { cxmake( fpsqrt( chi0r2 ), 0. ),    // hack: dummy[ieppV] is not used if pp[ieppV]==0
                                 ( cxternary( ( pp3 == 0. ),
                                              cxmake( -nh, 0. ),
-                                             cxmake( (fptype_amp)nh * pvec1, -pvec2 ) / fpsqrt( (fptype_amp)2. * ppDENOM * pp3DENOM ) ) ) }; // hack: dummy[ieppV] is not used if pp[ieppV]==0
-      const cxtype_amp_v foB_2 = sfomeg[1] * chi[imB];
-      const cxtype_amp_v foB_3 = sfomeg[1] * chi[ipB];
-      const cxtype_amp_v foB_4 = sfomeg[0] * chi[imB];
-      const cxtype_amp_v foB_5 = sfomeg[0] * chi[ipB];
+                                             cxmake( nh * pvec1, -pvec2 ) / fpsqrt( (fptype_amp)2. * ppDENOM * pp3DENOM ) ) ) }; // hack: dummy[ieppV] is not used if pp[ieppV]==0
+      const cxtype_amp_v chi_im = tp * chi[0] + tm * chi[1]; // chi[im]
+      const cxtype_amp_v chi_ip = tm * chi[0] + tp * chi[1]; // chi[ip]
+      const cxtype_amp_v foB_2 = sfomeg[1] * chi_im;
+      const cxtype_amp_v foB_3 = sfomeg[1] * chi_ip;
+      const cxtype_amp_v foB_4 = sfomeg[0] * chi_im;
+      const cxtype_amp_v foB_5 = sfomeg[0] * chi_ip;
       // Choose between the results from branch A and branch B
       const bool_v mask = ( pp == 0. );
       w[0] = cxternary( mask, foA_2, foB_2 );
@@ -775,7 +769,7 @@
       const cxtype_amp_v chi[2] = { cxmake( (fptype_v)sqp0p3, 0. ),
                                 cxternary( ( sqp0p3 == 0. ),
                                            cxmake( -nhel, 0. ) * fpsqrt( (fptype_amp)2. * pvec0 ),
-                                           cxmake( (fptype_amp)nh * pvec1, -pvec2 ) / (const fptype_amp_sv)sqp0p3DENOM ) }; // hack: dummy[ieppV] is not used if sqp0p3[ieppV]==0
+                                           cxmake( nh * pvec1, -pvec2 ) / (const fptype_amp_sv)sqp0p3DENOM ) }; // hack: dummy[ieppV] is not used if sqp0p3[ieppV]==0
 #else
       // pvec0+pvec3 is a cancelling difference of two ~|p| numbers for a
       // backward-moving massless fermion, and everything below divides by it.
@@ -788,22 +782,12 @@
                                           fptype_sv{ 0 },
                                           fpsqrt( p0p3stable ) * (fptype_amp)nsf );
       const cxtype_amp_sv chi[2] = { cxmake( sqp0p3, 0. ),
-                                 ( sqp0p3 == 0. ? cxmake( -nhel, 0. ) * fpsqrt( (fptype_amp)2. * pvec0 ) : cxmake( (fptype_amp)nh * pvec1, -pvec2 ) / sqp0p3 ) };
+                                 ( sqp0p3 == 0. ? cxmake( -nhel, 0. ) * fpsqrt( (fptype_amp)2. * pvec0 ) : cxmake( nh * pvec1, -pvec2 ) / sqp0p3 ) };
 #endif
-      if( nh == 1 )
-      {
-        w[0] = chi[0];
-        w[1] = chi[1];
-        w[2] = cxzero_sv();
-        w[3] = cxzero_sv();
-      }
-      else
-      {
-        w[0] = cxzero_sv();
-        w[1] = cxzero_sv();
-        w[2] = chi[1];
-        w[3] = chi[0];
-      }
+      w[0] = tp * chi[0];
+      w[1] = tp * chi[1];
+      w[2] = tm * chi[1];
+      w[3] = tm * chi[0];
     }
     mgDebug( 1, __FUNCTION__ );
     return;
@@ -817,7 +801,7 @@
   __host__ __device__ void
   opzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavour
           ALOHAOBJ & fo,          // output: wavefunctions
@@ -832,20 +816,17 @@
     fo.pvec[2] = fptype_momenta_sv{ 0 };
     fo.pvec[3] = pvec3_ * static_cast<fptype_momenta>(nsf);
     fo.flv_index = flv;
-    const int nh = nhel * nsf;
+    const fptype_amp_sv nh = nhel * static_cast<fptype_amp>( nsf );
+    // BRANCHLESS HELICITY: the helicity is not a branch condition but enters as the truth values
+    // tp (nh == +1) and tm (nh == -1), 1 or 0, which multiply the two alternatives. Different
+    // SIMD lanes or GPU threads can then carry different helicities and still run in lock-step.
+    const fptype_amp_sv tp = ( 1 + nh ) * (fptype_amp)0.5;
+    const fptype_amp_sv tm = 1 - tp;
     const cxtype_amp_sv csqp0p3 = cxmake( fpsqrt( (fptype_amp)2. * pvec3 ) * (fptype_amp)nsf, 0. );
+    w[0] = tp * csqp0p3;
     w[1] = cxzero_sv();
     w[2] = cxzero_sv();
-    if( nh == 1 )
-    {
-      w[0] = csqp0p3;
-      w[3] = cxzero_sv();
-    }
-    else
-    {
-      w[0] = cxzero_sv();
-      w[3] = csqp0p3;
-    }
+    w[3] = tm * csqp0p3;
     mgDebug( 1, __FUNCTION__ );
     return;
   }
@@ -858,7 +839,7 @@
   __host__ __device__ void
   omzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavour
           ALOHAOBJ & fo,          // output: wavefunctions
@@ -873,23 +854,17 @@
     fo.pvec[2] = fptype_momenta_sv{ 0 };
     fo.pvec[3] = pvec3_ * static_cast<fptype_momenta>(nsf);
     fo.flv_index = flv;
-    const int nh = nhel * nsf;
+    const fptype_amp_sv nh = nhel * static_cast<fptype_amp>( nsf );
+    // BRANCHLESS HELICITY: the helicity is not a branch condition but enters as the truth values
+    // tp (nh == +1) and tm (nh == -1), 1 or 0, which multiply the two alternatives. Different
+    // SIMD lanes or GPU threads can then carry different helicities and still run in lock-step.
+    const fptype_amp_sv tp = ( 1 + nh ) * (fptype_amp)0.5;
+    const fptype_amp_sv tm = 1 - tp;
     const cxtype_amp_sv chi1 = cxmake( -nhel, 0. ) * fpsqrt( -(fptype_amp)2. * pvec3 );
-    if( nh == 1 )
-    {
-      w[0] = cxzero_sv();
-      w[1] = chi1;
-      w[2] = cxzero_sv();
-      w[3] = cxzero_sv();
-    }
-    else
-    {
-      w[0] = cxzero_sv();
-      w[1] = cxzero_sv();
-      w[2] = chi1;
-      //w[3] = chi1; // AV: BUG!
-      w[3] = cxzero_sv(); // AV: BUG FIX
-    }
+    w[0] = cxzero_sv();
+    w[1] = tp * chi1;
+    w[2] = tm * chi1;
+    w[3] = cxzero_sv();
     mgDebug( 1, __FUNCTION__ );
     return;
   }
@@ -902,7 +877,7 @@
   __host__ __device__ void
   oxzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavour
           ALOHAOBJ & fo,          // output: wavefunctions
@@ -924,7 +899,12 @@
     fo.pvec[2] = pvec2_ * static_cast<fptype_momenta>(nsf);
     fo.pvec[3] = pvec3_ * static_cast<fptype_momenta>(nsf);
     fo.flv_index = flv;
-    const int nh = nhel * nsf;
+    const fptype_amp_sv nh = nhel * static_cast<fptype_amp>( nsf );
+    // BRANCHLESS HELICITY: the helicity is not a branch condition but enters as the truth values
+    // tp (nh == +1) and tm (nh == -1), 1 or 0, which multiply the two alternatives. Different
+    // SIMD lanes or GPU threads can then carry different helicities and still run in lock-step.
+    const fptype_amp_sv tp = ( 1 + nh ) * (fptype_amp)0.5;
+    const fptype_amp_sv tm = 1 - tp;
     //const float sqp0p3 = sqrtf( pvec0 + pvec3 ) * nsf; // AV: why force a float here?
     // pvec0+pvec3 is a cancelling difference of two ~|p| numbers for a
     // backward-moving massless fermion, and everything below divides by it.
@@ -936,21 +916,11 @@
                                             pvec0 + pvec3 );
     const fptype_amp_sv sqp0p3 = fpsqrt( p0p3stable ) * static_cast<fptype_amp>(nsf);
     const cxtype_amp_sv chi0 = cxmake( sqp0p3, 0. );
-    const cxtype_amp_sv chi1 = cxmake( (fptype_amp)nh * pvec1 / sqp0p3, -pvec2 / sqp0p3 );
-    if( nh == 1 )
-    {
-      w[0] = chi0;
-      w[1] = chi1;
-      w[2] = cxzero_sv();
-      w[3] = cxzero_sv();
-    }
-    else
-    {
-      w[0] = cxzero_sv();
-      w[1] = cxzero_sv();
-      w[2] = chi1;
-      w[3] = chi0;
-    }
+    const cxtype_amp_sv chi1 = cxmake( nh * pvec1 / sqp0p3, -pvec2 / sqp0p3 );
+    w[0] = tp * chi0;
+    w[1] = tp * chi1;
+    w[2] = tm * chi1;
+    w[3] = tm * chi0;
     mgDebug( 1, __FUNCTION__ );
     return;
   }
diff --git a/aloha/template_files/madmatrix/helas_fd.h b/aloha/template_files/madmatrix/helas_fd.h
index 2aed9342cf..723037a550 100644
--- a/aloha/template_files/madmatrix/helas_fd.h
+++ b/aloha/template_files/madmatrix/helas_fd.h
@@ -26,7 +26,7 @@
   __host__ __device__ INLINE void
   ixxxxx( const fptype_momenta momenta[], // input: momenta
           const fptype_amp fmass,     // input: fermion mass
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavour
           ALOHAOBJ & fi,          // output: aloha objects
@@ -41,7 +41,7 @@
   __host__ __device__ INLINE void
   ipzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavor index
           ALOHAOBJ & fi,          // output: wavefunctions
@@ -56,7 +56,7 @@
   __host__ __device__ INLINE void
   imzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavor index
           ALOHAOBJ & fi,          // output: wavefunctions
@@ -71,7 +71,7 @@
   __host__ __device__ INLINE void
   ixzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavor index
           ALOHAOBJ & fi,          // output: wavefunctions
@@ -85,7 +85,7 @@
   __host__ __device__ INLINE void
   vxxxxx( const fptype_momenta momenta[], // input: momenta
           const fptype_amp vmass,     // input: vector boson mass
-          const int nhel,         // input: -1, 0 (only if vmass!=0) or +1 (helicity of vector boson)
+          const fptype_amp_sv nhel, // input: -1, 0 (only if vmass!=0) or +1 (helicity of vector boson, per SIMD lane/GPU thread)
           const int nsv,          // input: +1 (final) or -1 (initial)
           const int flv,          // input: flavor index
           ALOHAOBJ & vc,          // output: wavefunctions
@@ -113,7 +113,7 @@
   __host__ __device__ INLINE void
   oxxxxx( const fptype_momenta momenta[], // input: momenta
           const fptype_amp fmass,     // input: fermion mass
-          const int nhel,         // input: -1, 0 (only if vmass!=0) or +1 (helicity of vector boson)
+          const fptype_amp_sv nhel, // input: -1, 0 (only if vmass!=0) or +1 (helicity of vector boson, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavor index
           ALOHAOBJ & fo,          // output: wavefunctions
@@ -128,7 +128,7 @@
   __host__ __device__ INLINE void
   opzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavor index
           ALOHAOBJ & fo,          // output: wavefunctions
@@ -143,7 +143,7 @@
   __host__ __device__ INLINE void
   omzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavor index
           ALOHAOBJ & fo,          // output: wavefunctions
@@ -157,7 +157,7 @@
   __host__ __device__ INLINE void
   oxzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavor index
           ALOHAOBJ & fo,          // output: wavefunctions
@@ -198,7 +198,7 @@
   __host__ __device__ void
   ixxxxx( const fptype_momenta momenta[], // input: momenta
           const fptype_amp fmass,     // input: fermion mass
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavour
           ALOHAOBJ & fi,          // output: wavefunctions
@@ -223,7 +223,12 @@
     fi.pvec[2] = -pvec2_ * static_cast<fptype_momenta>(nsf);
     fi.pvec[3] = -pvec3_ * static_cast<fptype_momenta>(nsf);
     fi.flv_index = flv;
-    const int nh = nhel * nsf;
+    const fptype_amp_sv nh = nhel * static_cast<fptype_amp>( nsf );
+    // BRANCHLESS HELICITY: the helicity is not a branch condition but enters as the truth values
+    // tp (nh == +1) and tm (nh == -1), 1 or 0, which multiply the two alternatives. Different
+    // SIMD lanes or GPU threads can then carry different helicities and still run in lock-step.
+    const fptype_amp_sv tp = ( 1 + nh ) * (fptype_amp)0.5;
+    const fptype_amp_sv tm = 1 - tp;
     if( fmass != 0. )
     {
 #ifndef MGONGPU_CPPSIMD
@@ -232,11 +237,9 @@
       volatile fptype_amp_sv p2 = pvec1 * pvec1 + pvec2 * pvec2 + pvec3 * pvec3; // volatile fixes #736
       const fptype_amp_sv pp = fpmin( pvec0, fpsqrt( p2 ) );
 #endif
-      // In C++ ixxxxx, use a single ip/im numbering that is valid both for pp==0 and pp>0, which have two numbering schemes in Fortran ixxxxx:
-      // for pp==0, Fortran sqm(0:1) has indexes 0,1 as in C++; but for Fortran pp>0, omega(2) has indexes 1,2 and not 0,1
-      // NB: this is only possible in ixxxx, but in oxxxxx two different numbering schemes must be used
-      const int ip = ( 1 + nh ) / 2; // NB: same as in Fortran pp==0, differs from Fortran pp>0, which is (3+nh)/2 because omega(2) has indexes 1,2
-      const int im = ( 1 - nh ) / 2; // NB: same as in Fortran pp==0, differs from Fortran pp>0, which is (3-nh)/2 because omega(2) has indexes 1,2
+      // The Fortran ip/im indices ( 1 + nh ) / 2 and ( 1 - nh ) / 2 (for pp > 0, ( 3 +- nh ) / 2 as omega(2) has
+      // indexes 1,2) are the truth values tp and tm: an array element x[ip] is tm * x[0] + tp * x[1], and
+      // ip * x[ip] is just tp * x[1] (and likewise for im)
 #ifndef MGONGPU_CPPSIMD
       if( pp == 0. )
       {
@@ -244,10 +247,10 @@
         fptype_amp sqm[2] = { fpsqrt( std::abs( fmass ) ), 0. }; // possibility of negative fermion masses
         //sqm[1] = ( fmass < 0. ? -abs( sqm[0] ) : abs( sqm[0] ) ); // AV: why abs here?
         sqm[1] = ( fmass < 0. ? -sqm[0] : sqm[0] ); // AV: removed an abs here
-        w[0] = cxmake( ip * sqm[ip], 0 );
-        w[1] = cxmake( im * nsf * sqm[ip], 0 );
-        w[2] = cxmake( ip * nsf * sqm[im], 0 );
-        w[3] = cxmake( im * sqm[im], 0 );
+        w[0] = cxmake( tp * sqm[1], 0 );
+        w[1] = cxmake( tm * nsf * sqm[0], 0 );
+        w[2] = cxmake( tp * nsf * sqm[0], 0 );
+        w[3] = cxmake( tm * sqm[1], 0 );
       }
       else
       {
@@ -255,30 +258,32 @@
                                fptype_amp( 1 + nsf - ( 1 - nsf ) * nh ) * (fptype_amp)0.5 };
         fptype_amp omega[2] = { fpsqrt( pvec0 + pp ), 0. };
         omega[1] = fmass / omega[0];
-        const fptype_amp sfomega[2] = { sf[0] * omega[ip], sf[1] * omega[im] };
+        const fptype_amp sfomega[2] = { sf[0] * ( tm * omega[0] + tp * omega[1] ), sf[1] * ( tp * omega[0] + tm * omega[1] ) };
         const fptype_amp pp3 = fpmax( fpternary( fpsignbit( pvec3 ), ( pvec1 * pvec1 + pvec2 * pvec2 ) / ( pp - pvec3 ), pp + pvec3 ), 0. );
         const cxtype_amp chi[2] = { cxmake( fpsqrt( pp3 * (fptype_amp)0.5 / pp ), 0. ),
                                 ( pp3 == 0. ? cxmake( -nh, 0. ) : cxmake( nh * pvec1, pvec2 ) / fpsqrt( (fptype_amp)2. * pp * pp3 ) ) };
-        w[0] = sfomega[0] * chi[im];
-        w[1] = sfomega[0] * chi[ip];
-        w[2] = sfomega[1] * chi[im];
-        w[3] = sfomega[1] * chi[ip];
+        const cxtype_amp chi_im = tp * chi[0] + tm * chi[1]; // chi[im]
+        const cxtype_amp chi_ip = tm * chi[0] + tp * chi[1]; // chi[ip]
+        w[0] = sfomega[0] * chi_im;
+        w[1] = sfomega[0] * chi_ip;
+        w[2] = sfomega[1] * chi_im;
+        w[3] = sfomega[1] * chi_ip;
       }
 #else
       // Branch A: pp == 0.
       // NB: Do not use "abs" for floats! It returns an integer with no build warning! Use std::abs!
       fptype_amp sqm[2] = { fpsqrt( std::abs( fmass ) ), 0 }; // possibility of negative fermion masses (NB: SCALAR!)
       sqm[1] = ( fmass < 0 ? -sqm[0] : sqm[0] );          // AV: removed an abs here (as above)
-      const cxtype_amp fiA_2 = ip * sqm[ip];                  // scalar cxtype: real part initialised from fptype, imag part = 0
-      const cxtype_amp fiA_3 = im * nsf * sqm[ip];            // scalar cxtype: real part initialised from fptype, imag part = 0
-      const cxtype_amp fiA_4 = ip * nsf * sqm[im];            // scalar cxtype: real part initialised from fptype, imag part = 0
-      const cxtype_amp fiA_5 = im * sqm[im];                  // scalar cxtype: real part initialised from fptype, imag part = 0
+      const cxtype_amp_v fiA_2 = cxmake( tp * sqm[1], 0 );
+      const cxtype_amp_v fiA_3 = cxmake( tm * (fptype_amp)nsf * sqm[0], 0 );
+      const cxtype_amp_v fiA_4 = cxmake( tp * (fptype_amp)nsf * sqm[0], 0 );
+      const cxtype_amp_v fiA_5 = cxmake( tm * sqm[1], 0 );
       // Branch B: pp != 0.
-      const fptype_amp sf[2] = { fptype_amp( 1 + nsf + ( 1 - nsf ) * nh ) * (fptype_amp)0.5,
-                             fptype_amp( 1 + nsf - ( 1 - nsf ) * nh ) * (fptype_amp)0.5 };
+      const fptype_amp_v sf[2] = { ( (fptype_amp)( 1 + nsf ) + (fptype_amp)( 1 - nsf ) * nh ) * (fptype_amp)0.5,
+                                   ( (fptype_amp)( 1 + nsf ) - (fptype_amp)( 1 - nsf ) * nh ) * (fptype_amp)0.5 };
       fptype_v omega[2] = { fpsqrt( pvec0 + pp ), 0 };
       omega[1] = fmass / omega[0];
-      const fptype_amp_v sfomega[2] = { sf[0] * omega[ip], sf[1] * omega[im] };
+      const fptype_amp_v sfomega[2] = { sf[0] * ( tm * omega[0] + tp * omega[1] ), sf[1] * ( tp * omega[0] + tm * omega[1] ) };
       const fptype_amp_v pp3 = fpmax( fpternary( fpsignbit( pvec3 ), ( pvec1 * pvec1 + pvec2 * pvec2 ) / ( pp - pvec3 ), pp + pvec3 ), 0 );
       volatile fptype_amp_v ppDENOM = fpternary( pp != 0, pp, 1. );    // hack: ppDENOM[ieppV]=1 if pp[ieppV]==0
       volatile fptype_amp_v pp3DENOM = fpternary( pp3 != 0, pp3, 1. ); // hack: pp3DENOM[ieppV]=1 if pp3[ieppV]==0
@@ -286,11 +291,13 @@
       const cxtype_amp_v chi[2] = { cxmake( fpsqrt( chi0r2 ), 0 ),     // hack: dummy[ieppV] is not used if pp[ieppV]==0
                                 cxternary( ( pp3 == 0. ),
                                            cxmake( -nh, 0 ),
-                                           cxmake( (fptype_amp)nh * pvec1, pvec2 ) / fpsqrt( (fptype_amp)2. * ppDENOM * pp3DENOM ) ) }; // hack: dummy[ieppV] is not used if pp[ieppV]==0
-      const cxtype_amp_v fiB_2 = sfomega[0] * chi[im];
-      const cxtype_amp_v fiB_3 = sfomega[0] * chi[ip];
-      const cxtype_amp_v fiB_4 = sfomega[1] * chi[im];
-      const cxtype_amp_v fiB_5 = sfomega[1] * chi[ip];
+                                           cxmake( nh * pvec1, pvec2 ) / fpsqrt( (fptype_amp)2. * ppDENOM * pp3DENOM ) ) }; // hack: dummy[ieppV] is not used if pp[ieppV]==0
+      const cxtype_amp_v chi_im = tp * chi[0] + tm * chi[1]; // chi[im]
+      const cxtype_amp_v chi_ip = tm * chi[0] + tp * chi[1]; // chi[ip]
+      const cxtype_amp_v fiB_2 = sfomega[0] * chi_im;
+      const cxtype_amp_v fiB_3 = sfomega[0] * chi_ip;
+      const cxtype_amp_v fiB_4 = sfomega[1] * chi_im;
+      const cxtype_amp_v fiB_5 = sfomega[1] * chi_ip;
       // Choose between the results from branch A and branch B
       const bool_v mask = ( pp == 0. );
       w[0] = cxternary( mask, fiA_2, fiB_2 );
@@ -316,8 +323,8 @@
       volatile fptype_amp_sv sqp0p3DENOM = fpternary( sqp0p3 != 0, (fptype_sv)sqp0p3, 1. ); // hack: dummy sqp0p3DENOM[ieppV]=1 if sqp0p3[ieppV]==0
       cxtype_sv chi[2] = { cxmake( (fptype_v)sqp0p3, 0. ),
                            cxternary( sqp0p3 == 0,
-                                      cxmake( -(fptype_amp)nhel * fpsqrt( (fptype_amp)2. * pvec0 ), 0. ),
-                                      cxmake( (fptype_amp)nh * pvec1, pvec2 ) / (const fptype_v)sqp0p3DENOM ) }; // hack: dummy[ieppV] is not used if sqp0p3[ieppV]==0
+                                      cxmake( -nhel * fpsqrt( (fptype_amp)2. * pvec0 ), 0. ),
+                                      cxmake( nh * pvec1, pvec2 ) / (const fptype_v)sqp0p3DENOM ) }; // hack: dummy[ieppV] is not used if sqp0p3[ieppV]==0
 #else
       // pvec0+pvec3 is a cancelling difference of two ~|p| numbers for a
       // backward-moving massless fermion, and everything below divides by it.
@@ -330,22 +337,12 @@
                                           fptype_sv{ 0 },
                                           fpsqrt( p0p3stable ) * (fptype_amp)nsf );
       const cxtype_amp_sv chi[2] = { cxmake( sqp0p3, 0. ),
-                                 ( sqp0p3 == 0. ? cxmake( -(fptype_amp)nhel * fpsqrt( (fptype_amp)2. * pvec0 ), 0. ) : cxmake( (fptype_amp)nh * pvec1, pvec2 ) / sqp0p3 ) };
+                                 ( sqp0p3 == 0. ? cxmake( -nhel * fpsqrt( (fptype_amp)2. * pvec0 ), 0. ) : cxmake( nh * pvec1, pvec2 ) / sqp0p3 ) };
 #endif
-      if( nh == 1 )
-      {
-        w[0] = cxzero_sv();
-        w[1] = cxzero_sv();
-        w[2] = chi[0];
-        w[3] = chi[1];
-      }
-      else
-      {
-        w[0] = chi[1];
-        w[1] = chi[0];
-        w[2] = cxzero_sv();
-        w[3] = cxzero_sv();
-      }
+      w[0] = tm * chi[1];
+      w[1] = tm * chi[0];
+      w[2] = tp * chi[0];
+      w[3] = tp * chi[1];
     }
     mgDebug( 1, __FUNCTION__ );
     return;
@@ -359,7 +356,7 @@
   __host__ __device__ void
   ipzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavour
           ALOHAOBJ & fi,          // output: wavefunctions
@@ -374,20 +371,17 @@
     fi.pvec[2] = fptype_momenta_sv{ 0 };
     fi.pvec[3] = -pvec3_ * static_cast<fptype_momenta>(nsf);
     fi.flv_index = flv;
-    const int nh = nhel * nsf;
+    const fptype_amp_sv nh = nhel * static_cast<fptype_amp>( nsf );
+    // BRANCHLESS HELICITY: the helicity is not a branch condition but enters as the truth values
+    // tp (nh == +1) and tm (nh == -1), 1 or 0, which multiply the two alternatives. Different
+    // SIMD lanes or GPU threads can then carry different helicities and still run in lock-step.
+    const fptype_amp_sv tp = ( 1 + nh ) * (fptype_amp)0.5;
+    const fptype_amp_sv tm = 1 - tp;
     const cxtype_amp_sv sqp0p3 = cxmake( fpsqrt( (fptype_amp)2. * pvec3 ) * (fptype_amp)nsf, 0. );
-    w[0] = cxmake( fpamp_of_mom( fi.pvec[1] ), fpamp_of_mom( fi.pvec[2] ) );
-    if( nh == 1 )
-    {
-      w[1] = cxmake( fpamp_of_mom( fi.pvec[1] ), fpamp_of_mom( fi.pvec[2] ) );
-      w[2] = sqp0p3;
-    }
-    else
-    {
-      w[1] = sqp0p3;
-      w[2] = cxmake( fpamp_of_mom( fi.pvec[1] ), fpamp_of_mom( fi.pvec[2] ) );
-    }
-    w[3] = cxmake( fpamp_of_mom( fi.pvec[1] ), fpamp_of_mom( fi.pvec[2] ) );
+    w[0] = cxzero_sv();
+    w[1] = tm * sqp0p3;
+    w[2] = tp * sqp0p3;
+    w[3] = cxzero_sv();
     mgDebug( 1, __FUNCTION__ );
     return;
   }
@@ -400,7 +394,7 @@
   __host__ __device__ void
   imzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavour
           ALOHAOBJ & fi,          // output: wavefunctions
@@ -415,20 +409,17 @@
     fi.pvec[2] = fptype_momenta_sv{ 0 };
     fi.pvec[3] = -pvec3_ * static_cast<fptype_momenta>(nsf);
     fi.flv_index = flv;
-    const int nh = nhel * nsf;
-    const cxtype_amp_sv chi = cxmake( -(fptype_amp)nhel * fpsqrt( -(fptype_amp)2. * pvec3 ), 0. );
+    const fptype_amp_sv nh = nhel * static_cast<fptype_amp>( nsf );
+    // BRANCHLESS HELICITY: the helicity is not a branch condition but enters as the truth values
+    // tp (nh == +1) and tm (nh == -1), 1 or 0, which multiply the two alternatives. Different
+    // SIMD lanes or GPU threads can then carry different helicities and still run in lock-step.
+    const fptype_amp_sv tp = ( 1 + nh ) * (fptype_amp)0.5;
+    const fptype_amp_sv tm = 1 - tp;
+    const cxtype_amp_sv chi = cxmake( -nhel * fpsqrt( -(fptype_amp)2. * pvec3 ), 0. );
+    w[0] = tm * chi;
     w[1] = cxzero_sv();
     w[2] = cxzero_sv();
-    if( nh == 1 )
-    {
-      w[0] = cxzero_sv();
-      w[3] = chi;
-    }
-    else
-    {
-      w[0] = chi;
-      w[3] = cxzero_sv();
-    }
+    w[3] = tp * chi;
     mgDebug( 1, __FUNCTION__ );
     return;
   }
@@ -441,7 +432,7 @@
   __host__ __device__ void
   ixzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavour
           ALOHAOBJ & fi,          // output: wavefunctions
@@ -463,7 +454,12 @@
     fi.pvec[3] = -pvec3_ * static_cast<fptype_momenta>(nsf);
     fi.flv_index = flv;
 
-    const int nh = nhel * nsf;
+    const fptype_amp_sv nh = nhel * static_cast<fptype_amp>( nsf );
+    // BRANCHLESS HELICITY: the helicity is not a branch condition but enters as the truth values
+    // tp (nh == +1) and tm (nh == -1), 1 or 0, which multiply the two alternatives. Different
+    // SIMD lanes or GPU threads can then carry different helicities and still run in lock-step.
+    const fptype_amp_sv tp = ( 1 + nh ) * (fptype_amp)0.5;
+    const fptype_amp_sv tm = 1 - tp;
     //const float sqp0p3 = sqrtf( pvec0 + pvec3 ) * nsf; // AV: why force a float here?
     // pvec0+pvec3 is a cancelling difference of two ~|p| numbers for a
     // backward-moving massless fermion, and everything below divides by it.
@@ -475,21 +471,11 @@
                                             pvec0 + pvec3 );
     const fptype_amp_sv sqp0p3 = fpsqrt( p0p3stable ) * static_cast<fptype_amp>(nsf);
     const cxtype_amp_sv chi0 = cxmake( sqp0p3, 0. );
-    const cxtype_amp_sv chi1 = cxmake( (fptype_amp)nh * pvec1 / sqp0p3, pvec2 / sqp0p3 );
-    if( nh == 1 )
-    {
-      w[0] = cxzero_sv();
-      w[1] = cxzero_sv();
-      w[2] = chi0;
-      w[3] = chi1;
-    }
-    else
-    {
-      w[0] = chi1;
-      w[1] = chi0;
-      w[2] = cxzero_sv();
-      w[3] = cxzero_sv();
-    }
+    const cxtype_amp_sv chi1 = cxmake( nh * pvec1 / sqp0p3, pvec2 / sqp0p3 );
+    w[0] = tm * chi1;
+    w[1] = tm * chi0;
+    w[2] = tp * chi0;
+    w[3] = tp * chi1;
     mgDebug( 1, __FUNCTION__ );
     return;
   }
@@ -501,7 +487,7 @@
   __host__ __device__ void
   vxxxxx( const fptype_momenta momenta[], // input: momenta
           const fptype_amp vmass,     // input: vector boson mass
-          const int nhel,         // input: -1, 0 (only if vmass!=0) or +1 (helicity of vector boson)
+          const fptype_amp_sv nhel, // input: -1, 0 (only if vmass!=0) or +1 (helicity of vector boson, per SIMD lane/GPU thread)
           const int nsv,          // input: +1 (final) or -1 (initial)
           const int flv,          // input: flavour
           ALOHAOBJ & vc,          // output: wavefunctions
@@ -527,7 +513,10 @@
     vc.pvec[3] = pvec3_ * static_cast<fptype_momenta>(nsv);
     vc.flv_index = flv;
     const fptype_amp sqh = fpsqrt( 0.5 ); // AV this is > 0!
-    const fptype_amp hel = nhel;
+    const fptype_amp_sv hel = nhel;
+    // BRANCHLESS HELICITY: |hel| is hel * hel for hel in { -1, 0, +1 }, so that the helicity can
+    // differ from one SIMD lane or GPU thread to the next with no branch and no std::abs
+    const fptype_amp_sv abshel = hel * hel;
 
     // FD gauge
      const cxtype_amp_sv cI = cxmake( 0. + fptype_amp_sv{ 0. },  1. + fptype_amp_sv{ 0. }  );
@@ -538,8 +527,8 @@
 
     if( vmass != 0. )
     {
-      const int nsvahl = nsv * std::abs( hel );
-      const fptype_amp hel0 = 1. - std::abs( hel );
+      const fptype_amp_sv nsvahl = static_cast<fptype_amp>( nsv ) * abshel;
+      const fptype_amp_sv hel0 = 1 - abshel; // truth value of hel == 0 (longitudinal polarisation)
 #ifndef MGONGPU_CPPSIMD
       const fptype_amp_sv pt2 = ( pvec1 * pvec1 ) + ( pvec2 * pvec2 );
       const fptype_amp_sv pp = fpmin( pvec0, fpsqrt( pt2 + ( pvec3 * pvec3 ) ) );
@@ -594,17 +583,14 @@
 
       nk = n[0]*pvec0 - n[1]*pvec1 - n[2]*pvec2 - n[3]*pvec3;
 
-      if ( abs(nhel) == 1)
-      {
-        w[4] = cxzero_sv<cxtype_amp_sv>();
-      }
-      else{
-        w[0] = cxmake( -vmass/nk * n[0], zero );
-        w[1] = cxmake( -vmass/nk * n[1], zero );
-        w[2] = cxmake( -vmass/nk * n[2], zero );
-        w[3] = cxmake( -vmass/nk * n[3], zero );
-        w[4] = -static_cast<fptype_amp>(nsv)*cI;
-      }
+      // BRANCHLESS HELICITY: |hel| = 1 keeps w[0..3] with w[4] = 0, hel = 0 (hel0 = 1) takes the
+      // FD-gauge longitudinal polarisation instead (nk > 0 for a massive momentum, so the unused
+      // alternative is finite and vanishes when multiplied by its zero truth value)
+      w[0] = abshel * w[0] + hel0 * cxmake( -vmass / nk * n[0], zero );
+      w[1] = abshel * w[1] + hel0 * cxmake( -vmass / nk * n[1], zero );
+      w[2] = abshel * w[2] + hel0 * cxmake( -vmass / nk * n[2], zero );
+      w[3] = abshel * w[3] + hel0 * cxmake( -vmass / nk * n[3], zero );
+      w[4] = hel0 * ( -static_cast<fptype_amp>( nsv ) * cI );
 
 #else
       volatile fptype_amp_sv pt2 = ( pvec1 * pvec1 ) + ( pvec2 * pvec2 );
@@ -613,9 +599,9 @@
       const fptype_amp_sv pt = fpmin( pp, fpsqrt( pt2 ) );
       // Branch A: pp == 0.
       const cxtype_amp vcA_2 = cxmake( 0, 0 );
-      const cxtype_amp vcA_3 = cxmake( -hel * sqh, 0 );
-      const cxtype_amp vcA_4 = cxmake( 0, nsvahl * sqh );
-      const cxtype_amp vcA_5 = cxmake( hel0, 0 );
+      const cxtype_amp_v vcA_3 = cxmake( -hel * sqh, 0 );
+      const cxtype_amp_v vcA_4 = cxmake( 0., nsvahl * sqh );
+      const cxtype_amp_v vcA_5 = cxmake( hel0, 0 );
       // Branch B: pp != 0.
       volatile fptype_amp_v ppDENOM = fpternary( pp != 0, pp, 1. ); // hack: ppDENOM[ieppV]=1 if pp[ieppV]==0
       const fptype_amp_v emp = pvec0 / ( vmass * ppDENOM );         // hack: dummy[ieppV] is not used if pp[ieppV]==0
@@ -624,11 +610,11 @@
       // Branch B1: pp != 0. and pt != 0.
       volatile fptype_amp_v ptDENOM = fpternary( pt != 0, pt, 1. );                                                     // hack: ptDENOM[ieppV]=1 if pt[ieppV]==0
       const fptype_amp_v pzpt = pvec3 / ( ppDENOM * ptDENOM ) * sqh * hel;                                              // hack: dummy[ieppV] is not used if pp[ieppV]==0
-      const cxtype_amp_v vcB1_3 = cxmake( hel0 * pvec1 * emp - pvec1 * pzpt, -fpamp_scalar(nsvahl) * pvec2 / ptDENOM * sqh ); // hack: dummy[ieppV] is not used if pt[ieppV]==0
-      const cxtype_amp_v vcB1_4 = cxmake( hel0 * pvec2 * emp - pvec2 * pzpt, fpamp_scalar(nsvahl) * pvec1 / ptDENOM * sqh );  // hack: dummy[ieppV] is not used if pt[ieppV]==0
+      const cxtype_amp_v vcB1_3 = cxmake( hel0 * pvec1 * emp - pvec1 * pzpt, -nsvahl * pvec2 / ptDENOM * sqh ); // hack: dummy[ieppV] is not used if pt[ieppV]==0
+      const cxtype_amp_v vcB1_4 = cxmake( hel0 * pvec2 * emp - pvec2 * pzpt, nsvahl * pvec1 / ptDENOM * sqh );  // hack: dummy[ieppV] is not used if pt[ieppV]==0
       // Branch B2: pp != 0. and pt == 0.
-      const cxtype_amp vcB2_3 = cxmake( -hel * sqh, 0. );
-      const cxtype_amp_v vcB2_4 = cxmake( 0., fpamp_scalar(nsvahl) * fpternary( ( pvec3 < 0 ), -sqh, sqh ) ); // AV: removed an abs here
+      const cxtype_amp_v vcB2_3 = cxmake( -hel * sqh, 0. );
+      const cxtype_amp_v vcB2_4 = cxmake( 0., nsvahl * fpternary( ( pvec3 < 0 ), -sqh, sqh ) ); // AV: removed an abs here
       // Choose between the results from branch A and branch B (and from branch B1 and branch B2)
       const bool_v mask = ( pp == 0. );
       const bool_v maskB = ( pt != 0. );
@@ -650,20 +636,14 @@
 
       nk = n[0]*pvec0 - n[1]*pvec1 - n[2]*pvec2 - n[3]*pvec3;
 
-      // nhel is a scalar: no need for a per-lane mask here (and a bool_v built
-      // from a single value would only set the first lane)
-      if ( abs(nhel) == 1 )
-      {
-        w[4] = zero;
-      }
-      else
-      {
-        w[0] = cxmake( -vmass/nk * n[0], zero );
-        w[1] = cxmake( -vmass/nk * n[1], zero );
-        w[2] = cxmake( -vmass/nk * n[2], zero );
-        w[3] = cxmake( -vmass/nk * n[3], zero );
-        w[4] = -static_cast<fptype_amp>(nsv)*cI;
-      }
+      // BRANCHLESS HELICITY: |hel| = 1 keeps w[0..3] with w[4] = 0, hel = 0 (hel0 = 1) takes the
+      // FD-gauge longitudinal polarisation instead (nk > 0 for a massive momentum, so the unused
+      // alternative is finite and vanishes when multiplied by its zero truth value)
+      w[0] = abshel * w[0] + hel0 * cxmake( -vmass / nk * n[0], zero );
+      w[1] = abshel * w[1] + hel0 * cxmake( -vmass / nk * n[1], zero );
+      w[2] = abshel * w[2] + hel0 * cxmake( -vmass / nk * n[2], zero );
+      w[3] = abshel * w[3] + hel0 * cxmake( -vmass / nk * n[3], zero );
+      w[4] = hel0 * ( -static_cast<fptype_amp>( nsv ) * cI );
 #endif
     }
     else
@@ -695,11 +675,11 @@
       // Branch A: pt != 0.
       volatile fptype_amp_v ptDENOM = fpternary( pt != 0, pt, 1. );                             // hack: ptDENOM[ieppV]=1 if pt[ieppV]==0
       const fptype_amp_v pzpt = pvec3 / ( pp * ptDENOM ) * sqh * hel;                           // hack: dummy[ieppV] is not used if pt[ieppV]==0
-      const cxtype_amp_v vcA_3 = cxmake( -pvec1 * pzpt, -fpamp_scalar(nsv) * pvec2 / ptDENOM * sqh ); // hack: dummy[ieppV] is not used if pt[ieppV]==0
-      const cxtype_amp_v vcA_4 = cxmake( -pvec2 * pzpt, fpamp_scalar(nsv) * pvec1 / ptDENOM * sqh );  // hack: dummy[ieppV] is not used if pt[ieppV]==0
+      const cxtype_amp_v vcA_3 = cxmake( -pvec1 * pzpt, -(fptype_amp)nsv * pvec2 / ptDENOM * sqh ); // hack: dummy[ieppV] is not used if pt[ieppV]==0
+      const cxtype_amp_v vcA_4 = cxmake( -pvec2 * pzpt, (fptype_amp)nsv * pvec1 / ptDENOM * sqh );  // hack: dummy[ieppV] is not used if pt[ieppV]==0
       // Branch B: pt == 0.
-      const cxtype_amp vcB_3 = cxmake( -(fptype_amp)hel * sqh, 0 );
-      const cxtype_amp_v vcB_4 = cxmake( 0, fpamp_scalar(nsv) * fpternary( ( pvec3 < 0 ), -sqh, sqh ) ); // AV: removed an abs here
+      const cxtype_amp_v vcB_3 = cxmake( -hel * sqh, 0 );
+      const cxtype_amp_v vcB_4 = cxmake( 0, (fptype_amp)nsv * fpternary( ( pvec3 < 0 ), -sqh, sqh ) ); // AV: removed an abs here
       // Choose between the results from branch A and branch B
       const bool_v mask = ( pt != 0. );
       w[1] = cxternary( mask, vcA_3, vcB_3 );
@@ -758,7 +738,7 @@
   __host__ __device__ void
   oxxxxx( const fptype_momenta momenta[], // input: momenta
           const fptype_amp fmass,     // input: fermion mass
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           int flv,                // input: flavour
           ALOHAOBJ & fo,          // output: wavefunctions
@@ -784,7 +764,12 @@
     fo.pvec[2] = pvec2_ * static_cast<fptype_momenta>(nsf);
     fo.pvec[3] = pvec3_ * static_cast<fptype_momenta>(nsf);
     fo.flv_index = flv;
-    const int nh = nhel * nsf;
+    const fptype_amp_sv nh = nhel * static_cast<fptype_amp>( nsf );
+    // BRANCHLESS HELICITY: the helicity is not a branch condition but enters as the truth values
+    // tp (nh == +1) and tm (nh == -1), 1 or 0, which multiply the two alternatives. Different
+    // SIMD lanes or GPU threads can then carry different helicities and still run in lock-step.
+    const fptype_amp_sv tp = ( 1 + nh ) * (fptype_amp)0.5;
+    const fptype_amp_sv tm = 1 - tp;
     if( fmass != 0. )
     {
 #ifndef MGONGPU_CPPSIMD
@@ -795,12 +780,12 @@
         fptype_amp sqm[2] = { fpsqrt( std::abs( fmass ) ), 0. }; // possibility of negative fermion masses
         //sqm[1] = ( fmass < 0. ? -abs( sqm[0] ) : abs( sqm[0] ) ); // AV: why abs here?
         sqm[1] = ( fmass < 0. ? -sqm[0] : sqm[0] ); // AV: removed an abs here
-        const int ip = -( ( 1 - nh ) / 2 ) * nhel;  // NB: Fortran sqm(0:1) also has indexes 0,1 as in C++
-        const int im = ( 1 + nh ) / 2 * nhel;       // NB: Fortran sqm(0:1) also has indexes 0,1 as in C++
-        w[0] = cxmake( im * sqm[std::abs( ip )], 0 );
-        w[1] = cxmake( ip * nsf * sqm[std::abs( ip )], 0 );
-        w[2] = cxmake( im * nsf * sqm[std::abs( im )], 0 );
-        w[3] = cxmake( ip * sqm[std::abs( im )], 0 );
+        // The Fortran ip = -( ( 1 - nh ) / 2 ) * nhel and im = ( 1 + nh ) / 2 * nhel are -tm * nhel and
+        // tp * nhel, so that |ip| = tm and |im| = tp: sqm[|ip|] is sqm[1] wherever ip != 0, and so on
+        w[0] = cxmake( tp * nhel * sqm[0], 0 );
+        w[1] = cxmake( -tm * nhel * nsf * sqm[1], 0 );
+        w[2] = cxmake( tp * nhel * nsf * sqm[1], 0 );
+        w[3] = cxmake( -tm * nhel * sqm[0], 0 );
       }
       else
       {
@@ -808,17 +793,18 @@
                                fptype_amp( 1 + nsf - ( 1 - nsf ) * nh ) * (fptype_amp)0.5 };
         fptype_amp omega[2] = { fpsqrt( pvec0 + pp ), 0. };
         omega[1] = fmass / omega[0];
-        const int ip = ( 1 + nh ) / 2; // NB: Fortran is (3+nh)/2 because omega(2) has indexes 1,2 and not 0,1
-        const int im = ( 1 - nh ) / 2; // NB: Fortran is (3-nh)/2 because omega(2) has indexes 1,2 and not 0,1
-        const fptype_amp sfomeg[2] = { sf[0] * omega[ip], sf[1] * omega[im] };
+        // The Fortran ip/im indices ( 3 +- nh ) / 2 (omega(2) has indexes 1,2) are the truth values tp/tm
+        const fptype_amp sfomeg[2] = { sf[0] * ( tm * omega[0] + tp * omega[1] ), sf[1] * ( tp * omega[0] + tm * omega[1] ) };
         const fptype_amp pp3 = fpmax( fpternary( fpsignbit( pvec3 ), ( pvec1 * pvec1 + pvec2 * pvec2 ) / ( pp - pvec3 ), pp + pvec3 ), 0. );
         const cxtype_amp chi[2] = { cxmake( fpsqrt( pp3 * (fptype_amp)0.5 / pp ), 0. ),
                                 ( ( pp3 == 0. ) ? cxmake( -nh, 0. )
                                                 : cxmake( nh * pvec1, -pvec2 ) / fpsqrt( (fptype_amp)2. * pp * pp3 ) ) };
-        w[0] = sfomeg[1] * chi[im];
-        w[1] = sfomeg[1] * chi[ip];
-        w[2] = sfomeg[0] * chi[im];
-        w[3] = sfomeg[0] * chi[ip];
+        const cxtype_amp chi_im = tp * chi[0] + tm * chi[1]; // chi[im]
+        const cxtype_amp chi_ip = tm * chi[0] + tp * chi[1]; // chi[ip]
+        w[0] = sfomeg[1] * chi_im;
+        w[1] = sfomeg[1] * chi_ip;
+        w[2] = sfomeg[0] * chi_im;
+        w[3] = sfomeg[0] * chi_ip;
       }
 #else
       volatile fptype_amp_sv p2 = pvec1 * pvec1 + pvec2 * pvec2 + pvec3 * pvec3; // volatile fixes #736
@@ -827,20 +813,17 @@
       // NB: Do not use "abs" for floats! It returns an integer with no build warning! Use std::abs!
       fptype_amp sqm[2] = { fpsqrt( std::abs( fmass ) ), 0 }; // possibility of negative fermion masses
       sqm[1] = ( fmass < 0 ? -sqm[0] : sqm[0] );          // AV: removed an abs here (as above)
-      const int ipA = -( ( 1 - nh ) / 2 ) * nhel;
-      const int imA = ( 1 + nh ) / 2 * nhel;
-      const cxtype_amp foA_2 = imA * sqm[std::abs( ipA )];
-      const cxtype_amp foA_3 = ipA * nsf * sqm[std::abs( ipA )];
-      const cxtype_amp foA_4 = imA * nsf * sqm[std::abs( imA )];
-      const cxtype_amp foA_5 = ipA * sqm[std::abs( imA )];
+      // (see the scalar code above for the truth values tp/tm in place of the Fortran ip/im indices)
+      const cxtype_amp_v foA_2 = cxmake( tp * nhel * sqm[0], 0 );
+      const cxtype_amp_v foA_3 = cxmake( -tm * nhel * (fptype_amp)nsf * sqm[1], 0 );
+      const cxtype_amp_v foA_4 = cxmake( tp * nhel * (fptype_amp)nsf * sqm[1], 0 );
+      const cxtype_amp_v foA_5 = cxmake( -tm * nhel * sqm[0], 0 );
       // Branch B: pp != 0.
-      const fptype_amp sf[2] = { fptype_amp( 1 + nsf + ( 1 - nsf ) * nh ) * (fptype_amp)0.5,
-                             fptype_amp( 1 + nsf - ( 1 - nsf ) * nh ) * (fptype_amp)0.5 };
+      const fptype_amp_v sf[2] = { ( (fptype_amp)( 1 + nsf ) + (fptype_amp)( 1 - nsf ) * nh ) * (fptype_amp)0.5,
+                                   ( (fptype_amp)( 1 + nsf ) - (fptype_amp)( 1 - nsf ) * nh ) * (fptype_amp)0.5 };
       fptype_v omega[2] = { fpsqrt( pvec0 + pp ), 0 };
       omega[1] = fmass / omega[0];
-      const int ipB = ( 1 + nh ) / 2;
-      const int imB = ( 1 - nh ) / 2;
-      const fptype_amp_v sfomeg[2] = { sf[0] * omega[ipB], sf[1] * omega[imB] };
+      const fptype_amp_v sfomeg[2] = { sf[0] * ( tm * omega[0] + tp * omega[1] ), sf[1] * ( tp * omega[0] + tm * omega[1] ) };
       const fptype_amp_v pp3 = fpmax( fpternary( fpsignbit( pvec3 ), ( pvec1 * pvec1 + pvec2 * pvec2 ) / ( pp - pvec3 ), pp + pvec3 ), 0. );
       volatile fptype_amp_v ppDENOM = fpternary( pp != 0, pp, 1. );    // hack: ppDENOM[ieppV]=1 if pp[ieppV]==0
       volatile fptype_amp_v pp3DENOM = fpternary( pp3 != 0, pp3, 1. ); // hack: pp3DENOM[ieppV]=1 if pp3[ieppV]==0
@@ -848,11 +831,13 @@
       const cxtype_amp_v chi[2] = { cxmake( fpsqrt( chi0r2 ), 0. ),    // hack: dummy[ieppV] is not used if pp[ieppV]==0
                                 ( cxternary( ( pp3 == 0. ),
                                              cxmake( -nh, 0. ),
-                                             cxmake( (fptype_amp)nh * pvec1, -pvec2 ) / fpsqrt( (fptype_amp)2. * ppDENOM * pp3DENOM ) ) ) }; // hack: dummy[ieppV] is not used if pp[ieppV]==0
-      const cxtype_amp_v foB_2 = sfomeg[1] * chi[imB];
-      const cxtype_amp_v foB_3 = sfomeg[1] * chi[ipB];
-      const cxtype_amp_v foB_4 = sfomeg[0] * chi[imB];
-      const cxtype_amp_v foB_5 = sfomeg[0] * chi[ipB];
+                                             cxmake( nh * pvec1, -pvec2 ) / fpsqrt( (fptype_amp)2. * ppDENOM * pp3DENOM ) ) ) }; // hack: dummy[ieppV] is not used if pp[ieppV]==0
+      const cxtype_amp_v chi_im = tp * chi[0] + tm * chi[1]; // chi[im]
+      const cxtype_amp_v chi_ip = tm * chi[0] + tp * chi[1]; // chi[ip]
+      const cxtype_amp_v foB_2 = sfomeg[1] * chi_im;
+      const cxtype_amp_v foB_3 = sfomeg[1] * chi_ip;
+      const cxtype_amp_v foB_4 = sfomeg[0] * chi_im;
+      const cxtype_amp_v foB_5 = sfomeg[0] * chi_ip;
       // Choose between the results from branch A and branch B
       const bool_v mask = ( pp == 0. );
       w[0] = cxternary( mask, foA_2, foB_2 );
@@ -879,7 +864,7 @@
       const cxtype_amp_v chi[2] = { cxmake( (fptype_v)sqp0p3, 0. ),
                                 cxternary( ( sqp0p3 == 0. ),
                                            cxmake( -nhel, 0. ) * fpsqrt( (fptype_amp)2. * pvec0 ),
-                                           cxmake( (fptype_amp)nh * pvec1, -pvec2 ) / (const fptype_amp_sv)sqp0p3DENOM ) }; // hack: dummy[ieppV] is not used if sqp0p3[ieppV]==0
+                                           cxmake( nh * pvec1, -pvec2 ) / (const fptype_amp_sv)sqp0p3DENOM ) }; // hack: dummy[ieppV] is not used if sqp0p3[ieppV]==0
 #else
       // pvec0+pvec3 is a cancelling difference of two ~|p| numbers for a
       // backward-moving massless fermion, and everything below divides by it.
@@ -892,22 +877,12 @@
                                           fptype_sv{ 0 },
                                           fpsqrt( p0p3stable ) * (fptype_amp)nsf );
       const cxtype_amp_sv chi[2] = { cxmake( sqp0p3, 0. ),
-                                 ( sqp0p3 == 0. ? cxmake( -nhel, 0. ) * fpsqrt( (fptype_amp)2. * pvec0 ) : cxmake( (fptype_amp)nh * pvec1, -pvec2 ) / sqp0p3 ) };
+                                 ( sqp0p3 == 0. ? cxmake( -nhel, 0. ) * fpsqrt( (fptype_amp)2. * pvec0 ) : cxmake( nh * pvec1, -pvec2 ) / sqp0p3 ) };
 #endif
-      if( nh == 1 )
-      {
-        w[0] = chi[0];
-        w[1] = chi[1];
-        w[2] = cxzero_sv();
-        w[3] = cxzero_sv();
-      }
-      else
-      {
-        w[0] = cxzero_sv();
-        w[1] = cxzero_sv();
-        w[2] = chi[1];
-        w[3] = chi[0];
-      }
+      w[0] = tp * chi[0];
+      w[1] = tp * chi[1];
+      w[2] = tm * chi[1];
+      w[3] = tm * chi[0];
     }
     mgDebug( 1, __FUNCTION__ );
     return;
@@ -921,7 +896,7 @@
   __host__ __device__ void
   opzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavour
           ALOHAOBJ & fo,          // output: wavefunctions
@@ -936,20 +911,17 @@
     fo.pvec[2] = fptype_momenta_sv{ 0 };
     fo.pvec[3] = pvec3_ * static_cast<fptype_momenta>(nsf);
     fo.flv_index = flv;
-    const int nh = nhel * nsf;
+    const fptype_amp_sv nh = nhel * static_cast<fptype_amp>( nsf );
+    // BRANCHLESS HELICITY: the helicity is not a branch condition but enters as the truth values
+    // tp (nh == +1) and tm (nh == -1), 1 or 0, which multiply the two alternatives. Different
+    // SIMD lanes or GPU threads can then carry different helicities and still run in lock-step.
+    const fptype_amp_sv tp = ( 1 + nh ) * (fptype_amp)0.5;
+    const fptype_amp_sv tm = 1 - tp;
     const cxtype_amp_sv csqp0p3 = cxmake( fpsqrt( (fptype_amp)2. * pvec3 ) * (fptype_amp)nsf, 0. );
+    w[0] = tp * csqp0p3;
     w[1] = cxzero_sv();
     w[2] = cxzero_sv();
-    if( nh == 1 )
-    {
-      w[0] = csqp0p3;
-      w[3] = cxzero_sv();
-    }
-    else
-    {
-      w[0] = cxzero_sv();
-      w[3] = csqp0p3;
-    }
+    w[3] = tm * csqp0p3;
     mgDebug( 1, __FUNCTION__ );
     return;
   }
@@ -962,7 +934,7 @@
   __host__ __device__ void
   omzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavour
           ALOHAOBJ & fo,          // output: wavefunctions
@@ -977,23 +949,17 @@
     fo.pvec[2] = fptype_momenta_sv{ 0 };
     fo.pvec[3] = pvec3_ * static_cast<fptype_momenta>(nsf);
     fo.flv_index = flv;
-    const int nh = nhel * nsf;
+    const fptype_amp_sv nh = nhel * static_cast<fptype_amp>( nsf );
+    // BRANCHLESS HELICITY: the helicity is not a branch condition but enters as the truth values
+    // tp (nh == +1) and tm (nh == -1), 1 or 0, which multiply the two alternatives. Different
+    // SIMD lanes or GPU threads can then carry different helicities and still run in lock-step.
+    const fptype_amp_sv tp = ( 1 + nh ) * (fptype_amp)0.5;
+    const fptype_amp_sv tm = 1 - tp;
     const cxtype_amp_sv chi1 = cxmake( -nhel, 0. ) * fpsqrt( -(fptype_amp)2. * pvec3 );
-    if( nh == 1 )
-    {
-      w[0] = cxzero_sv();
-      w[1] = chi1;
-      w[2] = cxzero_sv();
-      w[3] = cxzero_sv();
-    }
-    else
-    {
-      w[0] = cxzero_sv();
-      w[1] = cxzero_sv();
-      w[2] = chi1;
-      //w[3] = chi1; // AV: BUG!
-      w[3] = cxzero_sv(); // AV: BUG FIX
-    }
+    w[0] = cxzero_sv();
+    w[1] = tp * chi1;
+    w[2] = tm * chi1;
+    w[3] = cxzero_sv();
     mgDebug( 1, __FUNCTION__ );
     return;
   }
@@ -1006,7 +972,7 @@
   __host__ __device__ void
   oxzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavour
           ALOHAOBJ & fo,          // output: wavefunctions
@@ -1028,7 +994,12 @@
     fo.pvec[2] = pvec2_ * static_cast<fptype_momenta>(nsf);
     fo.pvec[3] = pvec3_ * static_cast<fptype_momenta>(nsf);
     fo.flv_index = flv;
-    const int nh = nhel * nsf;
+    const fptype_amp_sv nh = nhel * static_cast<fptype_amp>( nsf );
+    // BRANCHLESS HELICITY: the helicity is not a branch condition but enters as the truth values
+    // tp (nh == +1) and tm (nh == -1), 1 or 0, which multiply the two alternatives. Different
+    // SIMD lanes or GPU threads can then carry different helicities and still run in lock-step.
+    const fptype_amp_sv tp = ( 1 + nh ) * (fptype_amp)0.5;
+    const fptype_amp_sv tm = 1 - tp;
     //const float sqp0p3 = sqrtf( pvec0 + pvec3 ) * nsf; // AV: why force a float here?
     // pvec0+pvec3 is a cancelling difference of two ~|p| numbers for a
     // backward-moving massless fermion, and everything below divides by it.
@@ -1040,21 +1011,11 @@
                                             pvec0 + pvec3 );
     const fptype_amp_sv sqp0p3 = fpsqrt( p0p3stable ) * static_cast<fptype_amp>(nsf);
     const cxtype_amp_sv chi0 = cxmake( sqp0p3, 0. );
-    const cxtype_amp_sv chi1 = cxmake( (fptype_amp)nh * pvec1 / sqp0p3, -pvec2 / sqp0p3 );
-    if( nh == 1 )
-    {
-      w[0] = chi0;
-      w[1] = chi1;
-      w[2] = cxzero_sv();
-      w[3] = cxzero_sv();
-    }
-    else
-    {
-      w[0] = cxzero_sv();
-      w[1] = cxzero_sv();
-      w[2] = chi1;
-      w[3] = chi0;
-    }
+    const cxtype_amp_sv chi1 = cxmake( nh * pvec1 / sqp0p3, -pvec2 / sqp0p3 );
+    w[0] = tp * chi0;
+    w[1] = tp * chi1;
+    w[2] = tm * chi1;
+    w[3] = tm * chi0;
     mgDebug( 1, __FUNCTION__ );
     return;
   }
diff --git a/docs/source/madspace/examples/flow-training.rst b/docs/source/madspace/examples/flow-training.rst
index 43a58c67ff..fef6814a72 100644
--- a/docs/source/madspace/examples/flow-training.rst
+++ b/docs/source/madspace/examples/flow-training.rst
@@ -40,7 +40,7 @@ The mapping, cuts, matrix element, PDF and scale are exactly the same building b
     mapping = ms.PhaseSpaceMapping(masses, E_CM, mode="propagator", cuts=cuts)
 
     proc_dir = "PROC_ggttg"
-    me_path = os.path.join(proc_dir, "lib", "libmadmatrix_P0_gg_ttxg_scalar.so")
+    me_path = os.path.join(proc_dir, "lib", "libmadmatrix_sm_P0_gg_ttxg_scalar_m.so")
     param_card = os.path.join(proc_dir, "Cards", "param_card.dat")
 
     ctx = ms.default_context()
diff --git a/docs/source/madspace/examples/integrator-madnis.rst b/docs/source/madspace/examples/integrator-madnis.rst
index f5e8306025..3f1a329697 100644
--- a/docs/source/madspace/examples/integrator-madnis.rst
+++ b/docs/source/madspace/examples/integrator-madnis.rst
@@ -38,7 +38,7 @@ random numbers are now PyTorch tensors, since MadNIS trains with PyTorch:
     mapping = ms.PhaseSpaceMapping(masses, E_CM, mode="propagator", cuts=cuts)
 
     proc_dir = "PROC_ggttg"
-    me_path = os.path.join(proc_dir, "lib", "libmadmatrix_P0_gg_ttxg_scalar.so")
+    me_path = os.path.join(proc_dir, "lib", "libmadmatrix_sm_P0_gg_ttxg_scalar_m.so")
     param_card = os.path.join(proc_dir, "Cards", "param_card.dat")
 
     ctx = ms.default_context()
diff --git a/docs/source/madspace/examples/integrator.rst b/docs/source/madspace/examples/integrator.rst
index 4c135b12d7..0b91de83c0 100644
--- a/docs/source/madspace/examples/integrator.rst
+++ b/docs/source/madspace/examples/integrator.rst
@@ -42,7 +42,7 @@ examples, generated with ``output mg7 PROC_ggttg`` and ``make BACKEND=scalar`` b
 .. code-block:: python
 
     proc_dir = "PROC_ggttg"
-    me_path = os.path.join(proc_dir, "lib", "libmadmatrix_P0_gg_ttxg_scalar.so")
+    me_path = os.path.join(proc_dir, "lib", "libmadmatrix_sm_P0_gg_ttxg_scalar_m.so")
     param_card = os.path.join(proc_dir, "Cards", "param_card.dat")
 
     ctx = ms.default_context()
diff --git a/docs/source/madspace/examples/matrix-element.rst b/docs/source/madspace/examples/matrix-element.rst
index b42d06ddce..c4a53d92ef 100644
--- a/docs/source/madspace/examples/matrix-element.rst
+++ b/docs/source/madspace/examples/matrix-element.rst
@@ -37,7 +37,7 @@ Loading the matrix element
     import madspace as ms
 
     proc_dir = "PROC_ggttg"
-    me_path = os.path.join(proc_dir, "lib", "libmadmatrix_P0_gg_ttxg_scalar.so")
+    me_path = os.path.join(proc_dir, "lib", "libmadmatrix_sm_P0_gg_ttxg_scalar_m.so")
     param_card = os.path.join(proc_dir, "Cards", "param_card.dat")
 
     ctx = ms.default_context()
@@ -46,7 +46,9 @@ Loading the matrix element
 
 The library name also appears, alongside every other subprocess of the process, in
 ``PROC_ggttg/SubProcesses/subprocesses.json``, as ``me_path`` with a ``{device}`` placeholder
-for the backend.
+for the backend and an ``{fptype}`` placeholder for the floating-point precision the library
+was compiled with (``FPTYPE``, ``m`` by default). The model name (``sm``) is part of the
+library name too.
 
 :py:meth:`Context.load_matrix_element <madspace.Context.load_matrix_element>` returns a
 :py:class:`MatrixElementApi <madspace.MatrixElementApi>`, which reads the parameters from
diff --git a/madgraph/interface/common_run_interface.py b/madgraph/interface/common_run_interface.py
index 0afa5de747..8bcac2e38f 100755
--- a/madgraph/interface/common_run_interface.py
+++ b/madgraph/interface/common_run_interface.py
@@ -2181,6 +2181,20 @@ def check_multicore(self):
             raise self.ConfigurationError('''Can\'t load Reweight module.
             The variable mg5_path might not be correctly configured.''')
 
+        # the reweighting backend (madgraph/various/reweight_backend.py): this
+        # is the classic module; MadtRex runs from the MG7 output's launcher
+        backend = [a.split('=', 1)[1] for a in line.split()
+                   if a.startswith(('--backend=', '--reweight-backend='))]
+        if backend:
+            line = ' '.join(a for a in line.split()
+                            if not a.startswith(('--backend=', '--reweight-backend=')))
+            if backend[-1].lower() == 'madtrex':
+                raise self.InvalidCmd('--backend=madtrex: MadtRex needs an MG7 output; '
+                                      'this reweighting uses the classic module')
+            if backend[-1].lower() not in ('classic', 'auto'):
+                raise self.InvalidCmd('--backend=%s: expected classic, madtrex or auto'
+                                      % backend[-1])
+
         # plugin option
         plugin = False
         reweight_mode = None
diff --git a/madgraph/interface/madevent_interface.py b/madgraph/interface/madevent_interface.py
index 3d119549aa..f682aff010 100755
--- a/madgraph/interface/madevent_interface.py
+++ b/madgraph/interface/madevent_interface.py
@@ -523,8 +523,8 @@ class AskRun(cmd.ControlSwitch):
     def __init__(self, question, line_args=[], mode=None, force=False,
                                                                   *args, **opt):
         
-        self.check_available_module(opt['mother_interface'].options)
         self.me_dir = opt['mother_interface'].me_dir
+        self.check_available_module(opt['mother_interface'].options)
         super(AskRun,self).__init__(self.to_control, opt['mother_interface'],
                                      *args, **opt)
         
@@ -557,6 +557,18 @@ def check_available_module(self, options):
             if misc.has_f2py() or options['f2py_compiler']:
                 self.available_module.add('reweight')
 
+        # An MG7 output reweights with the classic module or with MadtRex
+        # (see madgraph/various/reweight_backend.py): reweighting is
+        # available when either one can run.
+        self.reweight_env = None
+        if not MADEVENT and os.path.exists(pjoin(self.me_dir, 'Cards', 'run_card.toml')):
+            from madgraph.various import reweight_backend
+            self.reweight_env = reweight_backend.probe_environment(
+                self.me_dir, dict(options, mg5_path=options.get('mg5_path') or MG5DIR),
+                has_f2py='reweight' in self.available_module)
+            if reweight_backend.available_backends(self.reweight_env):
+                self.available_module.add('reweight')
+
 #   old mode to activate the shower            
     def ans_parton(self, value=None):
         """None: means that the user type 'pythia'
@@ -907,7 +919,16 @@ def get_allowed_reweight(self):
         if 'reweight' not in self.available_module:
             self.allowed_reweight = []
             return self.allowed_reweight
-        self.allowed_reweight = ['OFF', 'ON', 'density']
+        if self.reweight_env is not None:
+            # MG7 output: ON is auto; then the backends that can run here
+            from madgraph.various import reweight_backend
+            backends = reweight_backend.available_backends(self.reweight_env)
+            self.allowed_reweight = ['OFF', 'ON'] + backends
+            if 'classic' not in backends:
+                return self.allowed_reweight
+            self.allowed_reweight.append('density')
+        else:
+            self.allowed_reweight = ['OFF', 'ON', 'density']
         
         # check for plugin mode
         plugin_path = self.mother_interface.plugin_path
@@ -920,13 +941,7 @@ def set_default_reweight(self):
         
         if 'reweight' in self.available_module:
             if os.path.exists(pjoin(self.me_dir,'Cards','reweight_card.dat')):
-                reweightcard = open(pjoin(self.me_dir,'Cards','reweight_card.dat'), 'r')
-                content = reweightcard.read()
-                if 'change particle_in_density_matrix' in content:
-                    self.switch['reweight'] = 'density'
-                else:
-                    self.switch['reweight'] = 'ON'
-                reweightcard.close()
+                self.switch['reweight'] = self.switch_value_from_card_reweight()
             else:
                 self.switch['reweight'] = 'OFF'
         else:
@@ -934,6 +949,9 @@ def set_default_reweight(self):
 
     def get_cardcmd_for_reweight(self, value):
         """set some command to run before allowing the user to modify the cards."""
+        if value in ['classic', 'madtrex']:
+            # the two backends of an MG7 output read the same card as ON
+            value = 'ON'
         if value in ['density']:
             content_rwgt_card = open(pjoin(self.me_dir, "Cards", "reweight_card.dat"), "r")
             if 'change particle_in_density_matrix' in content_rwgt_card.read():
@@ -961,14 +979,26 @@ def get_cardcmd_for_reweight(self, value):
 
     def switch_value_from_card_reweight(self):
         """re-derive the reweight switch (ON vs density) from the content of
-        reweight_card.dat (called after the user has edited the cards)."""
+        reweight_card.dat (called after the user has edited the cards). For
+        an MG7 output, a "change backend classic|madtrex" line of the card
+        gives that backend (when it can run here)."""
 
         path = pjoin(self.me_dir, 'Cards', 'reweight_card.dat')
         if not os.path.exists(path):
             return None
         with open(path) as fsock:
-            if 'change particle_in_density_matrix' in fsock.read():
-                return 'density'
+            content = fsock.read()
+        if 'change particle_in_density_matrix' in content:
+            return 'density'
+        if getattr(self, 'reweight_env', None) is not None:
+            # a backend picked in the question outranks the card's line
+            # (the run warns when the two disagree)
+            current = getattr(self, 'switch', {}).get('reweight')
+            if current in ('classic', 'madtrex'):
+                return current
+            match = re.search(r'^\s*change\s+backend\s+(\w+)', content, re.M | re.I)
+            if match and match.group(1).lower() in self.get_allowed_reweight():
+                return match.group(1).lower()
         return 'ON'
 
 
diff --git a/madgraph/interface/madgraph_interface.py b/madgraph/interface/madgraph_interface.py
index a67e23d574..a469441513 100755
--- a/madgraph/interface/madgraph_interface.py
+++ b/madgraph/interface/madgraph_interface.py
@@ -623,6 +623,14 @@ def help_install(self):
         logger.info("     --no-debug     Optimized build, no debug symbols, Release (source build, default).")
         logger.info("     --cuda-arch=ARCHS  Semicolon-separated CUDA compute capabilities, e.g. 75;80;86.")
         logger.info("     --hip-arch=ARCHS   Semicolon-separated HIP architectures, e.g. gfx900;gfx906.")
+        logger.info(" ")
+        logger.info("   \"install madtrex [options]\"",'$MG:BOLD')
+        logger.info("   Install MadtRex, the event-reweighting library used by MG7, together")
+        logger.info("   with its bundled Rex fork. Requires madspace to be installed first.")
+        logger.info("   Without options an interactive installer is launched. Available options:")
+        logger.info("     -y/--yes       Install non-interactively.")
+        logger.info("     -j N/--jobs=N  Parallel compilation jobs (defaults to the 'nb_core' option).")
+        logger.info("     --clean        Rebuild from scratch.")
 
     def help_display(self):
         logger.info("syntax: display " + "|".join(self._display_opts),'$MG:color:BLUE')
@@ -3749,6 +3757,12 @@ def complete_install(self, text, line, begidx, endidx):
                 if any(a.startswith(opt) for a in args[2:]):
                     options.remove(opt)
             return self.list_completion(text, options)
+        elif len(args) >= 2 and args[1] == 'madtrex':
+            options = ['-y', '--yes', '--clean', '--jobs=']
+            for opt in options[:]:
+                if any(a.startswith(opt) for a in args[2:]):
+                    options.remove(opt)
+            return self.list_completion(text, options)
         elif len(args)>=2 and args[1] in self._advanced_install_opts:
             options = ['--keep_source','--logging=']
             if args[1]=='pythia8':
@@ -3804,7 +3818,7 @@ def _tutorial_opts(self):
     _install_opts = ['Delphes', 'ExRootAnalysis',
                      'update', 'Golem95', 'QCDLoop', 'maddm', 'maddump',
                      'looptools', 'cuttools_iregi', 'MadSTR', 'RunningCoupling',
-                     'madspace']
+                     'madspace', 'madtrex']
     
     # The targets below are installed using the HEPToolsInstaller.py script
     _advanced_install_opts = ['pythia8','zlib','boost','lhapdf6','lhapdf5','collier',
@@ -8571,8 +8585,8 @@ def do_install(self, line, paths=None, additional_options=[]):
         elif args[0] == 'cuttools_iregi':
             self.install_cuttools_iregi()
             return
-        elif args[0] == 'madspace':
-            install_script = pjoin(MG5DIR, 'madspace', 'install.py')
+        elif args[0] in ('madspace', 'madtrex'):
+            install_script = pjoin(MG5DIR, args[0], 'install.py')
             install_args = args[1:]
             # The source build compiles in parallel only if it is told how many
             # jobs it may use (see set_build_parallelism in install.py), so hand
@@ -8580,7 +8594,11 @@ def do_install(self, line, paths=None, additional_options=[]):
             if not any(a == '-j' or a.startswith('-j') or a.startswith('--jobs')
                        for a in install_args):
                 install_args = install_args + ['-j', str(self.options['nb_core'])]
-            subprocess.run([sys.executable, install_script] + install_args)
+            result = subprocess.run([sys.executable, install_script] + install_args)
+            if result.returncode != 0:
+                raise self.InvalidCmd('install %s failed (%s exited with code %d): see '
+                                      'its output above' % (args[0], install_script,
+                                                            result.returncode))
             return
 
         plugin = self.install_plugin
diff --git a/madgraph/interface/reweight_interface.py b/madgraph/interface/reweight_interface.py
index a949b0e0c6..961b5d5062 100755
--- a/madgraph/interface/reweight_interface.py
+++ b/madgraph/interface/reweight_interface.py
@@ -62,6 +62,28 @@
 #lhapdf = None
 
 
+# The "change <option>" lines of a reweight card that ReweightInterface.do_change
+# understands, and those DensityInterface adds (its do_change_<option>
+# methods). The reweighting-backend selector (madgraph/various/
+# reweight_backend.py) uses them to tell which lines need this (the classic)
+# backend.
+REWEIGHT_CHANGE_OPTIONS = frozenset([
+    'backend', 'model', 'process', 'keep_ordering', 'use_eventid',
+    'allow_missing_finalstate', 'boost', 'virtual_path', 'tree_path',
+    'output', 'helicity', 'mode', 'rwgt_dir', 'systematics', 'soft_threshold',
+    'multicore', 'identical_particle_in_prod_and_decay', 'include_sudakov'])
+DENSITY_CHANGE_OPTIONS = frozenset([
+    'helicity_direction', 'boost_choice', 'order_helicities',
+    'symmetrise_initial_state', 'axis_referential', 'matrix_normalisation',
+    'particle_in_density_matrix'])
+# "change" options of the shared reweight-card vocabulary that only MadtRex
+# (MG7 outputs) understands; an error here rather than a discarded line.
+MADTREX_ONLY_CHANGE_OPTIONS = frozenset([
+    'fptype', 'verbose', 'max_batch', 'max_deferred_iterations', 'streaming',
+    'stream_batch_size', 'nb_core', 'min_batch', 'alpha_s_power',
+    'scale_variation_output'])
+
+
 class ReweightInterface(extended_cmd.Cmd):
     """Basic interface for reweighting operation"""
 
@@ -376,6 +398,8 @@ def complete_import(self, text, line, begidx, endidx):
     def help_change(self):
         """help for change command"""
 
+        print("change backend classic|auto: the reweighting program (MadtRex, the other")
+        print("               backend, needs an MG7 output)")
         print("change model X :use model X for the reweighting")
         print("change process p p > e+ e-: use a new process for the reweighting")
         print("change process p p > mu+ mu- --add : add one new process to existing ones")
@@ -390,7 +414,23 @@ def do_change(self, line):
         args = self.split_arg(line)
         if len(args)<2:
             logger.critical("not enough argument (need at least two). Discard line")
-        if args[0] == "model":
+        if args[0] in MADTREX_ONLY_CHANGE_OPTIONS or \
+                (args[0] == 'mode' and len(args) > 1 and args[1].lower() == 'pdf'):
+            raise self.InvalidCmd(
+                "'change %s' is MadtRex only: MadtRex reweights MG7 outputs "
+                "(change backend madtrex)" % ' '.join(args))
+        if args[0] not in REWEIGHT_CHANGE_OPTIONS:
+            logger.critical("unknown option! %s.  Discard line." % args[0])
+        elif args[0] == "backend":
+            # which backend reads the card (madgraph/various/reweight_backend.py)
+            if len(args) < 2 or args[1].lower() not in ('classic', 'auto', 'madtrex'):
+                raise self.InvalidCmd(
+                    "change backend: expected classic, madtrex or auto")
+            if args[1].lower() == 'madtrex':
+                raise self.InvalidCmd(
+                    "the reweight card asks for MadtRex ('change backend madtrex'): "
+                    "MadtRex needs an MG7 output; use 'change backend classic' or auto")
+        elif args[0] == "model":
             nb_f2py_module += 1 # tag to force the f2py to reload
             self.second_model = " ".join(args[1:])
             if self.has_standalone_dir:
@@ -463,8 +503,6 @@ def do_change(self, line):
             if args[1] == 'True':
                 self.inc_sudakov = True
                 self.rwgt_mode = 'LO'
-        else:
-            logger.critical("unknown option! %s.  Discard line." % args[0])
         
             
     def check_launch(self, args):
diff --git a/madgraph/iolibs/export_cpp.py b/madgraph/iolibs/export_cpp.py
index debabcae47..fdbcb07adc 100755
--- a/madgraph/iolibs/export_cpp.py
+++ b/madgraph/iolibs/export_cpp.py
@@ -121,7 +121,15 @@ def __init__(self, model, output_path, wanted_lorentz = [],
         self.params_dep = []   # base_objects.ModelVariable
         self.params_indep = [] # base_objects.ModelVariable
         self.coups_flv_dep = []    # (name, object, [couplings])
-        self.coups_flv_indep = []  # (name, object, [couplings]) 
+        self.coups_flv_indep = []  # (name, object, [couplings])
+        self.params_ext_names = [] # names of independent parameters read directly
+                                    # from the SLHA card (as opposed to internal
+                                    # parameters/couplings derived from them);
+                                    # these are the only ones settable by name
+                                    # at runtime (e.g. via the UMAMI interface)
+        self.params_ext_slha = []  # (name, lhablock.lower(), lhacode) for the same
+                                    # parameters, so they can also be addressed by
+                                    # their SLHA card location (block + index)
         self.p_to_cpp = parsers.UFOExpressionParserCPP()
 
         # Prepare parameters and couplings for writeout in C++
@@ -184,6 +192,18 @@ def prepare_parameters(self):
                               % (param.name, param.lhablock.lower(), param.value.real)
             else:
                 raise MadGraph5Error("Only support for SLHA blocks with 1 or 2 indices")
+            # Track which independent parameters were read directly from the
+            # SLHA card (as opposed to internal parameters derived from them):
+            # these are the only ones settable by name at runtime (e.g. via
+            # the UMAMI set_parameter interface). NB: this list is consumed
+            # only by write_set_parameters() below, not by param.expr itself,
+            # since other UFOModelConverterCPP subclasses (e.g. the MadMatrix/
+            # cudacpp backend) read param.expr directly for purposes (like
+            # generating hardcoded/constexpr parameter mirrors) that require
+            # it to remain a single plain assignment.
+            self.params_ext_names.append(param.name)
+            self.params_ext_slha.append((param.name, param.lhablock.lower(),
+                                          list(param.lhacode)))
             self.params_indep.insert(0,
                                    base_objects.ModelVariable(param.name,
                                                    expression,
@@ -314,6 +334,14 @@ def generate_parameters_class_files(self):
         replace_dict['set_flv_couplings'] = \
                                 self.write_flv_couplings(self.coups_flv_indep)
 
+        replace_dict['register_parameters'] = \
+                               self.write_register_parameters(
+                                   self.params_indep + self.params_dep + \
+                                   self.coups_indep + list(self.coups_dep.values()),
+                                   self.params_ext_names)
+        replace_dict['register_slha_keys'] = \
+                               self.write_register_slha_keys(self.params_ext_slha)
+
         replace_dict['print_independent_parameters'] = \
                                self.write_print_parameters(self.params_indep)
         replace_dict['print_independent_couplings'] = \
@@ -360,11 +388,21 @@ def write_parameters(self, params):
     def write_set_parameters(self, params):
         """Write out the lines of independent parameters"""
 
-        # For each parameter, write name = expr;
+        # For each parameter, write name = expr;, unless a runtime override
+        # was set by name for this (necessarily independent, SLHA-card-level)
+        # parameter (e.g. via the UMAMI set_parameter interface), in which
+        # case the override takes precedence over the value read from the
+        # param card
 
         res_strings = []
         for param in params:
-            res_strings.append("%s" % param.expr)
+            if param.name in self.params_ext_names:
+                res_strings.append(
+                    "if (m_param_overrides.find(\"%s\") != m_param_overrides.end())\n"
+                    "%s = m_param_overrides[\"%s\"];\nelse\n%s" %
+                    (param.name, param.name, param.name, param.expr))
+            else:
+                res_strings.append("%s" % param.expr)
 
         # Correct width sign for Majorana particles (where the width
         # and mass need to have the same sign)        
@@ -377,6 +415,55 @@ def write_set_parameters(self, params):
 
         return "\n".join(res_strings)
 
+    def write_register_parameters(self, params, ext_names, complex_map_name='m_complex_params'):
+        """Write out the code that registers name->pointer entries for every
+        scalar parameter and coupling (independent or dependent), and marks
+        which independent parameters were read directly from the SLHA card
+        (as opposed to being derived internally). This lookup-by-name table
+        backs runtime access to parameters/couplings, e.g. through the UMAMI
+        set_parameter/get_parameter interface. FLV_COUPLING entries are
+        skipped: they are not simple scalar values.
+
+        complex_map_name: which C++ map a complex-valued param's pointer
+        gets registered into (default 'm_complex_params'). Model parameters
+        (mdl_*) and derived couplings (GC_*) are *not* interchangeable
+        here in the mg7/madmatrix backend: couplings are declared cxtype
+        (which the caller may have templated to a non-double precision --
+        see model_handling.py's own 'cxsmpl<double>' -> 'cxtype' rewrite
+        for independent couplings specifically), while parameters always
+        stay plain cxsmpl<double> for numerical accuracy (matching
+        m_real_params' own always-double real-parameter pointers) -- so a
+        caller registering both groups needs two separate calls, one per
+        map, each with its own C++ declaration of matching pointer type;
+        passing both groups into a single call/map only compiles by
+        accident whenever cxtype happens to resolve to cxsmpl<double> too
+        (FPTYPE=m/d), and fails to compile at FPTYPE=f."""
+
+        lines = []
+        for param in params:
+            if not hasattr(param, 'type'):
+                continue
+            map_name = 'm_real_params' if param.type == 'real' \
+                       else complex_map_name
+            lines.append('%s["%s"] = &%s;' % (map_name, param.name, param.name))
+        for name in ext_names:
+            lines.append('m_settable_params.insert("%s");' % name)
+
+        return "\n".join(lines)
+
+    def write_register_slha_keys(self, params_ext_slha):
+        """Write out the code that registers a lookup from SLHA (block, index[,
+        index2]) key strings to parameter names, so that a settable parameter
+        can also be addressed by its location in the SLHA param card (e.g. "mass
+        6") rather than only by its model name (e.g. "mdl_MT")."""
+
+        lines = []
+        for name, block, lhacode in params_ext_slha:
+            key = block + "".join(" %d" % i for i in lhacode)
+            lines.append('m_slha_key_to_name["%s"] = "%s";' % (key, name))
+
+        return "\n".join(lines)
+
     def _assert_flv_couplings_supported(self, params):
         """Refuse, with a clear and actionable message, the merged-flavor
         coupling structures the C++ (mg7/standalone) backend cannot yet
@@ -1053,6 +1140,8 @@ def get_process_function_definitions(self, write=True):
         replace_dict['initProc_lines'] = \
                                 self.get_initProc_lines(self.matrix_elements[0],
                                                         color_amplitudes)
+        replace_dict['refreshMasses_lines'] = \
+                                self.get_refresh_masses_lines(self.matrix_elements[0])
         replace_dict['reset_jamp_lines'] = \
                                      self.get_reset_jamp_lines(color_amplitudes)
         replace_dict['sigmaKin_lines'], other_replace = \
@@ -1138,6 +1227,17 @@ def get_initProc_lines(self, matrix_element, color_amplitudes):
 
         return "\n".join(initProc_lines)
 
+    # Companion to get_initProc_lines: the same external-mass push_back list, without the
+    # jamp2 allocation (jamp2 is allocated once in the constructor and must not be
+    # reallocated every time CPPProcess::refreshMasses() re-populates mME).
+    def get_refresh_masses_lines(self, matrix_element):
+        """Get refreshMasses_lines for function definition for CPPProcess::refreshMasses"""
+
+        lines = []
+        for part in matrix_element.get_external_wavefunctions():
+            lines.append("mME.push_back(pars.%s);" % part.get('mass'))
+        return "\n".join(lines)
+
     def get_reset_jamp_lines(self, color_amplitudes):
         """Get lines to reset jamps"""
 
@@ -3251,9 +3351,20 @@ class ProcessExporterMG7(ProcessExporterCPP):
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        self.me_lib_format = args[1].get("me_lib_format", None)
+        self.me_lib_format = args[1].get("me_lib_format", None) or pjoin("SubProcesses", "{process_id}", "api.so")
         self.process_info = []
         self.merged_subprocesses = defaultdict(list)
+        # Lazy import: madgraph.iolibs.export_mg7 imports export_cpp itself, so a top-level
+        # import here would be circular. Without this, oneprocessclass silently falls back
+        # to the base ProcessExporterCPP.oneprocessclass (OneProcessExporterCPP), which lacks
+        # the .name attribute (and other MG7-specific behaviour) generate_subprocess_directory
+        # needs. Only apply the fallback when nothing more specific was set: subclasses such
+        # as madmatrix.output.ProcessExporterMadMatrix set their own oneprocessclass (with the
+        # coloramps.h/epoch_process_id.h/color_sum.cc/makefile-symlink logic in
+        # OneProcessExporterMadMatrix.generate_process_files()) and that must not be clobbered.
+        if self.oneprocessclass is OneProcessExporterCPP:
+            from madgraph.iolibs import export_mg7
+            self.oneprocessclass = export_mg7.OneProcessExporterMG7
 
     def generate_subprocess_directory(
         self, matrix_element, cpp_helas_call_writer, proc_number=None
@@ -3268,6 +3379,13 @@ def generate_subprocess_directory(
         # Create the directory PN_xx_xxxxx in the specified path
         proc_dir_name = process_exporter_mg7.name
         dirpath = pjoin(self.dir_path, 'SubProcesses', proc_dir_name)
+        # self.model/self.model_name aren't reliably set on this class by the
+        # time subprocess directories are generated (see eg
+        # write_check_sa_cpp()'s own "self.model may not be set yet" fallback
+        # above); the matrix element's own model is always available though,
+        # and is what %(model)s in template_Sub_make/template_src_make is
+        # substituted from too (via copy_template()'s model parameter).
+        model = matrix_element.get('processes')[0].get('model')
 
         try:
             os.mkdir(dirpath)
@@ -3284,7 +3402,6 @@ def generate_subprocess_directory(
         # Generate SVG Feynman diagrams (diagrams.svg + diagrams.json)
         if not self.opt.get('output_options', {}).get('noeps') == 'True':
             svg_stem = pjoin(dirpath, 'diagrams')
-            model = matrix_element.get('processes')[0].get('model')
             diagrams = matrix_element.get('base_amplitude').get('diagrams')
             logger.info('Generating Feynman diagrams for %s' %
                         matrix_element.get('processes')[0].nice_string())
@@ -3292,7 +3409,17 @@ def generate_subprocess_directory(
                                                   model=model, amplitude=True)
             plot.draw()
 
-        me_lib_path = self.me_lib_format.format(process_id = proc_dir_name)
+        # me_lib_format may (eg madmatrix/output.py's ProcessExporterMadMatrix)
+        # carry a {model} placeholder alongside {process_id}, resolved here to
+        # the same sanitized model name template_Sub_make's %(model)s uses;
+        # harmless (str.format ignores unused kwargs) for exporters whose
+        # me_lib_format has no {model} placeholder at all.
+        # kept for the exporters that name the libraries elsewhere too (the
+        # process table of madmatrix/output.py's standalone exporter)
+        self.lib_model_name = self.get_model_name(model.get('name'))
+        me_lib_path = self.me_lib_format.format(
+            process_id=proc_dir_name, model=self.lib_model_name
+        )
         # Store the path relative to the process directory (like me_path): an
         # absolute path would break as soon as the process directory is moved.
         rel_dirpath = pjoin('SubProcesses', proc_dir_name)
diff --git a/madgraph/iolibs/export_v4.py b/madgraph/iolibs/export_v4.py
index 90443a2e99..a1f5f7cc7d 100755
--- a/madgraph/iolibs/export_v4.py
+++ b/madgraph/iolibs/export_v4.py
@@ -5001,12 +5001,10 @@ def set_fortran_compiler(self, default_compiler, force=False):
         # Try to find the correct one.
         if default_compiler['f2py'] and misc.which(default_compiler['f2py']):
             f2py_compiler = default_compiler['f2py']
-        elif misc.which('f2py%d.%d' %(sys.version_info.major, sys.version_info.minor)):
-            f2py_compiler = 'f2py%d.%d' %(sys.version_info.major, sys.version_info.minor)
-        elif misc.which('f2py%d' %(sys.version_info.major)):
-            f2py_compiler = 'f2py%d' %(sys.version_info.major)            
-        elif misc.which('f2py'):
-            f2py_compiler = 'f2py'
+        elif misc.default_f2py():
+            # this Python's own f2py first: f2py3 on the PATH may build
+            # modules for another Python
+            f2py_compiler = misc.default_f2py()
 
 
         to_replace = {'fortran': f77_compiler, 'f2py': f2py_compiler}
diff --git a/madgraph/iolibs/template_files/cpp_model_parameters_cc.inc b/madgraph/iolibs/template_files/cpp_model_parameters_cc.inc
index 26ba3a3668..c3b0222785 100644
--- a/madgraph/iolibs/template_files/cpp_model_parameters_cc.inc
+++ b/madgraph/iolibs/template_files/cpp_model_parameters_cc.inc
@@ -3,16 +3,21 @@
 %(info_lines)s
 //==========================================================================
 
+#include <algorithm>
+#include <cctype>
 #include <iostream>
 #include <iomanip>
+#include <sstream>
 #include "Parameters_%(model_name)s.h"
 
     void Parameters_%(model_name)s::setIndependentParameters(SLHAReader& slha){
 	// Define "zero"
 	zero=0;
 	ZERO=0;
-	// Prepare a vector for indices 
+	// Prepare a vector for indices
 	vector<int> indices(2,0);
+	m_slha = slha;
+	registerParameters();
 	%(set_independent_parameters)s
     }
     void Parameters_%(model_name)s::setIndependentCouplings(){
@@ -44,3 +49,62 @@
 	%(print_dependent_couplings)s
     }
 
+    void Parameters_%(model_name)s::registerParameters(){
+	%(register_parameters)s
+	%(register_slha_keys)s
+    }
+
+    bool Parameters_%(model_name)s::resolveSlhaKey(const std::string& name, std::string& resolved) const {
+	std::istringstream iss(name);
+	std::string block;
+	if (!(iss >> block)) return false;
+	std::transform(block.begin(), block.end(), block.begin(),
+	               [](unsigned char c){ return std::tolower(c); });
+	std::string key = block;
+	bool any_index = false;
+	int index;
+	while (iss >> index) {
+	    key += " " + std::to_string(index);
+	    any_index = true;
+	}
+	if (!any_index) return false;
+	std::map<std::string, std::string>::const_iterator it = m_slha_key_to_name.find(key);
+	if (it == m_slha_key_to_name.end()) return false;
+	resolved = it->second;
+	return true;
+    }
+
+    bool Parameters_%(model_name)s::setParameterByName(std::string name, double real, double imag){
+	(void)imag; // settable (SLHA-card-level) parameters are always real
+	std::string resolved;
+	if (resolveSlhaKey(name, resolved)) name = resolved;
+	if (m_settable_params.find(name) == m_settable_params.end())
+	    return false;
+	m_param_overrides[name] = real;
+	setIndependentParameters(m_slha);
+	setIndependentCouplings();
+	return true;
+    }
+
+    bool Parameters_%(model_name)s::getParameterByName(std::string name, double* real, double* imag){
+	std::string resolved;
+	if (resolveSlhaKey(name, resolved)) name = resolved;
+	map<std::string, double*>::iterator real_it = m_real_params.find(name);
+	if (real_it != m_real_params.end()){
+	    *real = *(real_it->second);
+	    if (imag) *imag = 0.;
+	    return true;
+	}
+	map<std::string, std::complex<double> *>::iterator complex_it = m_complex_params.find(name);
+	if (complex_it != m_complex_params.end()){
+	    *real = complex_it->second->real();
+	    if (imag) *imag = complex_it->second->imag();
+	    return true;
+	}
+	return false;
+    }
+
+    void Parameters_%(model_name)s::resetParameterOverrides(){
+	m_param_overrides.clear();
+    }
+
diff --git a/madgraph/iolibs/template_files/cpp_model_parameters_h.inc b/madgraph/iolibs/template_files/cpp_model_parameters_h.inc
index 7fa3fb27e8..4f044b9f04 100644
--- a/madgraph/iolibs/template_files/cpp_model_parameters_h.inc
+++ b/madgraph/iolibs/template_files/cpp_model_parameters_h.inc
@@ -7,6 +7,9 @@
 #define Parameters_%(model_name)s_H
 
 #include <complex>
+#include <map>
+#include <set>
+#include <string>
 
 #include "read_slha.h"
 using namespace std;
@@ -48,6 +51,46 @@ void printIndependentCouplings();
 void printDependentParameters();
 // Print couplings that are changed event by event
 void printDependentCouplings();
+
+// Set the value of an independent, SLHA-card-level parameter by name at
+// runtime (overriding the value read from the param card), and recompute
+// every parameter/coupling that depends on it. Returns false if "name" is
+// not the name of a settable independent parameter (e.g. it names an
+// internal parameter derived from others, or a coupling, neither of which
+// can be set directly since they would just be recomputed from the
+// underlying parameters on the next call). "name" may be either the
+// parameter's model name (e.g. "mdl_MT") or its SLHA card location, given as
+// "<block> <index> [<index2>]" (e.g. "mass 6"); the block name is matched
+// case-insensitively.
+bool setParameterByName(std::string name, double real, double imag);
+// Retrieve the current value of any parameter or coupling by name (real or
+// complex, independent or dependent). As with setParameterByName, "name" may
+// be either the parameter's model name or its SLHA card location. Returns
+// false if "name" is unknown.
+bool getParameterByName(std::string name, double* real, double* imag);
+// Drop all overrides previously set via setParameterByName(), so that the
+// next setIndependentParameters(slha) call starts from a clean slate (e.g.
+// when (re-)initializing from a param card). Does not itself recompute
+// anything; the caller is expected to call setIndependentParameters()/
+// setIndependentCouplings() afterward (as CPPProcess's constructor does).
+void resetParameterOverrides();
+
+private:
+
+// If "name" is of the form "<block> <index> [<index2>]", resolve it (via
+// m_slha_key_to_name) to the corresponding parameter's model name and return
+// true; otherwise leave "resolved" untouched and return false.
+bool resolveSlhaKey(const std::string& name, std::string& resolved) const;
+
+// Name-keyed lookup tables populated by registerParameters(), backing
+// setParameterByName()/getParameterByName() above.
+void registerParameters();
+std::map<std::string, double*> m_real_params;
+std::map<std::string, std::complex<double> *> m_complex_params;
+std::set<std::string> m_settable_params;
+std::map<std::string, double> m_param_overrides;
+std::map<std::string, std::string> m_slha_key_to_name;
+SLHAReader m_slha;
 };
 
 #endif // Parameters_%(model_name)s_H
diff --git a/madgraph/iolibs/template_files/cpp_process_class.inc b/madgraph/iolibs/template_files/cpp_process_class.inc
index b64dfd4f2f..3739efb726 100644
--- a/madgraph/iolibs/template_files/cpp_process_class.inc
+++ b/madgraph/iolibs/template_files/cpp_process_class.inc
@@ -29,6 +29,12 @@ public:
 
   const vector<double>& getMasses() const {return mME;}
 
+  // Recompute mME (external particle on-shell masses) from the current model parameters
+  // (pars). mME is only populated once, in the constructor; call this after overriding a
+  // mass via pars.setParameterByName() (e.g. through umami_set_parameter) so getMasses()
+  // reflects the new value.
+  void refreshMasses();
+
   const double* getAmp2() const { return amp2; }
 
   Parameters_%(model_name)s& getParameters() { return pars; }
diff --git a/madgraph/iolibs/template_files/cpp_process_function_definitions.inc b/madgraph/iolibs/template_files/cpp_process_function_definitions.inc
index db3a44fee6..79a3be1664 100644
--- a/madgraph/iolibs/template_files/cpp_process_function_definitions.inc
+++ b/madgraph/iolibs/template_files/cpp_process_function_definitions.inc
@@ -10,13 +10,22 @@ CPPProcess::CPPProcess(string param_card_name) :
 {
 // Instantiate the model class and set parameters that stay fixed during run
     SLHAReader slha(param_card_name, false);
+    pars.resetParameterOverrides();
     pars.setIndependentParameters(slha);
     pars.setIndependentCouplings();
     %(initProc_lines)s
-} 
+}
+
+//--------------------------------------------------------------------------
+// Re-populate mME from the current model parameters (pars). See declaration in CPPProcess.h.
+
+void CPPProcess::refreshMasses() {
+    mME.clear();
+    %(refreshMasses_lines)s
+}
 
 //--------------------------------------------------------------------------
-// Evaluate |M|^2, part independent of incoming flavour. 
+// Evaluate |M|^2, part independent of incoming flavour.
 
 double CPPProcess::sigmaKin(int flavor_id) { 
     // Set the parameters which change event by event
diff --git a/madgraph/iolibs/template_files/madmatrix/backend/common/Density.cc b/madgraph/iolibs/template_files/madmatrix/backend/common/Density.cc
index f0a04883ba..0b8870a3d7 100644
--- a/madgraph/iolibs/template_files/madmatrix/backend/common/Density.cc
+++ b/madgraph/iolibs/template_files/madmatrix/backend/common/Density.cc
@@ -199,7 +199,7 @@ namespace madmatrix
             for( int n = 0; n < nChanging; n++ ) row[pos0[n]] = allowHel[icomb * nChanging + n];
             cxtype_amp_sv* jamp_sv = &s.jamps[(size_t)icomb * nParity * njampso];
             for( int i = 0; i < nParity * njampso; i++ ) jamp_sv[i] = cxzero_sv<cxtype_amp_sv>();
-            calculate_jamps( row, allmomenta, allcouplings, allflavors, jamp_sv, false, numerators, denominators, nullptr, ievt00 );
+            calculate_jamps( row, nullptr, allmomenta, allcouplings, allflavors, jamp_sv, false, numerators, denominators, nullptr, ievt00 );
           }
           bool nonzero = false;
           for( int iParity = 0; iParity < nParity; iParity++ )
diff --git a/madgraph/iolibs/template_files/madmatrix/backend/cpu/SigmaKin.cc b/madgraph/iolibs/template_files/madmatrix/backend/cpu/SigmaKin.cc
index b91982eeac..14ea811124 100644
--- a/madgraph/iolibs/template_files/madmatrix/backend/cpu/SigmaKin.cc
+++ b/madgraph/iolibs/template_files/madmatrix/backend/cpu/SigmaKin.cc
@@ -193,6 +193,29 @@ namespace madmatrix
 
   //--------------------------------------------------------------------------
 
+  // The helicities of the npar external legs for the events of one page (neppV events,
+  // one SIMD lane each, or one event without SIMD), as the external wavefunctions take
+  // them: the helicities hel[npar] for every event, or the helicity allihel[ievt] of each event.
+  static inline void
+  fillHelicities( fptype_amp_sv* hel_sv,  // output: hel_sv[npar] helicities of the external legs
+                  const short* hel,       // input: helicity of each leg, hel[npar], of all events (if allihel == nullptr)
+                  const int* allihel,     // input: helicity index [0,ncomb) of each event[nevt] (nullptr: hel for all events)
+                  const int ievt0 )       // input: first event number in current C++ event page
+  {
+#ifdef MGONGPU_CPPSIMD
+    for( int ieppV = 0; ieppV < neppV; ieppV++ )
+    {
+      const short* heli = ( allihel != nullptr ? cHel[allihel[ievt0 + ieppV]] : hel );
+      for( int ipar = 0; ipar < npar; ipar++ ) hel_sv[ipar][ieppV] = heli[ipar];
+    }
+#else
+    const short* heli = ( allihel != nullptr ? cHel[allihel[ievt0]] : hel );
+    for( int ipar = 0; ipar < npar; ipar++ ) hel_sv[ipar] = heli[ipar];
+#endif
+  }
+
+  //--------------------------------------------------------------------------
+
   // Accumulate a multichannel numerator contribution in place. In C++ each
   // event page is processed serially within the helicity loop, so a plain
   // sum suffices (CUDA needs atomicAdd instead: see backend/gpu/SigmaKin.cc).
@@ -204,9 +227,12 @@ namespace madmatrix
   // moved to a separate function/kernel). This function processes a single event "page"
   // or SIMD vector (or for two in "mixed" precision mode, nParity=2). Accepts a SCALAR
   // channelId because it is GUARANTEED that all events in a SIMD vector have the same
-  // channelId #898.
+  // channelId #898. The helicity, on the other hand, may differ from one event (SIMD lane)
+  // to the next: the external wavefunctions are branch-free in the helicity (see helas.h),
+  // so that the events of a page with different helicities are computed in lock-step.
   void
-  calculate_jamps( const short* hel,                 // input: helicity of each leg, hel[npar] (a row of cHel, or not: see Density.cc)
+  calculate_jamps( const short* hel,                   // input: helicity of each leg, hel[npar], of all events (if allihel == nullptr): a row of cHel, or not (see Density.cc)
+                   const int* allihel,                 // input: helicity index [0,ncomb) of each event[nevt] (nullptr: hel for all events)
                    const fptype_momenta* allmomenta,   // input: momenta[nevt*npar*4]
                    const fptype* allcouplings,         // input: couplings[nevt*ndcoup*2]
                    const unsigned int* iflavorVec,     // input: indices of the flavor combinations
@@ -227,16 +253,6 @@ namespace madmatrix
     using NUM_ACCESS = HostAccessNumerators;    // non-trivial access: buffer includes all events
     mgDebug( 0, __FUNCTION__ );
 
-    // EvaluateDiagrams.inc reads the helicities as cHel[ihel][ipar]. Shadow the
-    // helicity table with the one row this call evaluates, so that the row need
-    // not be one of cHel at all: the density matrix builds combinations that a
-    // polarised process leaves out of its table (sigmaKin_density, Density.cc).
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wshadow"
-    const short( *cHel )[npar] = reinterpret_cast<const short( * )[npar]>( hel );
-#pragma GCC diagnostic pop
-    constexpr int ihel = 0;
-
     // Local TEMPORARY variables for a subset of Feynman diagrams in the given C++ event
     // page (ipagV) [NB these variables are reused several times (and re-initialised each
     // time) within the same event or event page]. Create memory for both momenta and
@@ -292,6 +308,10 @@ namespace madmatrix
       for( size_t iicoup = 0; iicoup < nIPC; iicoup++ )
         COUPs[ndcoup + iicoup] = allCOUPs[ndcoup + iicoup]; // independent couplings, fixed for all events
       fptype_amp* numerators = NUM_ACCESS::ieventAccessRecord( allNumerators, ievt0 * ndiagrams );
+      // The helicities of the external legs for the events of this page, one per SIMD lane:
+      // the same helicities hel for all of them, or the helicity allihel of each event
+      fptype_amp_sv hel_sv[npar];
+      fillHelicities( hel_sv, hel, allihel, ievt0 );
       // Create an array of views over the Flavor Couplings
       FLV_COUPLING_ARRAY<nIPF, nMF> flvCOUPs{ cIPF_partner1, cIPF_partner2, cIPF_value };
 
@@ -414,7 +434,7 @@ namespace madmatrix
 #else
         cxtype_amp_sv jamp_sv[njampso] = {}; // all zeros
 #endif
-        calculate_jamps( cHel[ihel], allmomenta, allcouplings, hgFlavorVec, jamp_sv, false, allNumerators, allDenominators, jamp2_sv, ievt00 );
+        calculate_jamps( cHel[ihel], nullptr, allmomenta, allcouplings, hgFlavorVec, jamp_sv, false, allNumerators, allDenominators, jamp2_sv, ievt00 );
         color_sum_cpu( allMEs, jamp_sv, ievt00 );
         for( int ieppV = 0; ieppV < neppV; ++ieppV )
         {
@@ -569,7 +589,8 @@ namespace madmatrix
             unsigned int* allDiagramIdsOut,    // output: multichannel channelIds[nevt] (1 to #diagrams)
             bool mulChannelWeight,             // if true, multiply channel weight to ME output
             const int nevt,                    // input: #events (for cuda: nevt == ndim == gpublocks*gputhreads)
-            const bool sampleSignedHelicity )  // input: choose the helicity on |T_i| and return sign(T_i)*sum|T| (split orders only)
+            const bool sampleSignedHelicity,   // input: choose the helicity on |T_i| and return sign(T_i)*sum|T| (split orders only)
+            const int* allihel )               // input: helicity index [0,ncomb) of each event[nevt]; nullptr to sum over the good helicities
   {
     mgDebugInitialise();
 
@@ -606,6 +627,10 @@ namespace madmatrix
 
     // === PART 1 - HELICITY LOOP: CALCULATE WAVEFUNCTIONS ===
     // (using precomputed good helicities)
+    // With allihel, each event has a single helicity of its own instead: the "loop" is one
+    // evaluation in which the events of a SIMD page carry different helicities in lock-step,
+    // and everything below (color sum, color and channel choice) runs on that one helicity.
+    const int nhelEval = ( allihel != nullptr ? 1 : cNGoodHel );
 #if defined MGONGPU_CPPSIMD and defined MGONGPU_FPTYPE_DOUBLE and defined MGONGPU_FPTYPE2_FLOAT
     // Mixed fptypes #537: float for color algebra and double elsewhere
     // Delay color algebra and ME updates (only on even pages)
@@ -616,7 +641,7 @@ namespace madmatrix
 #endif
 #ifdef _OPENMP
     // OMP multithreading #575 (NB: tested only with gcc11 so far)
-#define _OMPLIST0 allcouplings, allMEs, allmomenta, allrndcol, allrndhel, allselcol, allselhel, cGoodHel, cNGoodHel, npagV2, sampleSignedHelicity
+#define _OMPLIST0 allcouplings, allMEs, allmomenta, allrndcol, allrndhel, allselcol, allselhel, allihel, cGoodHel, nhelEval, npagV2, sampleSignedHelicity
 #define _OMPLIST1 , allDenominators, allNumerators, allChannelIds, mgOnGpu::icolamp, mgOnGpu::channel2iconfig
 #pragma omp parallel for default( none ) shared( _OMPLIST0 _OMPLIST1 )
 #undef _OMPLIST0
@@ -645,31 +670,35 @@ namespace madmatrix
       if( ColorMatrixData::shouldUseBlas )
       {
         static thread_local std::vector<cxtype_sv> ghelJamp_sv( (size_t)ncomb * nParity * ncolor );
-        for( int ighel = 0; ighel < cNGoodHel; ighel++ )
+        for( int ighel = 0; ighel < nhelEval; ighel++ )
         {
-          const int ihel = cGoodHel[ighel];
+          const int ihel = ( allihel != nullptr ? -1 : cGoodHel[ighel] ); // unused with allihel
           cxtype_sv* jamp_sv = ghelJamp_sv.data() + (size_t)ighel * nParity * ncolor;
           for( int i = 0; i < nParity * ncolor; i++ ) jamp_sv[i] = cxzero_sv<cxtype_sv>(); // calculate_jamps accumulates into jamp_sv
           // **NB! in "mixed" precision, using SIMD, calculate_jamps computes MEs for TWO neppV pages with a single channelId! #924
           bool storeChannelWeights = allChannelIds != nullptr || allrnddiagram != nullptr;
-          calculate_jamps( cHel[ihel], allmomenta, allcouplings, iflavorVec, jamp_sv, storeChannelWeights, allNumerators, allDenominators, jamp2_sv, ievt00 );
+          calculate_jamps( ( allihel != nullptr ? nullptr : cHel[ihel] ), allihel, allmomenta, allcouplings, iflavorVec, jamp_sv, storeChannelWeights, allNumerators, allDenominators, jamp2_sv, ievt00 );
         }
+        // nhelEval may be 0 (no good helicity): the color sum of no helicity leaves allMEs at 0
+        if( nhelEval > 0 )
+        {
 #if defined MGONGPU_CPPSIMD and defined MGONGPU_FPTYPE_DOUBLE and defined MGONGPU_FPTYPE2_FLOAT
-        color_sum_cpu_blas( allMEs, MEs_ighel, MEs_ighel2, ghelJamp_sv.data(), cNGoodHel, ievt00 );
+          color_sum_cpu_blas( allMEs, MEs_ighel, MEs_ighel2, ghelJamp_sv.data(), nhelEval, ievt00 );
 #else
-        color_sum_cpu_blas( allMEs, MEs_ighel, nullptr, ghelJamp_sv.data(), cNGoodHel, ievt00 );
+          color_sum_cpu_blas( allMEs, MEs_ighel, nullptr, ghelJamp_sv.data(), nhelEval, ievt00 );
 #endif
+        }
       }
       else
 #endif // MGONGPU_CPP_HAS_BLAS
       {
-        for( int ighel = 0; ighel < cNGoodHel; ighel++ )
+        for( int ighel = 0; ighel < nhelEval; ighel++ )
         {
-          const int ihel = cGoodHel[ighel];
+          const int ihel = ( allihel != nullptr ? -1 : cGoodHel[ighel] ); // unused with allihel
           cxtype_amp_sv jamp_sv[nParity * njampso] = {}; // fixed nasty bug (omitting 'nParity' caused memory corruptions after calling calculate_jamps)
           // **NB! in "mixed" precision, using SIMD, calculate_jamps computes MEs for TWO neppV pages with a single channelId! #924
           bool storeChannelWeights = allChannelIds != nullptr || allrnddiagram != nullptr;
-          calculate_jamps( cHel[ihel], allmomenta, allcouplings, iflavorVec, jamp_sv, storeChannelWeights, allNumerators, allDenominators, jamp2_sv, ievt00 );
+          calculate_jamps( ( allihel != nullptr ? nullptr : cHel[ihel] ), allihel, allmomenta, allcouplings, iflavorVec, jamp_sv, storeChannelWeights, allNumerators, allDenominators, jamp2_sv, ievt00 );
           color_sum_cpu( allMEs, jamp_sv, ievt00 );
           MEs_ighel[ighel] = E_ACCESS::kernelAccess( E_ACCESS::ieventAccessRecord( allMEs, ievt00 ) );
 #if defined MGONGPU_CPPSIMD and defined MGONGPU_FPTYPE_DOUBLE and defined MGONGPU_FPTYPE2_FLOAT
@@ -682,6 +711,16 @@ namespace madmatrix
       for( int ieppV = 0; ieppV < neppV; ++ieppV )
       {
         const int ievt = ievt00 + ieppV;
+        if( allihel != nullptr )
+        {
+          // No choice: each event was evaluated for its given helicity only, whose |M|^2 (signed,
+          // with an interference) is the color sum above
+          allselhel[ievt] = allihel[ievt] + 1; // NB Fortran [1,ncomb], cudacpp [0,ncomb-1]
+#if defined MGONGPU_CPPSIMD and defined MGONGPU_FPTYPE_DOUBLE and defined MGONGPU_FPTYPE2_FLOAT
+          allselhel[ievt + neppV] = allihel[ievt + neppV] + 1;
+#endif
+          continue;
+        }
         if constexpr( nampso > 1 )
         {
           if( sampleSignedHelicity )
diff --git a/madgraph/iolibs/template_files/madmatrix/backend/cpu/SigmaKin.h b/madgraph/iolibs/template_files/madmatrix/backend/cpu/SigmaKin.h
index 6549e30940..7db8e7e00f 100644
--- a/madgraph/iolibs/template_files/madmatrix/backend/cpu/SigmaKin.h
+++ b/madgraph/iolibs/template_files/madmatrix/backend/cpu/SigmaKin.h
@@ -56,7 +56,11 @@ namespace madmatrix
             // input: when the |M|^2 keeps an interference (squared split orders, nampso > 1),
             // choose the helicity on |T_i| and return sign(T_i) * sum_j |T_j| as madevent does,
             // so that the chosen helicity is exact; false returns sum_j T_j (no effect otherwise)
-            const bool sampleSignedHelicity = false );
+            const bool sampleSignedHelicity = false,
+            // input: the helicity index [0,ncomb) of each event, which is then evaluated for that
+            // helicity alone (and returns its contribution to the helicity-summed |M|^2, see
+            // umami_matrix_element); nullptr, the default, sums over the good helicities
+            const int* allihel = nullptr );
 
   // Setters: called once by CPPProcess (P1-generated) to populate this file's
   // otherwise-internal storage, since it can no longer be written directly
@@ -74,7 +78,8 @@ namespace madmatrix
   // helicity table cHel[ncomb][npar] (row major) the combinations are built
   // from, and the identical-particle factor of a flavor combination.
   void
-  calculate_jamps( const short* hel,                 // input: helicity of each leg, hel[npar]
+  calculate_jamps( const short* hel,                 // input: helicity of each leg, hel[npar], of all events (if allihel == nullptr)
+                   const int* allihel,               // input: helicity index [0,ncomb) of each event[nevt] (nullptr: hel for all events)
                    const fptype_momenta* allmomenta, // input: momenta[nevt*npar*4]
                    const fptype* allcouplings,       // input: couplings[nevt*ndcoup*2]
                    const unsigned int* iflavorVec,   // input: indices of the flavor combinations
diff --git a/madgraph/iolibs/template_files/madmatrix/backend/cpu/umami.cc b/madgraph/iolibs/template_files/madmatrix/backend/cpu/umami.cc
index 419c0310e0..7c32452520 100644
--- a/madgraph/iolibs/template_files/madmatrix/backend/cpu/umami.cc
+++ b/madgraph/iolibs/template_files/madmatrix/backend/cpu/umami.cc
@@ -16,6 +16,7 @@
 #include <algorithm>
 #include <cfloat>
 #include <cmath>
+#include <cstdio>
 #include <string>
 #include <vector>
 #include <array>
@@ -245,8 +246,8 @@ extern "C"
   UmamiStatus umami_supported_inputs( bool const** supported, int* count )
   {
     // MOMENTA, ALPHA_S, FLAVOR_INDEX, RANDOM_COLOR, RANDOM_HELICITY, RANDOM_DIAGRAM,
-    // HELICITY_INDEX=false, DIAGRAM_INDEX=true, CHANNEL_INDEX=false
-    static const bool data[UMAMI_INPUT_KEY_COUNT] = { true, true, true, true, true, true, false, true };
+    // HELICITY_INDEX, DIAGRAM_INDEX=true, CHANNEL_INDEX=false
+    static const bool data[UMAMI_INPUT_KEY_COUNT] = { true, true, true, true, true, true, true, true };
     *supported = data;
     *count = UMAMI_INPUT_KEY_COUNT;
     return UMAMI_SUCCESS;
@@ -308,7 +309,36 @@ extern "C"
       static_cast<InterfaceInstance*>( handle )->summed_helicity = parameter_real != 0;
       return UMAMI_SUCCESS;
     }
+    // Parameters are process-wide (Parameters::getInstance()), not per-handle,
+    // matching how umami_initialize itself never stores the CPPProcess it
+    // constructs: initProc() only ever populates the Parameters singleton.
+#ifdef MGONGPU_HARDCODE_PARAM
+    // Hardcoded parameters are compile-time constants: there is nothing to set.
     return UMAMI_ERROR_NOT_IMPLEMENTED;
+#else
+    if( !Parameters::getInstance()->setParameterByName( name, parameter_real, parameter_imag ) )
+      return UMAMI_ERROR_UNKNOWN_PARAMETER;
+    // sigmaKin() and friends read cached copies of the independent parameters/couplings
+    // (cIPD/cIPC/cIPF/bsmIndepParam), not the Parameters singleton itself, for
+    // performance; refresh those caches now so the override actually takes effect.
+    CPPProcess::refreshIndependentParams();
+    // Keep g_externalMasses in sync with the (possibly updated) singleton.
+    // Emit a one-time warning if a mass parameter was changed, since any phase-space
+    // generator that cached masses from a prior umami_get_meta(UMAMI_META_MASSES,...) call
+    // will need to re-query them.
+    {
+      std::vector<fptype> updatedMasses;
+      CPPProcess::refreshExternalMasses( updatedMasses );
+      bool changed = ( updatedMasses.size() != g_externalMasses.size() );
+      for( size_t i = 0; !changed && i < updatedMasses.size(); ++i )
+        changed = ( static_cast<double>( updatedMasses[i] ) != g_externalMasses[i] );
+      if( changed )
+      {
+        g_externalMasses.assign( updatedMasses.begin(), updatedMasses.end() );
+      }
+    }
+    return UMAMI_SUCCESS;
+#endif
   }
 
   UmamiStatus umami_get_parameter(
@@ -317,7 +347,13 @@ extern "C"
     [[maybe_unused]] double* parameter_real,
     [[maybe_unused]] double* parameter_imag )
   {
+#ifdef MGONGPU_HARDCODE_PARAM
     return UMAMI_ERROR_NOT_IMPLEMENTED;
+#else
+    if( !Parameters::getInstance()->getParameterByName( name, parameter_real, parameter_imag ) )
+      return UMAMI_ERROR_UNKNOWN_PARAMETER;
+    return UMAMI_SUCCESS;
+#endif
   }
 
   UmamiStatus umami_matrix_element(
@@ -338,6 +374,7 @@ extern "C"
     const double* random_color_in = nullptr;
     const double* random_helicity_in = nullptr;
     const double* random_diagram_in = nullptr;
+    const int* helicity_in = nullptr;
     [[maybe_unused]] const int* diagram_in = nullptr; // TODO: unused
 
     for( std::size_t i = 0; i < input_count; ++i )
@@ -364,7 +401,8 @@ extern "C"
           random_diagram_in = static_cast<const double*>( input );
           break;
         case UMAMI_IN_HELICITY_INDEX:
-          return UMAMI_ERROR_UNSUPPORTED_INPUT;
+          helicity_in = static_cast<const int*>( input );
+          break;
         case UMAMI_IN_DIAGRAM_INDEX:
           diagram_in = static_cast<const int*>( input );
           break;
@@ -373,6 +411,20 @@ extern "C"
       }
     }
     if( !momenta_in ) return UMAMI_ERROR_MISSING_INPUT;
+    // UMAMI_IN_HELICITY_INDEX: evaluate each event for the one helicity it is given, an index
+    // into the full helicity table [0,ncomb) (the UMAMI_OUT_HELICITY_INDEX convention), instead
+    // of summing over the helicities. The |M|^2 is then the contribution of that helicity to
+    // the helicity-summed (and averaged) |M|^2, so that the sum over all ncomb helicities gives
+    // back the |M|^2 without this input; UMAMI_OUT_HELICITY_INDEX returns the given helicity,
+    // the color and diagram are chosen for it, and UMAMI_IN_RANDOM_HELICITY is not used.
+    if( helicity_in )
+    {
+      for( std::size_t i_event = 0; i_event < count; ++i_event )
+      {
+        const int ihel = helicity_in[i_event + offset];
+        if( ihel < 0 || ihel >= ProcessData::ncomb ) return UMAMI_ERROR;
+      }
+    }
 
     double* m2_out = nullptr;
     double* amp2_out = nullptr;
@@ -459,6 +511,7 @@ extern "C"
     HostBufferBase<fptype_amp, false> denominators( rounded_count );
     HostBufferBase<int, false> helicity_index( rounded_count );
     HostBufferBase<int, false> color_index( rounded_count );
+    HostBufferBase<int, false> helicity_given( rounded_count ); // as the other inputs, the padding lanes copy their source event
     // The input event of each lane. The padding lanes, which complete the last page
     // of a flavor, are copies of the first event of their page, as on GPUs. Left at
     // zero (momenta and g_s), they evaluated to nan, and the good-helicity filter of
@@ -485,6 +538,7 @@ extern "C"
       diagram_random[lane] = random_diagram_in ? random_diagram_in[i_event + offset] : 0.5;
       g_s[lane] = alpha_s_in ? sqrt( 4 * M_PI * alpha_s_in[i_event + offset] ) : 1.2177157847767195;
       flavor_indices[lane] = flavor_indices_in ? flavor_indices_in[i_event + offset] : 0;
+      if( helicity_in ) helicity_given[lane] = helicity_in[i_event + offset];
     }
     computeDependentCouplings( g_s.data(), couplings.data(), rounded_count );
 
@@ -522,7 +576,8 @@ extern "C"
       // an interference |M|^2, see select_helicity_signed, unless the instance was asked
       // to keep the helicity sum (interference_helicity_summed); without the random number
       // the |M|^2 stays the helicity sum (systematics re-evaluation, standalone checks)
-      random_helicity_in != nullptr && !instance->summed_helicity );
+      random_helicity_in != nullptr && !instance->summed_helicity,
+      helicity_in ? helicity_given.data() : nullptr );
 
     if ( sort_flavors )
     {
diff --git a/madgraph/iolibs/template_files/madmatrix/backend/gpu/SigmaKin.cc b/madgraph/iolibs/template_files/madmatrix/backend/gpu/SigmaKin.cc
index fe4c8cdfba..4c1e330d2a 100644
--- a/madgraph/iolibs/template_files/madmatrix/backend/gpu/SigmaKin.cc
+++ b/madgraph/iolibs/template_files/madmatrix/backend/gpu/SigmaKin.cc
@@ -183,18 +183,22 @@ namespace madmatrix
   // Also compute running sums over helicities adding jamp2, numerator, denominator
   // (NB: this function no longer handles matrix elements as the color sum has now been
   // moved to a separate function/kernel). This is a kernel function: it processes a
-  // single event (the CUDA thread), and takes a channelId array as input.
+  // single event (the CUDA thread), and takes a channelId array as input. The helicity
+  // may differ from one thread to the next (allihel): the external wavefunctions are
+  // branch-free in the helicity (see helas.h), so the threads of a warp do not diverge.
   __global__ void /* clang-format off */
-  calculate_jamps( int ihel,
+  calculate_jamps( int ihel,                           // input: helicity index [0,ncomb) of all events (if !processAllHelicities and allihel == nullptr)
+                   const int* allihel,                 // input: helicity index [0,ncomb) of each event[nevt] (nullptr: ihel, or the good helicity blockIdx.y)
                    const fptype_momenta* allmomenta,   // input: momenta[nevt*npar*4]
                    const fptype* allcouplings,         // input: couplings[nevt*ndcoup*2]
                    const unsigned int* iflavorVec,     // input: indices of the flavor combinations
-                   fptype_amp* allJamps,               // output: jamp[2*njampso*nevt] buffer for one helicity _within a super-buffer for dcNGoodHel helicities_
+                   fptype_amp* allJamps,               // output: jamp[2*njampso*nevt] buffer for one helicity _within a super-buffer for nhel helicities_
                    bool storeChannelWeights,
                    fptype_amp* allNumerators,          // input/output: multichannel numerators[nevt], add helicity ihel
                    fptype_amp* allDenominators,        // input/output: multichannel denominators[nevt], add helicity ihel
                    fptype_amp* colAllJamp2s,           // output: allJamp2s[ncolor_flow][nevt] super-buffer, sum over col/hel (nullptr to disable)
                    const int nevt,                     // input: #events (nevt == ndim == gpublocks*gputhreads)
+                   const int nhel,                     // input: #helicities in the allJamps super-buffer (nGoodHel, or 1 with allihel)
                    const bool processAllHelicities )   // input: if true, use blockIdx.y to index helicities
   /* clang-format on */
   {
@@ -208,15 +212,21 @@ namespace madmatrix
     using NUM_ACCESS = DeviceAccessNumerators;    // non-trivial access: buffer includes all events
     mgDebug( 0, __FUNCTION__ );
 
+    const int ievt = blockDim.x * blockIdx.x + threadIdx.x; // index of event (thread) in grid
     if( processAllHelicities )
     {
       int ighel = blockIdx.y;
-      ihel = dcGoodHel[ighel];
+      if( allihel == nullptr ) ihel = dcGoodHel[ighel];
       allJamps = allJamps + ighel * nevt;
       // NB: the numerators buffer has NO helicity dimension anymore: all good-helicity blocks
       // for a given event accumulate in place into the same [nevt][ndiagrams] slot via atomicAdd.
       // The denominators are no longer accumulated here (derived as the sum of numerators later).
     }
+    // One helicity per event (nhel == 1 and gridDim.y == 1): the given helicity of this event
+    if( allihel != nullptr ) ihel = allihel[ievt];
+    // The helicities of the external legs for this event, as the external wavefunctions take them
+    fptype_amp_sv hel_sv[npar];
+    for( int ipar = 0; ipar < npar; ipar++ ) hel_sv[ipar] = cHel[ihel][ipar];
 
     // Local TEMPORARY variables for a subset of Feynman diagrams in the given CUDA event
     // (ievt) [NB these variables are reused several times (and re-initialised each time)
@@ -271,7 +281,6 @@ namespace madmatrix
     const fptype_momenta* momenta = allmomenta;
     const fptype* COUPs[nxcoup > 0 ? nxcoup : 1];
     for( size_t ixcoup = 0; ixcoup < nxcoup; ixcoup++ ) COUPs[ixcoup] = allCOUPs[ixcoup];
-    const int ievt = blockDim.x * blockIdx.x + threadIdx.x; // index of event (thread) in grid
     fptype_amp* numerators = &allNumerators[ievt * ndiagrams];
     // Create an array of views over the Flavor Couplings
     FLV_COUPLING_ARRAY<nIPF, nMF> flvCOUPs{ cIPF_partner1, cIPF_partner2, cIPF_value };
@@ -323,13 +332,13 @@ namespace madmatrix
 
     // *** PREPARE OUTPUT JAMPS ***
     // allJamps already points at this helicity's slot (ighel * nevt) in the
-    // [2][ncolor][dcNGoodHel][nevt] super-buffer, hence ihel0 = 0; nhel stays
-    // dcNGoodHel because it is the color stride, and color_sum reads with it.
+    // [2][ncolor][nhel][nevt] super-buffer, hence ihel0 = 0; nhel stays the
+    // number of helicities because it is the color stride, and color_sum reads with it.
     {
       constexpr int ihel0 = 0;
       using J_ACCESS = DeviceAccessJamp;
       for( int icol = 0; icol < njampso; icol++ )
-        J_ACCESS::kernelAccessIcolIhelNhel( allJamps, icol, ihel0, dcNGoodHel ) = jamp_sv[icol];
+        J_ACCESS::kernelAccessIcolIhelNhel( allJamps, icol, ihel0, nhel ) = jamp_sv[icol];
     }
 
     mgDebug( 1, __FUNCTION__ );
@@ -370,12 +379,11 @@ namespace madmatrix
       const int gpublocks = 1;
       const int gputhreads = maxtry;
       constexpr int nOneHel = 1; // use a jamp buffer for a single helicity
-      gpuMemcpyToSymbol( dcNGoodHel, &nOneHel, sizeof( int ) );
       // NEW IMPLEMENTATION OF GETGOODHEL (#630): RESET THE RUNNING SUM OVER HELICITIES TO 0 BEFORE ADDING A NEW HELICITY
       gpuMemset( allMEs, 0, maxtry * sizeof( fptype ) );
       // NB: color_sum ADDS |M|^2 for one helicity to the running sum of |M|^2 over helicities for the given event(s)
       constexpr fptype_amp_sv* allJamp2s = nullptr; // no need for color selection during helicity filtering
-      gpuLaunchKernel( calculate_jamps, gpublocks, gputhreads, ihel, allmomenta, allcouplings, devFlavorVec, allJamps, false, allNumerators, allDenominators, allJamp2s, gpublocks * gputhreads, false );
+      gpuLaunchKernel( calculate_jamps, gpublocks, gputhreads, ihel, nullptr, allmomenta, allcouplings, devFlavorVec, allJamps, false, allNumerators, allDenominators, allJamp2s, gpublocks * gputhreads, nOneHel, false );
       gpuLaunchKernel( color_sum_kernel, gpublocks, gputhreads, allMEs, allJamps, nOneHel, 0 );
       gpuMemcpy( hstMEs, allMEs, maxtry * sizeof( fptype ), gpuMemcpyDeviceToHost );
       for( int ievt = 0; ievt < maxtry; ++ievt )
@@ -518,9 +526,18 @@ namespace madmatrix
                       fptype* ghelAllMEs,               // input/tmp: allMEs for nGoodHel <= ncomb individual/runningsum helicities (index is ighel)
                       fptype* allMEs,                   // output: allMEs[nevt], final sum over helicities
                       const int nevt,                   // input: #events (for cuda: nevt == ndim == gpublocks*gputhreads)
-                      const bool sampleSignedHelicity ) // input: see sigmaKin
+                      const bool sampleSignedHelicity,  // input: see sigmaKin
+                      const int* allihel )              // input: helicity index [0,ncomb) of each event[nevt] (nullptr: choose one)
   {
     const int ievt = blockDim.x * blockIdx.x + threadIdx.x; // index of event (thread)
+    if( allihel != nullptr )
+    {
+      // No sum and no choice: each event was evaluated for its given helicity only, whose
+      // |M|^2 (signed, with an interference) is the one entry of ghelAllMEs (nhel == 1)
+      allMEs[ievt] = ghelAllMEs[ievt];
+      allselhel[ievt] = allihel[ievt] + 1; // NB Fortran [1,ncomb], cudacpp [0,ncomb-1]
+      return;
+    }
     // no helicity (0) unless one is chosen below: a vanishing or NaN |M|^2 chooses none, and
     // allselhel is not cleared beforehand (the CPU buffers are zero-initialised)
     allselhel[ievt] = 0;
@@ -690,7 +707,8 @@ namespace madmatrix
             const bool async,
             const int gpublocks,                // input: cuda gpublocks
             const int gputhreads,               // input: cuda gputhreads
-            const bool sampleSignedHelicity )   // input: choose the helicity on |T_i| and return sign(T_i)*sum|T| (split orders only)
+            const bool sampleSignedHelicity,    // input: choose the helicity on |T_i| and return sign(T_i)*sum|T| (split orders only)
+            const int* allihel )                // input: helicity index [0,ncomb) of each event[nevt]; nullptr to sum over the good helicities
   /* clang-format on */
   {
     mgDebugInitialise();
@@ -711,15 +729,19 @@ namespace madmatrix
     // Reset the "matrix elements" - running sums of |M|^2 over helicities for the given event
     // (the sizes are computed in size_t, from sizeof: the int product could overflow)
     const int nevt = gpublocks * gputhreads;
+    // With allihel, each event has a single helicity of its own instead of the sum over the good
+    // helicities: one evaluation (nhel == 1) in which the threads carry different helicities in
+    // lock-step, and the color sum, the color and channel choice all run on that one helicity.
+    const int nhel = ( allihel != nullptr ? 1 : cNGoodHel );
     gpuMemsetAsync( allMEs, 0, sizeof( fptype ) * nevt, stream );
-    gpuMemsetAsync( ghelAllJamps, 0, sizeof( fptype_amp ) * cNGoodHel * njampso * mgOnGpu::nx2 * nevt, stream );
+    gpuMemsetAsync( ghelAllJamps, 0, sizeof( fptype_amp ) * nhel * njampso * mgOnGpu::nx2 * nevt, stream );
     gpuMemsetAsync( colAllJamp2s, 0, sizeof( fptype_amp ) * ncolor_flow * nevt, stream );
     // The numerators buffer has NO helicity dimension: all good helicities accumulate in place via
     // atomicAdd, so it is zeroed once as [nevt][ndiagrams]. The denominators are derived from the
     // numerators in normalise_output, so the buffer is just [nevt].
     gpuMemsetAsync( ghelAllNumerators, 0, sizeof( fptype_amp ) * ndiagrams * nevt, stream );
     gpuMemsetAsync( ghelAllDenominators, 0, sizeof( fptype_amp ) * nevt, stream );
-    gpuMemsetAsync( ghelAllMEs, 0, sizeof( fptype ) * cNGoodHel * nevt, stream );
+    gpuMemsetAsync( ghelAllMEs, 0, sizeof( fptype ) * nhel * nevt, stream );
 
     // === PART 1 - HELICITY LOOP: CALCULATE WAVEFUNCTIONS (one event per GPU thread) ===
 
@@ -730,28 +752,28 @@ namespace madmatrix
     // every Wilson coefficient at zero): nothing to compute, the |M|^2 stays 0 as on the CPU (a
     // launch with gridDim.y = 0 is an invalid configuration)
     bool storeChannelWeights = allChannelIds != nullptr || allrnddiagram != nullptr;
-    if( cNGoodHel > 0 && async )
+    if( nhel > 0 && async )
     {
-      gpuLaunchKernel2D( calculate_jamps, gpublocks, cNGoodHel, gputhreads, ghelStreams[0], 0, allmomenta, allcouplings, iflavorVec, ghelAllJamps, storeChannelWeights, ghelAllNumerators, ghelAllDenominators, colAllJamp2s, nevt, true );
-      color_sum_gpu( ghelAllMEs, ghelAllJamps, ghelAllBlasTmp, pBlasHandle, ghelStreams, cNGoodHel, gpublocks, gputhreads, true );
+      gpuLaunchKernel2D( calculate_jamps, gpublocks, nhel, gputhreads, ghelStreams[0], 0, allihel, allmomenta, allcouplings, iflavorVec, ghelAllJamps, storeChannelWeights, ghelAllNumerators, ghelAllDenominators, colAllJamp2s, nevt, nhel, true );
+      color_sum_gpu( ghelAllMEs, ghelAllJamps, ghelAllBlasTmp, pBlasHandle, ghelStreams, nhel, gpublocks, gputhreads, true );
     }
-    else if( cNGoodHel > 0 )
+    else if( nhel > 0 )
     {
-      for( int ighel = 0; ighel < cNGoodHel; ighel++ )
+      for( int ighel = 0; ighel < nhel; ighel++ )
       {
-        const int ihel = cGoodHel[ighel];
+        const int ihel = ( allihel != nullptr ? 0 : cGoodHel[ighel] ); // unused with allihel
         fptype_amp* hAllJamps = ghelAllJamps + ighel * nevt; // HACK: bypass DeviceAccessJamp (consistent with layout defined there)
         // NB: the numerators buffer has no helicity dimension: every helicity stream accumulates in place
         // into the same [nevt][ndiagrams] slot via atomicAdd. The denominators are derived later.
-        gpuLaunchKernelStream( calculate_jamps, gpublocks, gputhreads, ghelStreams[ighel], ihel, allmomenta, allcouplings, iflavorVec, hAllJamps, storeChannelWeights, ghelAllNumerators, ghelAllDenominators, colAllJamp2s, nevt, false );
+        gpuLaunchKernelStream( calculate_jamps, gpublocks, gputhreads, ghelStreams[ighel], ihel, allihel, allmomenta, allcouplings, iflavorVec, hAllJamps, storeChannelWeights, ghelAllNumerators, ghelAllDenominators, colAllJamp2s, nevt, nhel, false );
       }
       // (2) Then compute the ME for that helicity from the color sum of QCD partial amplitudes jamps
-      color_sum_gpu( ghelAllMEs, ghelAllJamps, ghelAllBlasTmp, pBlasHandle, ghelStreams, cNGoodHel, gpublocks, gputhreads, false );
+      color_sum_gpu( ghelAllMEs, ghelAllJamps, ghelAllBlasTmp, pBlasHandle, ghelStreams, nhel, gpublocks, gputhreads, false );
       checkGpu( gpuDeviceSynchronize() ); // do not start helicity/color selection until the loop over helicities has completed
       // (3) Wait for all helicity streams to complete, then finally compute the ME sum over all helicities and choose one helicity and one color
     }
     // Event-by-event random choice of helicity #403 and ME sum over helicities (defer this after the helicity loop to avoid breaking streams parallelism)
-    gpuLaunchKernelStream( add_and_select_hel, gpublocks, gputhreads, stream, allselhel, allrndhel, ghelAllMEs, allMEs, gpublocks * gputhreads, sampleSignedHelicity );
+    gpuLaunchKernelStream( add_and_select_hel, gpublocks, gputhreads, stream, allselhel, allrndhel, ghelAllMEs, allMEs, gpublocks * gputhreads, sampleSignedHelicity, allihel );
 
     gpuLaunchKernelStream( normalise_output, gpublocks, gputhreads, stream, allMEs, iflavorVec, ghelAllNumerators, ghelAllDenominators, allChannelIds, storeChannelWeights, mulChannelWeight, helcolDenominators[0] );
 
diff --git a/madgraph/iolibs/template_files/madmatrix/backend/gpu/SigmaKin.h b/madgraph/iolibs/template_files/madmatrix/backend/gpu/SigmaKin.h
index dad3d88bbb..83fb696625 100644
--- a/madgraph/iolibs/template_files/madmatrix/backend/gpu/SigmaKin.h
+++ b/madgraph/iolibs/template_files/madmatrix/backend/gpu/SigmaKin.h
@@ -65,7 +65,11 @@ namespace madmatrix
             // input: when the |M|^2 keeps an interference (squared split orders, nampso > 1),
             // choose the helicity on |T_i| and return sign(T_i) * sum_j |T_j| as madevent does,
             // so that the chosen helicity is exact; false returns sum_j T_j (no effect otherwise)
-            const bool sampleSignedHelicity = false );
+            const bool sampleSignedHelicity = false,
+            // input: the helicity index [0,ncomb) of each event (a device array), which is then
+            // evaluated for that helicity alone (and returns its contribution to the helicity-summed
+            // |M|^2, see umami_matrix_element); nullptr, the default, sums over the good helicities
+            const int* allihel = nullptr );
 
   // Setters: called once by CPPProcess (P1-generated) to populate this file's
   // otherwise-internal storage, since it can no longer be written directly
diff --git a/madgraph/iolibs/template_files/madmatrix/backend/gpu/umami.cc b/madgraph/iolibs/template_files/madmatrix/backend/gpu/umami.cc
index 74335b434b..9d46212f3e 100644
--- a/madgraph/iolibs/template_files/madmatrix/backend/gpu/umami.cc
+++ b/madgraph/iolibs/template_files/madmatrix/backend/gpu/umami.cc
@@ -17,6 +17,7 @@
 #include <algorithm>
 #include <cfloat>
 #include <cmath>
+#include <cstdio>
 #include <string>
 #include <vector>
 #include <array>
@@ -120,12 +121,14 @@ namespace
     const double* diagram_random_in,
     const double* alpha_s_in,
     const unsigned int* flavor_indices_in,
+    const int* helicity_in,
     fptype_momenta* momenta,
     fptype* helicity_random,
     fptype* color_random,
     fptype* diagram_random,
     fptype* g_s,
     unsigned int* flavor_indices,
+    int* helicity_given,
     std::size_t count,
     std::size_t stride,
     std::size_t offset )
@@ -144,6 +147,14 @@ namespace
     color_random[i_event] = color_random_in ? color_random_in[i_in + offset] : 0.5;
     g_s[i_event] = alpha_s_in ? sqrt( 4 * M_PI * alpha_s_in[i_in + offset] ) : 1.2177157847767195;
     flavor_indices[i_event] = flavor_indices_in ? flavor_indices_in[i_in + offset] : 0;
+    if( helicity_in )
+    {
+      // an index out of [0,ncomb) cannot be reported from here (umami_matrix_element returns
+      // before the kernels run): evaluate helicity 0 instead, never out of the helicity table,
+      // and let copy_outputs flag the event (see umami_matrix_element)
+      const int ihel = helicity_in[i_in + offset];
+      helicity_given[i_event] = ( ihel >= 0 && ihel < ProcessData::ncomb ? ihel : 0 );
+    }
   }
 
   __global__ void copy_outputs(
@@ -158,12 +169,24 @@ namespace
     int* diagram_out,
     int* color_out,
     int* helicity_out,
+    const int* helicity_in,
     std::size_t count,
     std::size_t stride,
     std::size_t offset )
   {
     std::size_t i_event = blockDim.x * blockIdx.x + threadIdx.x;
     if( i_event >= count ) return;
+    if( helicity_in )
+    {
+      const int ihel = helicity_in[i_event + offset];
+      if( ihel < 0 || ihel >= ProcessData::ncomb )
+      {
+        // an invalid given helicity: no |M|^2 and no helicity, see umami_matrix_element
+        if( m2_out ) m2_out[i_event + offset] = 0;
+        if( helicity_out ) helicity_out[i_event + offset] = -1;
+        return;
+      }
+    }
 
     if( m2_out ) m2_out[i_event + offset] = matrix_elements[i_event];
     if( amp2_out )
@@ -333,8 +356,8 @@ extern "C"
   UmamiStatus umami_supported_inputs( bool const** supported, int* count )
   {
     // MOMENTA, ALPHA_S, FLAVOR_INDEX, RANDOM_COLOR, RANDOM_HELICITY, RANDOM_DIAGRAM,
-    // HELICITY_INDEX=false, DIAGRAM_INDEX=true, CHANNEL_INDEX=false
-    static const bool data[UMAMI_INPUT_KEY_COUNT] = { true, true, true, true, true, true, false, true };
+    // HELICITY_INDEX, DIAGRAM_INDEX=true, CHANNEL_INDEX=false
+    static const bool data[UMAMI_INPUT_KEY_COUNT] = { true, true, true, true, true, true, true, true };
     *supported = data;
     *count = UMAMI_INPUT_KEY_COUNT;
     return UMAMI_SUCCESS;
@@ -397,7 +420,36 @@ extern "C"
       static_cast<InterfaceInstance*>( handle )->summed_helicity = parameter_real != 0;
       return UMAMI_SUCCESS;
     }
+    // Parameters are process-wide (Parameters::getInstance()), not per-handle,
+    // matching how umami_initialize itself never stores the CPPProcess it
+    // constructs: initProc() only ever populates the Parameters singleton.
+#ifdef MGONGPU_HARDCODE_PARAM
+    // Hardcoded parameters are compile-time constants: there is nothing to set.
     return UMAMI_ERROR_NOT_IMPLEMENTED;
+#else
+    if( !Parameters::getInstance()->setParameterByName( name, parameter_real, parameter_imag ) )
+      return UMAMI_ERROR_UNKNOWN_PARAMETER;
+    // sigmaKin() and friends read cached copies of the independent parameters/couplings
+    // (cIPD/cIPC/cIPF/bsmIndepParam), not the Parameters singleton itself, for
+    // performance; refresh those caches now so the override actually takes effect.
+    CPPProcess::refreshIndependentParams();
+    // Keep g_externalMasses in sync with the (possibly updated) singleton.
+    // Emit a one-time warning if a mass parameter was changed, since any phase-space
+    // generator that cached masses from a prior umami_get_meta(UMAMI_META_MASSES,...) call
+    // will need to re-query them.
+    {
+      std::vector<fptype> updatedMasses;
+      CPPProcess::refreshExternalMasses( updatedMasses );
+      bool changed = ( updatedMasses.size() != g_externalMasses.size() );
+      for( size_t i = 0; !changed && i < updatedMasses.size(); ++i )
+        changed = ( static_cast<double>( updatedMasses[i] ) != g_externalMasses[i] );
+      if( changed )
+      {
+        g_externalMasses.assign( updatedMasses.begin(), updatedMasses.end() );
+      }
+    }
+    return UMAMI_SUCCESS;
+#endif
   }
 
   UmamiStatus umami_get_parameter(
@@ -406,7 +458,13 @@ extern "C"
     [[maybe_unused]] double* parameter_real,
     [[maybe_unused]] double* parameter_imag )
   {
+#ifdef MGONGPU_HARDCODE_PARAM
     return UMAMI_ERROR_NOT_IMPLEMENTED;
+#else
+    if( !Parameters::getInstance()->getParameterByName( name, parameter_real, parameter_imag ) )
+      return UMAMI_ERROR_UNKNOWN_PARAMETER;
+    return UMAMI_SUCCESS;
+#endif
   }
 
   UmamiStatus umami_matrix_element(
@@ -427,6 +485,7 @@ extern "C"
     const double* random_color_in = nullptr;
     const double* random_helicity_in = nullptr;
     const double* random_diagram_in = nullptr;
+    const int* helicity_in = nullptr;
     [[maybe_unused]] const int* diagram_in = nullptr; // TODO: unused
 
     for( std::size_t i = 0; i < input_count; ++i )
@@ -453,7 +512,8 @@ extern "C"
           random_diagram_in = static_cast<const double*>( input );
           break;
         case UMAMI_IN_HELICITY_INDEX:
-          return UMAMI_ERROR_UNSUPPORTED_INPUT;
+          helicity_in = static_cast<const int*>( input );
+          break;
         case UMAMI_IN_DIAGRAM_INDEX:
           diagram_in = static_cast<const int*>( input );
           break;
@@ -462,6 +522,14 @@ extern "C"
       }
     }
     if( !momenta_in ) return UMAMI_ERROR_MISSING_INPUT;
+    // UMAMI_IN_HELICITY_INDEX: evaluate each event for the one helicity it is given, an index
+    // into the full helicity table [0,ncomb) (the UMAMI_OUT_HELICITY_INDEX convention), instead
+    // of summing over the helicities. The |M|^2 is then the contribution of that helicity to
+    // the helicity-summed (and averaged) |M|^2, so that the sum over all ncomb helicities gives
+    // back the |M|^2 without this input; UMAMI_OUT_HELICITY_INDEX returns the given helicity,
+    // the color and diagram are chosen for it, and UMAMI_IN_RANDOM_HELICITY is not used. The
+    // indices live on the device and are only read by the kernels, so that an index out of
+    // [0,ncomb) cannot fail the call as on the CPU: that event gets |M|^2 = 0 and helicity -1.
 
     gpuStream_t gpu_stream = nullptr;
     double* m2_out = nullptr;
@@ -507,11 +575,11 @@ extern "C"
     fptype *couplings, *g_s, *helicity_random, *color_random, *diagram_random;
     fptype *matrix_elements, *ghel_matrix_elements;
     fptype_amp *color_jamps, *ghel_jamps;
-    int *helicity_index, *color_index;
+    int *helicity_index, *color_index, *helicity_given;
     unsigned int *flavor_indices, *diagram_index;
 
     std::size_t n_coup = madmatrix::Parameters_dependentCouplings::ndcoup;
-    std::array<std::pair<void**, std::size_t>, 16> ptrs_and_sizes = {{
+    std::array<std::pair<void**, std::size_t>, 17> ptrs_and_sizes = {{
         {reinterpret_cast<void**>(&momenta), rounded_count * ProcessData::npar * 4 * sizeof( fptype_momenta )},
         {reinterpret_cast<void**>(&couplings), rounded_count * n_coup * 2 * sizeof( fptype )},
         {reinterpret_cast<void**>(&g_s), rounded_count * sizeof( fptype )},
@@ -530,6 +598,7 @@ extern "C"
         {reinterpret_cast<void**>(&denominators), rounded_count * sizeof( fptype_amp )},
         {reinterpret_cast<void**>(&helicity_index), rounded_count * sizeof( int )},
         {reinterpret_cast<void**>(&color_index), rounded_count * sizeof( int )},
+        {reinterpret_cast<void**>(&helicity_given), rounded_count * sizeof( int )},
         {reinterpret_cast<void**>(&ghel_matrix_elements), rounded_count * ProcessData::ncomb * sizeof( fptype )},
         // njampso = ncolor per amplitude split order (just ncolor without split orders)
         {reinterpret_cast<void**>(&ghel_jamps), rounded_count * ProcessData::ncomb * ProcessData::njampso * mgOnGpu::nx2 * sizeof( fptype_amp )},
@@ -557,12 +626,14 @@ extern "C"
       random_diagram_in,
       alpha_s_in,
       flavor_indices_in,
+      helicity_in,
       momenta,
       helicity_random,
       color_random,
       diagram_random,
       g_s,
       flavor_indices,
+      helicity_given,
       count,
       stride,
       offset );
@@ -611,7 +682,8 @@ extern "C"
       // an interference |M|^2 (see add_and_select_hel), unless the instance was asked to
       // keep the helicity sum (interference_helicity_summed); without the random number
       // the |M|^2 stays the helicity sum (systematics re-evaluation, standalone checks)
-      random_helicity_in != nullptr && !instance->summed_helicity );
+      random_helicity_in != nullptr && !instance->summed_helicity,
+      helicity_in ? helicity_given : nullptr );
 
     copy_outputs<<<n_blocks, n_threads, 0, gpu_stream>>>(
       denominators,
@@ -625,6 +697,7 @@ extern "C"
       diagram_out,
       color_out,
       helicity_out,
+      helicity_in,
       count,
       stride,
       offset );
diff --git a/madgraph/iolibs/template_files/madmatrix/backend/simd/SigmaKin.cc b/madgraph/iolibs/template_files/madmatrix/backend/simd/SigmaKin.cc
index bf75465f03..1a84d81d36 100644
--- a/madgraph/iolibs/template_files/madmatrix/backend/simd/SigmaKin.cc
+++ b/madgraph/iolibs/template_files/madmatrix/backend/simd/SigmaKin.cc
@@ -222,21 +222,21 @@ namespace madmatrix
 
   template<class M_ACCESS, class W_ACCESS>
   inline void
-  ixxxxx( const fptype_momenta momenta[], const fptype_amp_v fmass, const int nhel, const int nsf, const int flv, ALOHAOBJ& fi, const int ipar )
+  ixxxxx( const fptype_momenta momenta[], const fptype_amp_v fmass, const fptype_amp_sv nhel, const int nsf, const int flv, ALOHAOBJ& fi, const int ipar )
   {
     wavefunction_per_lane_mass<W_ACCESS>( fmass, fi, [&]( const fptype_amp m, ALOHAOBJ& out ) { ixxxxx<M_ACCESS, W_ACCESS>( momenta, m, nhel, nsf, flv, out, ipar ); } );
   }
 
   template<class M_ACCESS, class W_ACCESS>
   inline void
-  oxxxxx( const fptype_momenta momenta[], const fptype_amp_v fmass, const int nhel, const int nsf, const int flv, ALOHAOBJ& fo, const int ipar )
+  oxxxxx( const fptype_momenta momenta[], const fptype_amp_v fmass, const fptype_amp_sv nhel, const int nsf, const int flv, ALOHAOBJ& fo, const int ipar )
   {
     wavefunction_per_lane_mass<W_ACCESS>( fmass, fo, [&]( const fptype_amp m, ALOHAOBJ& out ) { oxxxxx<M_ACCESS, W_ACCESS>( momenta, m, nhel, nsf, flv, out, ipar ); } );
   }
 
   template<class M_ACCESS, class W_ACCESS>
   inline void
-  vxxxxx( const fptype_momenta momenta[], const fptype_amp_v vmass, const int nhel, const int nsv, const int flv, ALOHAOBJ& vc, const int ipar )
+  vxxxxx( const fptype_momenta momenta[], const fptype_amp_v vmass, const fptype_amp_sv nhel, const int nsv, const int flv, ALOHAOBJ& vc, const int ipar )
   {
     wavefunction_per_lane_mass<W_ACCESS>( vmass, vc, [&]( const fptype_amp m, ALOHAOBJ& out ) { vxxxxx<M_ACCESS, W_ACCESS>( momenta, m, nhel, nsv, flv, out, ipar ); } );
   }
@@ -257,6 +257,29 @@ namespace madmatrix
 
   //--------------------------------------------------------------------------
 
+  // The helicities of the npar external legs for the events of one page (neppV events,
+  // one SIMD lane each, or one event without SIMD), as the external wavefunctions take
+  // them: the helicities hel[npar] for every event, or the helicity allihel[ievt] of each event.
+  static inline void
+  fillHelicities( fptype_amp_sv* hel_sv,  // output: hel_sv[npar] helicities of the external legs
+                  const short* hel,       // input: helicity of each leg, hel[npar], of all events (if allihel == nullptr)
+                  const int* allihel,     // input: helicity index [0,ncomb) of each event[nevt] (nullptr: hel for all events)
+                  const int ievt0 )       // input: first event number in current C++ event page
+  {
+#ifdef MGONGPU_CPPSIMD
+    for( int ieppV = 0; ieppV < neppV; ieppV++ )
+    {
+      const short* heli = ( allihel != nullptr ? cHel[allihel[ievt0 + ieppV]] : hel );
+      for( int ipar = 0; ipar < npar; ipar++ ) hel_sv[ipar][ieppV] = heli[ipar];
+    }
+#else
+    const short* heli = ( allihel != nullptr ? cHel[allihel[ievt0]] : hel );
+    for( int ipar = 0; ipar < npar; ipar++ ) hel_sv[ipar] = heli[ipar];
+#endif
+  }
+
+  //--------------------------------------------------------------------------
+
   // Accumulate a multichannel numerator contribution in place. In C++ each
   // event page is processed serially within the helicity loop, so a plain
   // sum suffices (CUDA needs atomicAdd instead: see backend/gpu/SigmaKin.cc).
@@ -268,9 +291,12 @@ namespace madmatrix
   // moved to a separate function/kernel). This function processes a single event "page"
   // or SIMD vector (or for two in "mixed" precision mode, nParity=2). Accepts a SCALAR
   // channelId because it is GUARANTEED that all events in a SIMD vector have the same
-  // channelId #898.
+  // channelId #898. The helicity, on the other hand, may differ from one event (SIMD lane)
+  // to the next: the external wavefunctions are branch-free in the helicity (see helas.h),
+  // so that the events of a page with different helicities are computed in lock-step.
   void
-  calculate_jamps( const short* hel,                 // input: helicity of each leg, hel[npar] (a row of cHel, or not: see Density.cc)
+  calculate_jamps( const short* hel,                   // input: helicity of each leg, hel[npar], of all events (if allihel == nullptr): a row of cHel, or not (see Density.cc)
+                   const int* allihel,                 // input: helicity index [0,ncomb) of each event[nevt] (nullptr: hel for all events)
                    const fptype_momenta* allmomenta,   // input: momenta[nevt*npar*4]
                    const fptype* allcouplings,         // input: couplings[nevt*ndcoup*2]
                    const unsigned int* iflavorVec,     // input: indices of the flavor combinations
@@ -291,16 +317,6 @@ namespace madmatrix
     using NUM_ACCESS = HostAccessNumerators;    // non-trivial access: buffer includes all events
     mgDebug( 0, __FUNCTION__ );
 
-    // EvaluateDiagrams.inc reads the helicities as cHel[ihel][ipar]. Shadow the
-    // helicity table with the one row this call evaluates, so that the row need
-    // not be one of cHel at all: the density matrix builds combinations that a
-    // polarised process leaves out of its table (sigmaKin_density, Density.cc).
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wshadow"
-    const short( *cHel )[npar] = reinterpret_cast<const short( * )[npar]>( hel );
-#pragma GCC diagnostic pop
-    constexpr int ihel = 0;
-
     // Local TEMPORARY variables for a subset of Feynman diagrams in the given C++ event
     // page (ipagV) [NB these variables are reused several times (and re-initialised each
     // time) within the same event or event page]. Create memory for both momenta and
@@ -356,6 +372,10 @@ namespace madmatrix
       for( size_t iicoup = 0; iicoup < nIPC; iicoup++ )
         COUPs[ndcoup + iicoup] = allCOUPs[ndcoup + iicoup]; // independent couplings, fixed for all events
       fptype_amp* numerators = NUM_ACCESS::ieventAccessRecord( allNumerators, ievt0 * ndiagrams );
+      // The helicities of the external legs for the events of this page, one per SIMD lane:
+      // the same helicities hel for all of them, or the helicity allihel of each event
+      fptype_amp_sv hel_sv[npar];
+      fillHelicities( hel_sv, hel, allihel, ievt0 );
       // Create an array of views over the Flavor Couplings
       FLV_COUPLING_ARRAY<nIPF, nMF> flvCOUPs{ cIPF_partner1, cIPF_partner2, cIPF_value };
 
@@ -478,7 +498,7 @@ namespace madmatrix
 #else
         cxtype_amp_sv jamp_sv[njampso] = {}; // all zeros
 #endif
-        calculate_jamps( cHel[ihel], allmomenta, allcouplings, hgFlavorVec, jamp_sv, false, allNumerators, allDenominators, jamp2_sv, ievt00 );
+        calculate_jamps( cHel[ihel], nullptr, allmomenta, allcouplings, hgFlavorVec, jamp_sv, false, allNumerators, allDenominators, jamp2_sv, ievt00 );
         color_sum_cpu( allMEs, jamp_sv, ievt00 );
         for( int ieppV = 0; ieppV < neppV; ++ieppV )
         {
@@ -633,7 +653,8 @@ namespace madmatrix
             unsigned int* allDiagramIdsOut,    // output: multichannel channelIds[nevt] (1 to #diagrams)
             bool mulChannelWeight,             // if true, multiply channel weight to ME output
             const int nevt,                    // input: #events (for cuda: nevt == ndim == gpublocks*gputhreads)
-            const bool sampleSignedHelicity )  // input: choose the helicity on |T_i| and return sign(T_i)*sum|T| (split orders only)
+            const bool sampleSignedHelicity,   // input: choose the helicity on |T_i| and return sign(T_i)*sum|T| (split orders only)
+            const int* allihel )               // input: helicity index [0,ncomb) of each event[nevt]; nullptr to sum over the good helicities
   {
     mgDebugInitialise();
 
@@ -670,6 +691,10 @@ namespace madmatrix
 
     // === PART 1 - HELICITY LOOP: CALCULATE WAVEFUNCTIONS ===
     // (using precomputed good helicities)
+    // With allihel, each event has a single helicity of its own instead: the "loop" is one
+    // evaluation in which the events of a SIMD page carry different helicities in lock-step,
+    // and everything below (color sum, color and channel choice) runs on that one helicity.
+    const int nhelEval = ( allihel != nullptr ? 1 : cNGoodHel );
 #if defined MGONGPU_CPPSIMD and defined MGONGPU_FPTYPE_DOUBLE and defined MGONGPU_FPTYPE2_FLOAT
     // Mixed fptypes #537: float for color algebra and double elsewhere
     // Delay color algebra and ME updates (only on even pages)
@@ -680,7 +705,7 @@ namespace madmatrix
 #endif
 #ifdef _OPENMP
     // OMP multithreading #575 (NB: tested only with gcc11 so far)
-#define _OMPLIST0 allcouplings, allMEs, allmomenta, allrndcol, allrndhel, allselcol, allselhel, cGoodHel, cNGoodHel, npagV2, sampleSignedHelicity
+#define _OMPLIST0 allcouplings, allMEs, allmomenta, allrndcol, allrndhel, allselcol, allselhel, allihel, cGoodHel, nhelEval, npagV2, sampleSignedHelicity
 #define _OMPLIST1 , allDenominators, allNumerators, allChannelIds, mgOnGpu::icolamp, mgOnGpu::channel2iconfig
 #pragma omp parallel for default( none ) shared( _OMPLIST0 _OMPLIST1 )
 #undef _OMPLIST0
@@ -709,31 +734,35 @@ namespace madmatrix
       if( ColorMatrixData::shouldUseBlas )
       {
         static thread_local std::vector<cxtype_sv> ghelJamp_sv( (size_t)ncomb * nParity * ncolor );
-        for( int ighel = 0; ighel < cNGoodHel; ighel++ )
+        for( int ighel = 0; ighel < nhelEval; ighel++ )
         {
-          const int ihel = cGoodHel[ighel];
+          const int ihel = ( allihel != nullptr ? -1 : cGoodHel[ighel] ); // unused with allihel
           cxtype_sv* jamp_sv = ghelJamp_sv.data() + (size_t)ighel * nParity * ncolor;
           for( int i = 0; i < nParity * ncolor; i++ ) jamp_sv[i] = cxzero_sv<cxtype_sv>(); // calculate_jamps accumulates into jamp_sv
           // **NB! in "mixed" precision, using SIMD, calculate_jamps computes MEs for TWO neppV pages with a single channelId! #924
           bool storeChannelWeights = allChannelIds != nullptr || allrnddiagram != nullptr;
-          calculate_jamps( cHel[ihel], allmomenta, allcouplings, iflavorVec, jamp_sv, storeChannelWeights, allNumerators, allDenominators, jamp2_sv, ievt00 );
+          calculate_jamps( ( allihel != nullptr ? nullptr : cHel[ihel] ), allihel, allmomenta, allcouplings, iflavorVec, jamp_sv, storeChannelWeights, allNumerators, allDenominators, jamp2_sv, ievt00 );
         }
+        // nhelEval may be 0 (no good helicity): the color sum of no helicity leaves allMEs at 0
+        if( nhelEval > 0 )
+        {
 #if defined MGONGPU_CPPSIMD and defined MGONGPU_FPTYPE_DOUBLE and defined MGONGPU_FPTYPE2_FLOAT
-        color_sum_cpu_blas( allMEs, MEs_ighel, MEs_ighel2, ghelJamp_sv.data(), cNGoodHel, ievt00 );
+          color_sum_cpu_blas( allMEs, MEs_ighel, MEs_ighel2, ghelJamp_sv.data(), nhelEval, ievt00 );
 #else
-        color_sum_cpu_blas( allMEs, MEs_ighel, nullptr, ghelJamp_sv.data(), cNGoodHel, ievt00 );
+          color_sum_cpu_blas( allMEs, MEs_ighel, nullptr, ghelJamp_sv.data(), nhelEval, ievt00 );
 #endif
+        }
       }
       else
 #endif // MGONGPU_CPP_HAS_BLAS
       {
-        for( int ighel = 0; ighel < cNGoodHel; ighel++ )
+        for( int ighel = 0; ighel < nhelEval; ighel++ )
         {
-          const int ihel = cGoodHel[ighel];
+          const int ihel = ( allihel != nullptr ? -1 : cGoodHel[ighel] ); // unused with allihel
           cxtype_amp_sv jamp_sv[nParity * njampso] = {}; // fixed nasty bug (omitting 'nParity' caused memory corruptions after calling calculate_jamps)
           // **NB! in "mixed" precision, using SIMD, calculate_jamps computes MEs for TWO neppV pages with a single channelId! #924
           bool storeChannelWeights = allChannelIds != nullptr || allrnddiagram != nullptr;
-          calculate_jamps( cHel[ihel], allmomenta, allcouplings, iflavorVec, jamp_sv, storeChannelWeights, allNumerators, allDenominators, jamp2_sv, ievt00 );
+          calculate_jamps( ( allihel != nullptr ? nullptr : cHel[ihel] ), allihel, allmomenta, allcouplings, iflavorVec, jamp_sv, storeChannelWeights, allNumerators, allDenominators, jamp2_sv, ievt00 );
           color_sum_cpu( allMEs, jamp_sv, ievt00 );
           MEs_ighel[ighel] = E_ACCESS::kernelAccess( E_ACCESS::ieventAccessRecord( allMEs, ievt00 ) );
 #if defined MGONGPU_CPPSIMD and defined MGONGPU_FPTYPE_DOUBLE and defined MGONGPU_FPTYPE2_FLOAT
@@ -746,6 +775,16 @@ namespace madmatrix
       for( int ieppV = 0; ieppV < neppV; ++ieppV )
       {
         const int ievt = ievt00 + ieppV;
+        if( allihel != nullptr )
+        {
+          // No choice: each event was evaluated for its given helicity only, whose |M|^2 (signed,
+          // with an interference) is the color sum above
+          allselhel[ievt] = allihel[ievt] + 1; // NB Fortran [1,ncomb], cudacpp [0,ncomb-1]
+#if defined MGONGPU_CPPSIMD and defined MGONGPU_FPTYPE_DOUBLE and defined MGONGPU_FPTYPE2_FLOAT
+          allselhel[ievt + neppV] = allihel[ievt + neppV] + 1;
+#endif
+          continue;
+        }
         if constexpr( nampso > 1 )
         {
           if( sampleSignedHelicity )
diff --git a/madgraph/iolibs/template_files/madmatrix/backend/simd/SigmaKin.h b/madgraph/iolibs/template_files/madmatrix/backend/simd/SigmaKin.h
index 60a5af11bc..8e0d0f94ab 100644
--- a/madgraph/iolibs/template_files/madmatrix/backend/simd/SigmaKin.h
+++ b/madgraph/iolibs/template_files/madmatrix/backend/simd/SigmaKin.h
@@ -57,7 +57,11 @@ namespace madmatrix
             // input: when the |M|^2 keeps an interference (squared split orders, nampso > 1),
             // choose the helicity on |T_i| and return sign(T_i) * sum_j |T_j| as madevent does,
             // so that the chosen helicity is exact; false returns sum_j T_j (no effect otherwise)
-            const bool sampleSignedHelicity = false );
+            const bool sampleSignedHelicity = false,
+            // input: the helicity index [0,ncomb) of each event, which is then evaluated for that
+            // helicity alone (and returns its contribution to the helicity-summed |M|^2, see
+            // umami_matrix_element); nullptr, the default, sums over the good helicities
+            const int* allihel = nullptr );
 
   // Setters: called once by CPPProcess (P1-generated) to populate this file's
   // otherwise-internal storage, since it can no longer be written directly
@@ -75,7 +79,8 @@ namespace madmatrix
   // helicity table cHel[ncomb][npar] (row major) the combinations are built
   // from, and the identical-particle factor of a flavor combination.
   void
-  calculate_jamps( const short* hel,                 // input: helicity of each leg, hel[npar]
+  calculate_jamps( const short* hel,                 // input: helicity of each leg, hel[npar], of all events (if allihel == nullptr)
+                   const int* allihel,               // input: helicity index [0,ncomb) of each event[nevt] (nullptr: hel for all events)
                    const fptype_momenta* allmomenta, // input: momenta[nevt*npar*4]
                    const fptype* allcouplings,       // input: couplings[nevt*ndcoup*2]
                    const unsigned int* iflavorVec,   // input: indices of the flavor combinations
diff --git a/madgraph/iolibs/template_files/madmatrix/backend/simd/umami.cc b/madgraph/iolibs/template_files/madmatrix/backend/simd/umami.cc
index 6ca9d8a37a..00730a836a 100644
--- a/madgraph/iolibs/template_files/madmatrix/backend/simd/umami.cc
+++ b/madgraph/iolibs/template_files/madmatrix/backend/simd/umami.cc
@@ -16,6 +16,7 @@
 #include <algorithm>
 #include <cfloat>
 #include <cmath>
+#include <cstdio>
 #include <string>
 #include <vector>
 #include <array>
@@ -246,8 +247,8 @@ extern "C"
   UmamiStatus umami_supported_inputs( bool const** supported, int* count )
   {
     // MOMENTA, ALPHA_S, FLAVOR_INDEX, RANDOM_COLOR, RANDOM_HELICITY, RANDOM_DIAGRAM,
-    // HELICITY_INDEX=false, DIAGRAM_INDEX=true, CHANNEL_INDEX=false
-    static const bool data[UMAMI_INPUT_KEY_COUNT] = { true, true, true, true, true, true, false, true };
+    // HELICITY_INDEX, DIAGRAM_INDEX=true, CHANNEL_INDEX=false
+    static const bool data[UMAMI_INPUT_KEY_COUNT] = { true, true, true, true, true, true, true, true };
     *supported = data;
     *count = UMAMI_INPUT_KEY_COUNT;
     return UMAMI_SUCCESS;
@@ -309,7 +310,36 @@ extern "C"
       static_cast<InterfaceInstance*>( handle )->summed_helicity = parameter_real != 0;
       return UMAMI_SUCCESS;
     }
+    // Parameters are process-wide (Parameters::getInstance()), not per-handle,
+    // matching how umami_initialize itself never stores the CPPProcess it
+    // constructs: initProc() only ever populates the Parameters singleton.
+#ifdef MGONGPU_HARDCODE_PARAM
+    // Hardcoded parameters are compile-time constants: there is nothing to set.
     return UMAMI_ERROR_NOT_IMPLEMENTED;
+#else
+    if( !Parameters::getInstance()->setParameterByName( name, parameter_real, parameter_imag ) )
+      return UMAMI_ERROR_UNKNOWN_PARAMETER;
+    // sigmaKin() and friends read cached copies of the independent parameters/couplings
+    // (cIPD/cIPC/cIPF/bsmIndepParam), not the Parameters singleton itself, for
+    // performance; refresh those caches now so the override actually takes effect.
+    CPPProcess::refreshIndependentParams();
+    // Keep g_externalMasses in sync with the (possibly updated) singleton.
+    // Emit a one-time warning if a mass parameter was changed, since any phase-space
+    // generator that cached masses from a prior umami_get_meta(UMAMI_META_MASSES,...) call
+    // will need to re-query them.
+    {
+      std::vector<fptype> updatedMasses;
+      CPPProcess::refreshExternalMasses( updatedMasses );
+      bool changed = ( updatedMasses.size() != g_externalMasses.size() );
+      for( size_t i = 0; !changed && i < updatedMasses.size(); ++i )
+        changed = ( static_cast<double>( updatedMasses[i] ) != g_externalMasses[i] );
+      if( changed )
+      {
+        g_externalMasses.assign( updatedMasses.begin(), updatedMasses.end() );
+      }
+    }
+    return UMAMI_SUCCESS;
+#endif
   }
 
   UmamiStatus umami_get_parameter(
@@ -318,7 +348,13 @@ extern "C"
     [[maybe_unused]] double* parameter_real,
     [[maybe_unused]] double* parameter_imag )
   {
+#ifdef MGONGPU_HARDCODE_PARAM
     return UMAMI_ERROR_NOT_IMPLEMENTED;
+#else
+    if( !Parameters::getInstance()->getParameterByName( name, parameter_real, parameter_imag ) )
+      return UMAMI_ERROR_UNKNOWN_PARAMETER;
+    return UMAMI_SUCCESS;
+#endif
   }
 
   UmamiStatus umami_matrix_element(
@@ -339,6 +375,7 @@ extern "C"
     const double* random_color_in = nullptr;
     const double* random_helicity_in = nullptr;
     const double* random_diagram_in = nullptr;
+    const int* helicity_in = nullptr;
     [[maybe_unused]] const int* diagram_in = nullptr; // TODO: unused
 
     for( std::size_t i = 0; i < input_count; ++i )
@@ -365,7 +402,8 @@ extern "C"
           random_diagram_in = static_cast<const double*>( input );
           break;
         case UMAMI_IN_HELICITY_INDEX:
-          return UMAMI_ERROR_UNSUPPORTED_INPUT;
+          helicity_in = static_cast<const int*>( input );
+          break;
         case UMAMI_IN_DIAGRAM_INDEX:
           diagram_in = static_cast<const int*>( input );
           break;
@@ -374,6 +412,20 @@ extern "C"
       }
     }
     if( !momenta_in ) return UMAMI_ERROR_MISSING_INPUT;
+    // UMAMI_IN_HELICITY_INDEX: evaluate each event for the one helicity it is given, an index
+    // into the full helicity table [0,ncomb) (the UMAMI_OUT_HELICITY_INDEX convention), instead
+    // of summing over the helicities. The |M|^2 is then the contribution of that helicity to
+    // the helicity-summed (and averaged) |M|^2, so that the sum over all ncomb helicities gives
+    // back the |M|^2 without this input; UMAMI_OUT_HELICITY_INDEX returns the given helicity,
+    // the color and diagram are chosen for it, and UMAMI_IN_RANDOM_HELICITY is not used.
+    if( helicity_in )
+    {
+      for( std::size_t i_event = 0; i_event < count; ++i_event )
+      {
+        const int ihel = helicity_in[i_event + offset];
+        if( ihel < 0 || ihel >= ProcessData::ncomb ) return UMAMI_ERROR;
+      }
+    }
 
     double* m2_out = nullptr;
     double* amp2_out = nullptr;
@@ -460,6 +512,7 @@ extern "C"
     HostBufferBase<fptype_amp, false> denominators( rounded_count );
     HostBufferBase<int, false> helicity_index( rounded_count );
     HostBufferBase<int, false> color_index( rounded_count );
+    HostBufferBase<int, false> helicity_given( rounded_count ); // as the other inputs, the padding lanes copy their source event
     // The input event of each lane. The padding lanes, which complete the last page
     // of a flavor, are copies of the first event of their page, as on GPUs. Left at
     // zero (momenta and g_s), they evaluated to nan, and the good-helicity filter of
@@ -486,6 +539,7 @@ extern "C"
       diagram_random[lane] = random_diagram_in ? random_diagram_in[i_event + offset] : 0.5;
       g_s[lane] = alpha_s_in ? sqrt( 4 * M_PI * alpha_s_in[i_event + offset] ) : 1.2177157847767195;
       flavor_indices[lane] = flavor_indices_in ? flavor_indices_in[i_event + offset] : 0;
+      if( helicity_in ) helicity_given[lane] = helicity_in[i_event + offset];
     }
     computeDependentCouplings( g_s.data(), couplings.data(), rounded_count );
 
@@ -523,7 +577,8 @@ extern "C"
       // an interference |M|^2, see select_helicity_signed, unless the instance was asked
       // to keep the helicity sum (interference_helicity_summed); without the random number
       // the |M|^2 stays the helicity sum (systematics re-evaluation, standalone checks)
-      random_helicity_in != nullptr && !instance->summed_helicity );
+      random_helicity_in != nullptr && !instance->summed_helicity,
+      helicity_in ? helicity_given.data() : nullptr );
 
     if ( sort_flavors )
     {
diff --git a/madgraph/iolibs/template_files/madmatrix/check_sa.cc b/madgraph/iolibs/template_files/madmatrix/check_sa.cc
index 501f94ddbd..c3e2991b78 100644
--- a/madgraph/iolibs/template_files/madmatrix/check_sa.cc
+++ b/madgraph/iolibs/template_files/madmatrix/check_sa.cc
@@ -176,7 +176,7 @@ namespace
     return true;
   }
 
-  // Raw binary per-event values (--alphas <file>, --flavors, --rnd-hel, --rnd-col): one
+  // Raw binary per-event values (--alphas <file>, --flavors, --rnd-hel, --rnd-col, --hel-index): one
   // value of type T per event, with no header, as array.tofile writes them in Python.
   template<typename T>
   bool read_binary_values( const std::string& path, const char* what, std::vector<T>& values )
@@ -206,6 +206,7 @@ namespace
     std::string flavorsFile;                                // --flavors: a file of 32-bit unsigned ints
     std::string rndHelFile;                                 // --rnd-hel: a file of doubles
     std::string rndColFile;                                 // --rnd-col: a file of doubles
+    std::string helIndexFile;                               // --hel-index: a file of 32-bit ints
     std::string dumpHelFile;                                // --dump-hel: write the helicity indices
     std::string dumpColFile;                                // --dump-col: write the color indices
     std::vector<std::pair<std::string, double>> parameters; // --param <name>=<value>
@@ -254,6 +255,8 @@ namespace
       << "  --rnd-hel <file>  Random numbers of the helicity choice (umami otherwise uses 0.5\n"
       << "                    and returns the helicity sum for an interference |M|^2).\n"
       << "  --rnd-col <file>  Random numbers of the color choice (otherwise 0.5).\n"
+      << "  --hel-index <file>  Helicity index [0,ncomb) of every event: the |M|^2 of that one\n"
+      << "                    helicity instead of the helicity sum (UMAMI_IN_HELICITY_INDEX).\n"
       << "  --dump-hel <file>  Write the chosen helicity index of every event (-1: none).\n"
       << "  --dump-col <file>  Write the chosen color index of every event (-1: none).\n"
       << "  --param <name>=<value>  umami_set_parameter after umami_initialize (repeatable).\n"
@@ -648,6 +651,7 @@ namespace
     double alphasValue = param_card_alpha_s();
     std::vector<double> alphasIn, rndHelIn, rndColIn;
     std::vector<unsigned int> flavorsIn;
+    std::vector<int> helIndexIn;
     if( is_float( io.alphas.c_str() ) )
       alphasValue = atof( io.alphas.c_str() );
     else if( !io.alphas.empty() && !read_binary_values( io.alphas, "alpha_s", alphasIn ) )
@@ -655,11 +659,13 @@ namespace
     if( !io.flavorsFile.empty() && !read_binary_values( io.flavorsFile, "flavor", flavorsIn ) ) return 2;
     if( !io.rndHelFile.empty() && !read_binary_values( io.rndHelFile, "helicity random number", rndHelIn ) ) return 2;
     if( !io.rndColFile.empty() && !read_binary_values( io.rndColFile, "color random number", rndColIn ) ) return 2;
+    if( !io.helIndexFile.empty() && !read_binary_values( io.helIndexFile, "helicity index", helIndexIn ) ) return 2;
     const std::size_t nevtRun = eventsFromFile ? lheEvents.size() : (std::size_t)niter * nevt;
     for( auto [size, what] : { std::make_pair( alphasIn.size(), "alpha_s" ),
                                std::make_pair( flavorsIn.size(), "flavor" ),
                                std::make_pair( rndHelIn.size(), "helicity random number" ),
-                               std::make_pair( rndColIn.size(), "color random number" ) } )
+                               std::make_pair( rndColIn.size(), "color random number" ),
+                               std::make_pair( helIndexIn.size(), "helicity index" ) } )
     {
       if( size != 0 && size < nevtRun )
       {
@@ -684,6 +690,7 @@ namespace
     std::vector<double> alphasVec( nevt, alphasValue );
     std::vector<double> rndHelVec( nevt, 0.5 ), rndColVec( nevt, 0.5 );
     std::vector<int> helVec( nevt ), colVec( nevt );
+    std::vector<int> helInVec( nevt, 0 );
 #ifdef MGONGPUCPP_GPUIMPL
     timermap.start( "00 GpuInit" );
     GpuRuntime gpuRuntime( false );
@@ -700,6 +707,7 @@ namespace
     DeviceBufferBase<double> devAlphaS( nevt );
     DeviceBufferBase<double> devRndHel( nevt ), devRndCol( nevt );
     DeviceBufferBase<int> devHel( nevt ), devCol( nevt );
+    DeviceBufferBase<int> devHelIn( nevt );
     std::vector<double> hstUmamiMEs( nevt );
     // the inputs that are the same for every event are copied to the device once
     gpuMemcpy( devFlv.data(), flvVec.data(), nevt * sizeof( unsigned int ), gpuMemcpyHostToDevice );
@@ -712,6 +720,7 @@ namespace
     double* rndColPtr = devRndCol.data();
     int* helPtr = devHel.data();
     int* colPtr = devCol.data();
+    int* helInPtr = devHelIn.data();
     gpuStream_t stream = nullptr; // nullptr: the default stream
     if( io.nonblockingStream ) gpuStreamCreateNonBlocking( &stream );
 #else
@@ -728,6 +737,7 @@ namespace
     double* rndColPtr = rndColVec.data();
     int* helPtr = helVec.data();
     int* colPtr = colVec.data();
+    int* helInPtr = helInVec.data();
 #endif
 
     // The keys and buffers of the umami call, the same for every batch
@@ -745,6 +755,11 @@ namespace
       in_keys.push_back( UMAMI_IN_RANDOM_COLOR );
       inputs.push_back( rndColPtr );
     }
+    if( !helIndexIn.empty() )
+    {
+      in_keys.push_back( UMAMI_IN_HELICITY_INDEX );
+      inputs.push_back( helInPtr );
+    }
     if( dumpHel.is_open() )
     {
       out_keys.push_back( UMAMI_OUT_HELICITY_INDEX );
@@ -922,11 +937,17 @@ namespace
         for( unsigned int ievt = 0; ievt < nevt; ++ievt )
           flvVec[ievt] = flavorsIn[base + std::min( ievt, nreal - 1 )];
       }
+      if( !helIndexIn.empty() )
+      {
+        for( unsigned int ievt = 0; ievt < nevt; ++ievt )
+          helInVec[ievt] = helIndexIn[base + std::min( ievt, nreal - 1 )];
+      }
 #ifdef MGONGPUCPP_GPUIMPL
       if( !alphasIn.empty() ) gpuMemcpy( alphasPtr, alphasVec.data(), nevt * sizeof( double ), gpuMemcpyHostToDevice );
       if( !rndHelIn.empty() ) gpuMemcpy( rndHelPtr, rndHelVec.data(), nevt * sizeof( double ), gpuMemcpyHostToDevice );
       if( !rndColIn.empty() ) gpuMemcpy( rndColPtr, rndColVec.data(), nevt * sizeof( double ), gpuMemcpyHostToDevice );
       if( !flavorsIn.empty() ) gpuMemcpy( flvPtr, flvVec.data(), nevt * sizeof( unsigned int ), gpuMemcpyHostToDevice );
+      if( !helIndexIn.empty() ) gpuMemcpy( helInPtr, helInVec.data(), nevt * sizeof( int ), gpuMemcpyHostToDevice );
       // A non-blocking stream does not wait for the inputs copied and transposed on the
       // default stream: finish them first, as any caller on such a stream must
       if( stream != nullptr ) checkGpu( gpuDeviceSynchronize() );
@@ -1146,10 +1167,11 @@ int main( int argc, char** argv )
     else if( arg == "--energy" && argn + 1 < argc && is_float( argv[argn + 1] ) )
       perfEnergy = atof( argv[++argn] );
     else if( ( arg == "--alphas" || arg == "--flavors" || arg == "--rnd-hel" || arg == "--rnd-col" ||
-               arg == "--dump-hel" || arg == "--dump-col" ) && argn + 1 < argc )
+               arg == "--hel-index" || arg == "--dump-hel" || arg == "--dump-col" ) && argn + 1 < argc )
     {
       std::string& value = ( arg == "--alphas" ? io.alphas : arg == "--flavors" ? io.flavorsFile :
                              arg == "--rnd-hel" ? io.rndHelFile : arg == "--rnd-col" ? io.rndColFile :
+                             arg == "--hel-index" ? io.helIndexFile :
                              arg == "--dump-hel" ? io.dumpHelFile : io.dumpColFile );
       value = argv[++argn];
       mode = MODE_PERF;
diff --git a/madgraph/iolibs/template_files/madmatrix/cpp_model_parameters_cc.inc b/madgraph/iolibs/template_files/madmatrix/cpp_model_parameters_cc.inc
index 5d465f831e..094964fea4 100644
--- a/madgraph/iolibs/template_files/madmatrix/cpp_model_parameters_cc.inc
+++ b/madgraph/iolibs/template_files/madmatrix/cpp_model_parameters_cc.inc
@@ -13,8 +13,11 @@
 
 #include "Parameters.h"
 
+#include <algorithm>
+#include <cctype>
 #include <iomanip>
 #include <iostream>
+#include <sstream>
 
 using namespace madmatrix;
 
@@ -38,6 +41,8 @@ Parameters::setIndependentParameters( SLHAReader& slha )
   zero = 0;                         // define "zero"
   ZERO = 0;                         // define "zero"
   std::vector<int> indices( 2, 0 ); // prepare a vector for indices
+  m_slha = slha;
+  registerParameters();
   %(set_independent_parameters)s
 }
 
@@ -48,6 +53,94 @@ Parameters::setIndependentCouplings()
   %(set_flv_couplings)s
 }
 
+void
+Parameters::registerParameters()
+{
+  %(register_parameters)s
+  %(register_slha_keys)s
+}
+
+bool
+Parameters::resolveSlhaKey( const std::string& name, std::string& resolved ) const
+{
+  std::istringstream iss( name );
+  std::string block;
+  if( !( iss >> block ) ) return false;
+  std::transform( block.begin(), block.end(), block.begin(),
+                  []( unsigned char c ) { return std::tolower( c ); } );
+  std::string key = block;
+  bool any_index = false;
+  int index;
+  while( iss >> index )
+  {
+    key += " " + std::to_string( index );
+    any_index = true;
+  }
+  if( !any_index ) return false;
+  std::map<std::string, std::string>::const_iterator it = m_slha_key_to_name.find( key );
+  if( it == m_slha_key_to_name.end() ) return false;
+  resolved = it->second;
+  return true;
+}
+
+bool
+Parameters::setParameterByName( std::string name, double real, double imag )
+{
+  (void)imag; // settable (SLHA-card-level) parameters are always real
+  std::string resolved;
+  if( resolveSlhaKey( name, resolved ) ) name = resolved;
+  if( m_settable_params.find( name ) == m_settable_params.end() )
+    return false;
+  m_param_overrides[name] = real;
+  setIndependentParameters( m_slha );
+  setIndependentCouplings();
+  return true;
+}
+
+bool
+Parameters::getParameterByName( std::string name, double* real, double* imag )
+{
+  std::string resolved;
+  if( resolveSlhaKey( name, resolved ) ) name = resolved;
+  std::map<std::string, double*>::iterator real_it = m_real_params.find( name );
+  if( real_it != m_real_params.end() )
+  {
+    *real = *( real_it->second );
+    if( imag ) *imag = 0.;
+    return true;
+  }
+  // cxtype (not hardcoded cxsmpl<double>): must match m_complex_params'
+  // own value_type (see its declaration's own comment) -- real()/imag()
+  // below still convert safely into *real/*imag (always plain double*
+  // here) whatever cxtype's own underlying precision is.
+  std::map<std::string, cxtype*>::iterator complex_it = m_complex_params.find( name );
+  if( complex_it != m_complex_params.end() )
+  {
+    *real = complex_it->second->real();
+    if( imag ) *imag = complex_it->second->imag();
+    return true;
+  }
+  // Independent model parameters (as opposed to derived couplings just
+  // above) -- see m_complex_model_params' own declaration comment for why
+  // this is a second, separate map/type rather than sharing
+  // m_complex_params.
+  std::map<std::string, cxsmpl<double>*>::iterator complex_model_it =
+      m_complex_model_params.find( name );
+  if( complex_model_it != m_complex_model_params.end() )
+  {
+    *real = complex_model_it->second->real();
+    if( imag ) *imag = complex_model_it->second->imag();
+    return true;
+  }
+  return false;
+}
+
+void
+Parameters::resetParameterOverrides()
+{
+  m_param_overrides.clear();
+}
+
 /*
 void
 Parameters::setDependentParameters() // now computed event-by-event (running alphas #373)
diff --git a/madgraph/iolibs/template_files/madmatrix/cpp_model_parameters_h.inc b/madgraph/iolibs/template_files/madmatrix/cpp_model_parameters_h.inc
index ded7bf84d5..8bcfa3e425 100644
--- a/madgraph/iolibs/template_files/madmatrix/cpp_model_parameters_h.inc
+++ b/madgraph/iolibs/template_files/madmatrix/cpp_model_parameters_h.inc
@@ -34,6 +34,10 @@
 
 #include "read_slha.h"
 
+#include <map>
+#include <set>
+#include <string>
+
 //One namespace. Split ber backend.
 namespace madmatrix
 {
@@ -105,8 +109,60 @@ namespace madmatrix
     static constexpr int nBsmIndepParam = %(nbsmip)i;
     %(hasbsmip)sdouble mdl_bsmIndepParam[nBsmIndepParam];
 
+    // Set the value of an independent, SLHA-card-level parameter by name at
+    // runtime (overriding the value read from the param card), and recompute
+    // every parameter/coupling that depends on it. Returns false if "name" is
+    // not the name of a settable independent parameter (e.g. it names an
+    // internal parameter derived from others, or a coupling, neither of which
+    // can be set directly since they would just be recomputed from the
+    // underlying parameters on the next call).
+    // "name" may be either the parameter's model name (e.g. "mdl_MT") or its
+    // SLHA card location, given as "<block> <index> [<index2>]" (e.g. "mass
+    // 6"); the block name is matched case-insensitively.
+    bool setParameterByName( std::string name, double real, double imag );
+    // Retrieve the current value of any independent parameter or coupling by
+    // name (real or complex; "name" accepts the same two forms as
+    // setParameterByName). NB: dependent (aS-dependent) parameters and
+    // couplings are not stored here at all in this backend (they are
+    // recomputed fresh every event batch by computeDependentCouplings()), so
+    // they cannot be retrieved by name; only independent ones can. Returns
+    // false if "name" is unknown.
+    bool getParameterByName( std::string name, double* real, double* imag );
+    // Drop all overrides previously set via setParameterByName(), so that the
+    // next setIndependentParameters(slha) call starts from a clean slate (e.g.
+    // when (re-)initializing from a param card). Does not itself recompute
+    // anything; the caller is expected to call setIndependentParameters()/
+    // setIndependentCouplings() afterward (as initProc() does).
+    void resetParameterOverrides();
+
   private:
 
+    // If "name" is of the form "<block> <index> [<index2>]", resolve it (via
+    // m_slha_key_to_name) to the corresponding parameter's model name and
+    // return true; otherwise leave "resolved" untouched and return false.
+    bool resolveSlhaKey( const std::string& name, std::string& resolved ) const;
+
+    // Name-keyed lookup tables populated by registerParameters(), backing
+    // setParameterByName()/getParameterByName() above.
+    void registerParameters();
+    std::map<std::string, double*> m_real_params;
+    // Two separate maps, not one: independent *couplings* (eg GC_72 below)
+    // are declared cxtype, which resolves to cxsmpl<float> at FPTYPE=f, so
+    // this map's value_type has to be cxtype too (a mismatched hardcoded
+    // cxsmpl<double> here fails to compile the moment any coupling is
+    // actually registered below, eg complex effective couplings like
+    // ggH's) -- while independent model *parameters* (eg mdl_I1x33) always
+    // stay plain cxsmpl<double>, matching m_real_params' own always-double
+    // real parameters, and go into m_complex_model_params below instead;
+    // see write_register_parameters()'s own doc comment (export_cpp.py)
+    // for why these can't share one map/type.
+    std::map<std::string, cxtype*> m_complex_params;
+    std::map<std::string, cxsmpl<double>*> m_complex_model_params;
+    std::set<std::string> m_settable_params;
+    std::map<std::string, double> m_param_overrides;
+    std::map<std::string, std::string> m_slha_key_to_name;
+    SLHAReader m_slha;
+
     static Parameters* instance;
   };
 
diff --git a/madgraph/iolibs/template_files/madmatrix/madmatrix.mk b/madgraph/iolibs/template_files/madmatrix/madmatrix.mk
index c60d486507..ca0b79971d 100644
--- a/madgraph/iolibs/template_files/madmatrix/madmatrix.mk
+++ b/madgraph/iolibs/template_files/madmatrix/madmatrix.mk
@@ -409,6 +409,9 @@ else
 endif
 export BACKENDDIR
 
+# Export FPTYPE (used e.g. to name the common library, see MADMATRIX_COMMONLIB below)
+export FPTYPE
+
 #-------------------------------------------------------------------------------
 
 #=== Configure ccache for C++ and CUDA/HIP builds
@@ -469,11 +472,16 @@ ifeq ($(PCH),1)
 endif
 
 # Dependency on src directory
-# The common library name carries the full BACKEND suffix so each vectorisation/GPU variant is distinct.
+# The common library name carries the model name and the full BACKEND suffix -- see the
+# matching comment in madmatrix_src.mk for why the model name is required (dlopen
+# collisions across process directories regenerated in different models) -- and, purely
+# for naming consistency with MADMATRIX_LIB, the FPTYPE too; unlike MADMATRIX_LIB this
+# isn't consumed by anything outside these madmatrix makefiles (subprocesses.json never
+# references it directly), so no further propagation is needed for this one.
 # LIBTAG (empty by default) goes into the name of every library of this output: the
 # dynamic loader identifies a library by its name, so two outputs loaded into one
 # process (MadSpin's production and decay matrix elements) need different ones.
-MADMATRIX_COMMONLIB = madmatrix_common$(LIBTAG)_$(BACKEND)
+MADMATRIX_COMMONLIB = madmatrix_common$(LIBTAG)_%(model)s_$(BACKEND)_$(FPTYPE)
 LIBFLAGS = -L$(LIBDIR) -l$(MADMATRIX_COMMONLIB)
 INCFLAGS += -I$(SRC) -I$(SRC)/rambo
 
@@ -814,7 +822,16 @@ override RUNTIME =
 processid_short=$(shell basename $(CURDIR))
 ###$(info processid_short=$(processid_short))
 
-MADMATRIX_LIB = madmatrix$(LIBTAG)_$(processid_short)_$(BACKEND)
+# The process library name carries LIBTAG (see MADMATRIX_COMMONLIB above), the model
+# name (same %(model)s substitution, and the same dlopen-collision rationale, as
+# MADMATRIX_COMMONLIB above), the process id, the hardware backend, and the
+# floating-point type -- so that libraries built for different models/precisions/backends
+# out of the same process directory never collide, and so that a loader can tell, from
+# the filename alone, exactly which of those a given library on disk was built with
+# (subprocesses.json itself stays agnostic to BACKEND and FPTYPE -- see me_lib_format in
+# madmatrix/output.py -- since a given process directory can be compiled for several of
+# each without regenerating the JSON).
+MADMATRIX_LIB = madmatrix$(LIBTAG)_%(model)s_$(processid_short)_$(BACKEND)_$(FPTYPE)
 objects_lib=$(BUILDDIR)/CPPProcess.o $(BUILDDIR)/color_sum.o $(BUILDDIR)/MatrixElementKernels.o $(BUILDDIR)/CrossSectionKernels.o $(BUILDDIR)/umami.o $(BUILDDIR)/SigmaKin.o
 # The spin-density matrices for MadSpin (backend/common/Density.h): C++ backends,
 # and only in the builds madmatrix/density.py asks for (DENSITY=1, the scalar
@@ -917,7 +934,7 @@ ifeq ($(MADMATRIX_COMMONLIB_EXTERNAL),1)
 $(LIBDIR)/lib$(MADMATRIX_COMMONLIB).so: ;
 else
 $(LIBDIR)/lib$(MADMATRIX_COMMONLIB).so: $(SRC)/*.h $(SRC)/*.cc $(BUILDDIR)/.build.$(TAG)
-	$(MAKE) -C $(SRC) BACKEND=$(BACKEND) LIBDIR=$(LIBDIR)
+	$(MAKE) -C $(SRC) BACKEND=$(BACKEND) FPTYPE=$(FPTYPE) LIBDIR=$(LIBDIR)
 endif
 
 #-------------------------------------------------------------------------------
@@ -1020,20 +1037,26 @@ $(addprefix commonlib.,$(BLDBACKENDS)): commonlib.%%:
 clean:
 ifeq ($(USEBUILDDIR),1)
 	rm -rf $(BUILDDIR)
+ifneq ($(MADMATRIX_COMMONLIB_EXTERNAL),1)
+	$(MAKE) -C $(SRC) clean BACKEND=$(BACKEND) FPTYPE=$(FPTYPE) LIBDIR=$(LIBDIR)
+endif
 else
 	rm -f $(BUILDDIR)/.build.* $(BUILDDIR)/*.o
 	rm -f $(LIBDIR)/lib$(MADMATRIX_LIB).so
 	rm -f $(BACKEND_LOG)
 ifneq ($(MADMATRIX_COMMONLIB_EXTERNAL),1)
-	$(MAKE) -C $(SRC) clean BACKEND=$(BACKEND) LIBDIR=$(LIBDIR)
+	$(MAKE) -C $(SRC) clean BACKEND=$(BACKEND) FPTYPE=$(FPTYPE) LIBDIR=$(LIBDIR)
 endif
 endif
 
-# cleanall: remove objects and libraries for ALL backends.
+# cleanall: remove objects and libraries for ALL backends and FPTYPEs.
+# (NB: MADMATRIX_LIB itself only names one BACKEND/FPTYPE combination -- see
+# its own definition above -- so this globs for all of them explicitly,
+# rather than just removing $(MADMATRIX_LIB) as 'clean' above does.)
 cleanall:
 	rm -rf build.*
 	rm -f ./.build.* ./*.o
-	rm -f $(LIBDIR)/libmadmatrix_$(processid_short)_*.so
+	rm -f $(LIBDIR)/libmadmatrix_%(model)s_$(processid_short)_*.so
 ifneq ($(MADMATRIX_COMMONLIB_EXTERNAL),1)
 	$(MAKE) -C $(SRC) cleanall LIBDIR=$(LIBDIR)
 endif
@@ -1042,7 +1065,7 @@ endif
 # for the SubProcesses dispatcher, which owns the common library while the P*
 # directories are cleaned in parallel with MADMATRIX_COMMONLIB_EXTERNAL=1.
 cleancommon:
-	$(MAKE) -C $(SRC) clean BACKEND=$(BACKEND) LIBDIR=$(LIBDIR)
+	$(MAKE) -C $(SRC) clean BACKEND=$(BACKEND) FPTYPE=$(FPTYPE) LIBDIR=$(LIBDIR)
 
 cleanallcommon:
 	$(MAKE) -C $(SRC) cleanall LIBDIR=$(LIBDIR)
diff --git a/madgraph/iolibs/template_files/madmatrix/madmatrix_src.mk b/madgraph/iolibs/template_files/madmatrix/madmatrix_src.mk
index 1a68e606a5..f918e2a6c7 100644
--- a/madgraph/iolibs/template_files/madmatrix/madmatrix_src.mk
+++ b/madgraph/iolibs/template_files/madmatrix/madmatrix_src.mk
@@ -66,9 +66,17 @@ endif
 #=== Makefile TARGETS and build rules below
 #===============================================================================
 
-# The common library name carries the full BACKEND suffix so each vectorisation/GPU variant is distinct
-# (and LIBTAG, see madmatrix.mk).
-MADMATRIX_COMMONLIB = madmatrix_common$(LIBTAG)_$(BACKEND)
+# The common library name carries the model name and the full BACKEND suffix, so each
+# vectorisation/GPU variant is distinct AND libraries from different regenerated model
+# directories (eg the same process regenerated in several models and loaded into one
+# program) never collide: dlopen resolves a bare NEEDED entry by matching any
+# already-loaded library of that name anywhere in the process, regardless of RUNPATH,
+# so two "libmadmatrix_common_<backend>.so" from different models would otherwise
+# silently alias -- the second model's process code ends up bound to the first
+# model's Parameters implementation. FPTYPE is appended too, purely for naming
+# consistency with MADMATRIX_LIB (SubProcesses/madmatrix.mk); FPTYPE is exported from
+# there, same as BACKEND. LIBTAG as well, see madmatrix.mk.
+MADMATRIX_COMMONLIB = madmatrix_common$(LIBTAG)_%(model)s_$(BACKEND)_$(FPTYPE)
 
 # Name recorded in the library itself, and hence in whatever links against it.
 # Without it the linker bakes in the absolute build path (LIBDIR), so the process
diff --git a/madgraph/iolibs/template_files/madmatrix/process_class.inc b/madgraph/iolibs/template_files/madmatrix/process_class.inc
index 51aecbc4dd..c43d0876e0 100644
--- a/madgraph/iolibs/template_files/madmatrix/process_class.inc
+++ b/madgraph/iolibs/template_files/madmatrix/process_class.inc
@@ -25,6 +25,21 @@
     // Initialize process (read model parameters from file)
     virtual void initProc( const std::string& param_card_name );
 
+#ifndef MGONGPU_HARDCODE_PARAM
+    // Re-synchronize the cached independent parameters/couplings (cIPD/cIPC/cIPF) and the
+    // BSM-independent-parameter buffer (bsmIndepParam) with the current values in the
+    // Parameters singleton. sigmaKin() and friends read these cached copies, not the
+    // singleton itself, for performance, so updating the singleton alone (e.g. via
+    // Parameters::setParameterByName()) has no effect until this is called again; it is
+    // called once automatically by initProc().
+    static void refreshIndependentParams();
+    // Re-populate a caller-supplied mass vector from the current Parameters singleton,
+    // mirroring what initProc() stores in m_masses. Call after umami_set_parameter
+    // changes a mass parameter so callers of umami_get_meta(UMAMI_META_MASSES,...) see
+    // the updated values.
+    static void refreshExternalMasses( std::vector<fptype>& masses );
+#endif
+
     // PDG id for a given flavor combination index and particle index (0-based).
     // Useful for tools that need to enumerate flavor combinations without going through UMAMI.
     static int flavorPDG( int iflavor, int ipar );
diff --git a/madgraph/iolibs/template_files/madmatrix/process_function_definitions.inc b/madgraph/iolibs/template_files/madmatrix/process_function_definitions.inc
index 2f2a273d02..3e3875927c 100644
--- a/madgraph/iolibs/template_files/madmatrix/process_function_definitions.inc
+++ b/madgraph/iolibs/template_files/madmatrix/process_function_definitions.inc
@@ -65,6 +65,12 @@ namespace madmatrix
   {
     // Instantiate the model class and set parameters that stay fixed during run
     m_pars = Parameters::getInstance();
+    // Parameters is a process-wide singleton (not owned per-instance), so a fresh
+    // initProc() must explicitly drop any overrides left by a previous
+    // setParameterByName() call -- otherwise re-initializing from a (possibly
+    // different) param card would silently carry forward stale overrides instead
+    // of giving a clean slate.
+    m_pars->resetParameterOverrides();
     SLHAReader slha( param_card_name, m_verbose );
     m_pars->setIndependentParameters( slha );
     m_pars->setIndependentCouplings();
@@ -79,6 +85,14 @@ namespace madmatrix
     }
     %(initProc_lines)s
     createNormalizedColorMatrix(); // no-op on cpu/simd, pushes to device memory on gpu
+    refreshIndependentParams();
+  }
+
+  // See declaration in CPPProcess.h for why this exists and when it must be called.
+  void
+  CPPProcess::refreshIndependentParams()
+  {
+    Parameters* m_pars = Parameters::getInstance();
     // Read physics parameters like masses and couplings from user configuration files,
     // then hand them to the backend-owned storage via the setters (see SigmaKin.cc).
     %(cipdassign)s
@@ -86,6 +100,15 @@ namespace madmatrix
     %(cipfassign)s
     %(bsmassign)s
   }
+
+  // See declaration in CPPProcess.h for why this exists and when it must be called.
+  void
+  CPPProcess::refreshExternalMasses( std::vector<fptype>& masses )
+  {
+    Parameters* m_pars = Parameters::getInstance();
+    masses.clear();
+%(refreshExternalMasses_lines)s
+  }
 #else
   // Initialize process (with hardcoded parameters)
   void
diff --git a/madgraph/iolibs/template_files/mg7/api.cpp b/madgraph/iolibs/template_files/mg7/api.cpp
index 6961afc7bd..d8198ce285 100644
--- a/madgraph/iolibs/template_files/mg7/api.cpp
+++ b/madgraph/iolibs/template_files/mg7/api.cpp
@@ -1,7 +1,17 @@
 #include "CPPProcess.h"
 #include "api.h"
+#include <cstdio>
 #include <vector>
 
+namespace {
+    // Masses reported by umami_get_meta(UMAMI_META_MASSES, ...), which (unlike every other
+    // UMAMI entry point in this file) takes no handle: model parameters are stored per-
+    // instance in this backend (CPPProcess::pars), so this can only ever reflect the most
+    // recently initialized-or-modified instance, not a specific one. See the caveat on
+    // umami_get_meta() in api.h.
+    std::vector<double> g_masses;
+}
+
 extern "C" {
 
 UmamiStatus umami_get_meta(UmamiMetaKey meta_key, void* result) {
@@ -20,6 +30,12 @@ UmamiStatus umami_get_meta(UmamiMetaKey meta_key, void* result) {
         break;
     case UMAMI_META_COLOR_COUNT:
         return UMAMI_ERROR_UNSUPPORTED_META;
+    case UMAMI_META_MASSES: {
+        if (g_masses.size() != (size_t)CPPProcess::nexternal) return UMAMI_ERROR_UNINITIALIZED_META;
+        for (int i = 0; i < CPPProcess::nexternal; ++i)
+            static_cast<double*>(result)[i] = g_masses[i];
+        break;
+    }
     default:
         return UMAMI_ERROR_UNSUPPORTED_META;
     }
@@ -54,6 +70,7 @@ UmamiStatus umami_initialize(UmamiHandle* handle, char const* param_card_path) {
     CPPProcess* process = new CPPProcess(param_card_path);
     std::vector<double*>& momenta = process->getMomenta();
     for (int i = 0; i < CPPProcess::nexternal; ++i) momenta.push_back(new double[4]());
+    g_masses.assign(process->getMasses().begin(), process->getMasses().end());
     *handle = process;
     return UMAMI_SUCCESS;
 }
@@ -65,7 +82,18 @@ UmamiStatus umami_set_parameter(
     double parameter_real,
     double parameter_imag
 ) {
-    return UMAMI_ERROR_NOT_IMPLEMENTED;
+    CPPProcess* process = static_cast<CPPProcess*>(handle);
+    if (!process->getParameters().setParameterByName(name, parameter_real, parameter_imag))
+        return UMAMI_ERROR_UNKNOWN_PARAMETER;
+    process->refreshMasses();
+    const std::vector<double>& masses = process->getMasses();
+    bool changed = (masses.size() != g_masses.size());
+    for (size_t i = 0; !changed && i < masses.size(); ++i)
+        changed = (masses[i] != g_masses[i]);
+    if (changed) {
+        g_masses.assign(masses.begin(), masses.end());
+    }
+    return UMAMI_SUCCESS;
 }
 
 
@@ -75,7 +103,10 @@ UmamiStatus umami_get_parameter(
     double* parameter_real,
     double* parameter_imag
 ) {
-    return UMAMI_ERROR_NOT_IMPLEMENTED;
+    CPPProcess* process = static_cast<CPPProcess*>(handle);
+    if (!process->getParameters().getParameterByName(name, parameter_real, parameter_imag))
+        return UMAMI_ERROR_UNKNOWN_PARAMETER;
+    return UMAMI_SUCCESS;
 }
 
 UmamiStatus umami_matrix_element(
diff --git a/madgraph/iolibs/template_files/mg7/api.h b/madgraph/iolibs/template_files/mg7/api.h
index 426a342000..16fe3d514b 100644
--- a/madgraph/iolibs/template_files/mg7/api.h
+++ b/madgraph/iolibs/template_files/mg7/api.h
@@ -40,6 +40,8 @@ typedef enum {
     UMAMI_ERROR_UNSUPPORTED_OUTPUT,
     UMAMI_ERROR_UNSUPPORTED_META,
     UMAMI_ERROR_MISSING_INPUT,
+    UMAMI_ERROR_UNKNOWN_PARAMETER,
+    UMAMI_ERROR_UNINITIALIZED_META,
 } UmamiStatus;
 
 typedef enum {
@@ -54,6 +56,7 @@ typedef enum {
     UMAMI_META_DIAGRAM_COUNT,
     UMAMI_META_HELICITY_COUNT,
     UMAMI_META_COLOR_COUNT,
+    UMAMI_META_MASSES,
 } UmamiMetaKey;
 
 typedef enum {
@@ -97,7 +100,12 @@ typedef void* UmamiHandle;
  *     pointer to an instance of the subprocess. Has to be cleaned up by
  *     the caller with `free_subprocess`.
  * @return
- *     UMAMI_SUCCESS on success, error code otherwise
+ *     UMAMI_SUCCESS on success, error code otherwise. UMAMI_META_MASSES returns
+ *     UMAMI_ERROR_UNINITIALIZED_META if no instance has been created yet with
+ *     umami_initialize(). NB: unlike other meta keys, UMAMI_META_MASSES can vary between
+ *     instances (parameters are stored per-instance in this backend, and are settable via
+ *     umami_set_parameter); since this function takes no handle, it always reports the
+ *     masses of the most recently initialized-or-modified instance, not a specific one.
  */
 UmamiStatus umami_get_meta(UmamiMetaKey meta_key, void* result);
 
@@ -143,18 +151,24 @@ UmamiStatus umami_supported_outputs(bool const** supported, int* count);
 UmamiStatus umami_initialize(UmamiHandle* handle, char const* param_card_path);
 
 /**
- * Sets the value of a model parameter
+ * Sets the value of a model parameter. Only independent, SLHA-card-level
+ * parameters (i.e. those normally read from the param card) can be set this
+ * way; parameters or couplings derived from them are recomputed
+ * automatically and cannot be overridden directly.
  *
  * @param handle
  *     handle of a matrix element instance
  * @param name
- *     name of the parameter
+ *     name of the parameter, given either as its model name (e.g. "mdl_MT")
+ *     or as its SLHA card location, "<block> <index> [<index2>]" (e.g.
+ *     "mass 6"); the block name is matched case-insensitively
  * @param parameter_real
  *     real part of the parameter value
  * @param parameter_imag
  *     imaginary part of the parameter value. Ignored for real valued parameters.
  * @return
- *     UMAMI_SUCCESS on success, error code otherwise
+ *     UMAMI_SUCCESS on success, UMAMI_ERROR_UNKNOWN_PARAMETER if "name" is not
+ *     a settable parameter, error code otherwise
  */
 UmamiStatus umami_set_parameter(
     UmamiHandle handle,
@@ -169,14 +183,16 @@ UmamiStatus umami_set_parameter(
  * @param handle
  *     handle of a matrix element instance
  * @param name
- *     name of the parameter
+ *     name of the parameter; accepts the same two forms as
+ *     umami_set_parameter (model name, or SLHA "<block> <index> [<index2>]")
  * @param parameter_real
  *     pointer to double to return real part of the parameter value
  * @param parameter_imag
  *     pointer to double to return imaginary part of the parameter value. Ignored
  *     for real-valued parameters (i.e. you may pass a null pointer)
  * @return
- *     UMAMI_SUCCESS on success, error code otherwise
+ *     UMAMI_SUCCESS on success, UMAMI_ERROR_UNKNOWN_PARAMETER if "name" is not
+ *     a known parameter or coupling, error code otherwise
  */
 UmamiStatus umami_get_parameter(
     UmamiHandle handle,
diff --git a/madgraph/iolibs/template_files/mg7/bootstrap.py b/madgraph/iolibs/template_files/mg7/bootstrap.py
index 90fab5bcd4..b134a63f8c 100644
--- a/madgraph/iolibs/template_files/mg7/bootstrap.py
+++ b/madgraph/iolibs/template_files/mg7/bootstrap.py
@@ -32,6 +32,10 @@
 MG_ROOT = Path(_mg_pkg.__file__).parents[1]
 MADSPACE_DIR = MG_ROOT / "madspace"
 INSTALL_DIR = MADSPACE_DIR / "install"
+# MadtRex (event reweighting) is installed separately, on top of madspace;
+# nothing here installs it -- see madtrex/install.py.
+MADTREX_DIR = MG_ROOT / "madtrex"
+MADTREX_INSTALL_DIR = MADTREX_DIR / "install"
 
 
 def madspace_is_installed() -> bool:
diff --git a/madgraph/iolibs/template_files/mg7/generic_reweight_card.py b/madgraph/iolibs/template_files/mg7/generic_reweight_card.py
new file mode 100644
index 0000000000..f6afa969a0
--- /dev/null
+++ b/madgraph/iolibs/template_files/mg7/generic_reweight_card.py
@@ -0,0 +1,218 @@
+"""Parsing for MadtRex's generic reweight card -- the card that lets one
+card file mix parameter (matrix-element) and PDF reweighting, which
+otherwise run as two entirely independent passes (see
+param_reweight_card.py/pdf_reweight_card.py's own module docstrings, and
+madtrex_driver.py's run_parameter_reweighting()/run_pdf_reweighting()) --
+and the reweight-card vocabulary both MadtRex card readers share with the
+classic reweighting module (madgraph/interface/reweight_interface.py).
+
+Vocabulary: Cards/reweight_card.dat is read by either backend, so it uses
+the classic module's spelling throughout -- "change <option> <value>" for
+options of the reweighting itself, "set ..." only for physics parameters
+(and, in a PDF section, the PDF set and scales). Where both backends have
+the same option, it has the classic name and the classic default
+("change helicity True"). The MadtRex spellings this replaces ("set
+helicity_reweighting", "set FPTYPE", "set mode pdf", ...) are still
+accepted, with a deprecation warning. The directives only the classic
+module understands (see CLASSIC_ONLY_CHANGE_OPTIONS) are an error here
+rather than silently ignored: they need "change backend classic".
+
+The section grammar is deliberately minimal -- this module does not
+understand either mode's own vocabulary, only:
+
+- "change mode <LO|pdf>": switches which mode every following line belongs
+  to, until the next "change mode" line or end of file. LO (the classic
+  module's default mode) is parameter reweighting, pdf is PDF/scale
+  reweighting. May appear more than once (switching back and forth), so
+  launches of each mode can be interleaved in whatever order is most
+  natural to write. Lines before the card's first "change mode" belong to
+  LO, the default mode -- so a card without any "change mode" line is a
+  plain parameter card. The former spelling "set mode <parameter|pdf>" is
+  still accepted;
+- everything else is opaque here: split_by_mode() just buckets each line's
+  raw text by whichever mode is currently active, and hands the
+  concatenated per-mode text (preserving each mode's own original line
+  order) to that mode's real parser (param_reweight_card.parse_reweight_card()/
+  pdf_reweight_card.parse_pdf_reweight_card()) completely unchanged -- so a
+  "change mode LO" section's lines must themselves already be valid
+  parameter-card syntax (change model/launch/set BLOCK ID VALUE/...), and
+  likewise for "change mode pdf" (launch/set PDF_SET/set MUF/...).
+  Concatenating same-mode chunks in file order preserves each mode's own
+  preamble-vs-after-first-launch rules correctly, as long as (as is the
+  natural way to write one) a mode's preamble-only directives appear in that
+  mode's *first* section, before its own first "launch". An LO section
+  without any "launch" (eg only "change helicity False" before a "change
+  mode pdf") is dropped when another section has launches.
+
+madtrex_driver.py's run_generic_reweighting() runs whichever mode(s)
+split_by_mode() finds (both, if both are present) through their own real
+routine, then coalesces the results back into a single output sample.
+"""
+
+import logging
+import re
+
+logger = logging.getLogger("madgraph7")
+
+MODES = frozenset({"parameter", "pdf"})
+
+# "change mode" values -> the section they select. LO is the classic
+# module's name for parameter reweighting; "parameter" is the former
+# "set mode parameter".
+_MODE_NAMES = {"lo": "parameter", "parameter": "parameter", "pdf": "pdf"}
+
+# The "change backend" values (see madgraph/various/reweight_backend.py).
+BACKENDS = frozenset({"classic", "madtrex", "auto"})
+
+# "change" options only the classic reweighting module understands --
+# reweight_interface.REWEIGHT_CHANGE_OPTIONS/DENSITY_CHANGE_OPTIONS minus the
+# shared ones (backend, model, process, helicity, mode). Kept here since this
+# module ships in gridpacks without MadGraph; a unit test keeps the two in
+# sync.
+CLASSIC_ONLY_CHANGE_OPTIONS = frozenset({
+    "keep_ordering", "use_eventid", "allow_missing_finalstate", "boost",
+    "virtual_path", "tree_path", "output", "rwgt_dir", "systematics",
+    "soft_threshold", "multicore", "identical_particle_in_prod_and_decay",
+    "include_sudakov",
+    # density-matrix reweighting (reweight_interface.DensityInterface)
+    "helicity_direction", "boost_choice", "order_helicities",
+    "symmetrise_initial_state", "axis_referential", "matrix_normalisation",
+    "particle_in_density_matrix",
+})
+# "change mode" values only the classic module understands.
+CLASSIC_ONLY_MODES = frozenset({"nlo", "nlo_tree", "lo+nlo"})
+
+_MODE_RE = re.compile(r"^\s*(?P<verb>change|set)\s+mode\s+(?P<mode>\S+)\s*$", re.I)
+_LAUNCH_RE = re.compile(r"^\s*launch\b", re.I)
+
+
+# A "!" comment: from a "!" at the start of the line or after whitespace (the
+# reweight cards' examples use both "#" and "!" for trailing comments).
+_BANG_COMMENT_RE = re.compile(r"(?:^|\s)!.*$")
+
+
+def strip_comment(line: str) -> str:
+    """line without its comment: what follows a "#", or a "!" at the start
+    of the line or after whitespace."""
+    return _BANG_COMMENT_RE.sub("", line.split("#", 1)[0])
+
+
+def deprecated(old: str, new: str) -> None:
+    """Warns that the reweight-card spelling old is replaced by new."""
+    logger.warning("reweight card: '%s' is deprecated, use '%s' instead", old, new)
+
+
+def classic_only_reason(line: str) -> str | None:
+    """Why line (comment already stripped) needs the classic reweighting
+    module, or None when MadtRex may understand it."""
+    words = line.split()
+    if len(words) < 2 or words[0].lower() != "change":
+        if words and words[0].lower() not in ("set", "change", "launch") and (
+            "/" in words[0] or words[0].endswith(".dat")
+        ):
+            return "a param_card given by path"
+        return None
+    name = words[1].lower()
+    if name in CLASSIC_ONLY_CHANGE_OPTIONS:
+        return f"'change {name}' is a classic reweighting option"
+    if name == "mode" and len(words) > 2 and words[2].lower() in CLASSIC_ONLY_MODES:
+        return "NLO reweighting"
+    if name == "process" and words[-1] == "--add":
+        return "several processes for one model"
+    return None
+
+
+def check_not_classic_only(line: str, raw_line: str) -> None:
+    """Raises if line needs the classic reweighting module."""
+    reason = classic_only_reason(line)
+    if reason is not None:
+        raise ValueError(
+            f"reweight card line {raw_line.strip()!r} needs the classic reweighting "
+            f"module ({reason}), not MadtRex: set 'change backend classic' (or remove "
+            f"the line)"
+        )
+
+
+def check_backend_line(words: list[str], raw_line: str) -> None:
+    """Checks a "change backend <value>" line read by MadtRex."""
+    value = words[2].lower() if len(words) > 2 else ""
+    if value not in BACKENDS:
+        raise ValueError(
+            f"reweight card line {raw_line.strip()!r}: unknown backend {value!r}; "
+            f"expected one of {sorted(BACKENDS)}"
+        )
+    if value == "classic":
+        raise ValueError(
+            f"reweight card line {raw_line.strip()!r} asks for the classic reweighting "
+            f"module, not MadtRex"
+        )
+
+
+CARD_KINDS = ("param", "pdf", "generic")
+# "set" names and "change" options of a PDF reweighting card
+_PDF_SET_NAMES = frozenset({
+    "pdf_set", "beam1_pdf_set", "beam2_pdf_set", "muf", "mur", "muf_factor", "mur_factor",
+})
+_PDF_CHANGE_OPTIONS = frozenset({"alpha_s_power", "scale_variation_output"})
+
+
+def card_kind(text: str) -> str:
+    """Which MadtRex runner reads the card text: "generic" when it has a
+    "change mode" (or former "set mode") line, "pdf" when it names a PDF set,
+    scale or PDF-reweighting option, else "param". Same rule as
+    madgraph/various/reweight_backend.classify_card()."""
+    has_pdf = False
+    for raw_line in text.splitlines():
+        words = strip_comment(raw_line).split()
+        if len(words) < 2:
+            continue
+        keyword, name = words[0].lower(), words[1].lower()
+        if keyword in ("change", "set") and name == "mode":
+            return "generic"
+        if (keyword == "set" and len(words) == 3 and name in _PDF_SET_NAMES | _PDF_CHANGE_OPTIONS) \
+                or (keyword == "change" and name in _PDF_CHANGE_OPTIONS):
+            has_pdf = True
+    return "pdf" if has_pdf else "param"
+
+
+def split_by_mode(text: str) -> dict[str, str]:
+    """Splits text into {mode: sub_card_text} ("parameter" for LO, "pdf"),
+    one entry per mode actually used (in first-appearance order), each
+    sub_card_text being every line that belonged to that mode, concatenated
+    in original file order and ready to hand to that mode's own parser
+    as-is. Lines before the first "change mode" are LO.
+
+    Raises if a "change mode" line names anything other than LO/pdf (or the
+    former "set mode parameter|pdf")."""
+    chunks: dict[str, list[str]] = {"parameter": []}
+    has_launch = {"parameter": False}
+    current_mode = "parameter"
+
+    for raw_line in text.splitlines():
+        line = strip_comment(raw_line)
+
+        mode_match = _MODE_RE.match(line) if line.strip() else None
+        if mode_match:
+            value = mode_match.group("mode").strip()
+            mode = _MODE_NAMES.get(value.lower())
+            if mode is None:
+                if value.lower() in CLASSIC_ONLY_MODES:
+                    check_not_classic_only(line, raw_line)
+                raise ValueError(
+                    f"generic reweight card: unknown reweighting mode {value!r} in "
+                    f"{raw_line!r}; expected 'change mode LO' or 'change mode pdf'"
+                )
+            if mode_match.group("verb").lower() == "set":
+                deprecated(f"set mode {value}", "change mode " + ("LO" if mode == "parameter" else "pdf"))
+            current_mode = mode
+            chunks.setdefault(current_mode, [])
+            has_launch.setdefault(current_mode, False)
+            continue
+
+        if _LAUNCH_RE.match(line):
+            has_launch[current_mode] = True
+        chunks[current_mode].append(raw_line)
+
+    if not has_launch["parameter"] and len(chunks) > 1:
+        del chunks["parameter"]
+    return {mode: "\n".join(lines) for mode, lines in chunks.items()}
diff --git a/madgraph/iolibs/template_files/mg7/gridpack.py b/madgraph/iolibs/template_files/mg7/gridpack.py
index 2f1ab4f6cd..05f37d3283 100644
--- a/madgraph/iolibs/template_files/mg7/gridpack.py
+++ b/madgraph/iolibs/template_files/mg7/gridpack.py
@@ -29,6 +29,13 @@
         raise RuntimeError("madspace installation failed — see output above")
     sys.path.insert(0, str(_LOCAL_INSTALL_DIR))
 
+# MadtRex (event reweighting), if the gridpack was made with include_madtrex:
+# put it on the path so it can be imported next to madspace. Nothing in this
+# script imports it.
+_LOCAL_MADTREX_INSTALL_DIR = _GRIDPACK_DIR / "madtrex" / "install"
+if (_LOCAL_MADTREX_INSTALL_DIR / "madtrex").is_dir():
+    sys.path.insert(1, str(_LOCAL_MADTREX_INSTALL_DIR))
+
 import madspace as ms
 from custom_function_loader import load_custom_function
 import glob
@@ -36,6 +43,15 @@
 import tomllib
 import argparse
 
+_PRECISION_2_FPTYPE = {"all64": "d", "all32": "f", "color32": "m", "denom64": "v"}
+
+
+def fptype_of(precision: str) -> str:
+    """[run] precision -> FPTYPE the matrix elements were compiled with (as
+    fptype_of() in launch.py, which this standalone script can't import)."""
+    return _PRECISION_2_FPTYPE.get(str(precision).lower(), str(precision).lower())
+
+
 def resolve_verbosity(verbosity: str) -> str:
     """Resolve the run_card "auto" verbosity to "pretty"/"log" depending on
     whether stdout is attached to a terminal; other values pass through
@@ -288,6 +304,8 @@ def resolve_dir(cli_value, card_value):
     # run-time matrix-element parameters (bwcutoff) of the run that made the
     # gridpack; gridpacks written before they were recorded used the default
     me_parameters = madspace_data.get("me_parameters", {})
+    # the FPTYPE the matrix elements were compiled with ({fptype} in me_path)
+    fptype = fptype_of(run_args["precision"])
     globals_path = os.path.join("data", "globals.tar")
     for context in contexts:
         context.load_globals(globals_path)
@@ -301,7 +319,7 @@ def resolve_dir(cli_value, card_value):
     for context, backend in zip(contexts, backends):
         for me_path in madspace_data["matrix_elements"]:
             context.load_matrix_element(
-                me_path.format(device=backend), param_card_path, me_parameters
+                me_path.format(device=backend, fptype=fptype), param_card_path, me_parameters
             )
 
     # set up generators
@@ -457,7 +475,9 @@ def load_systematics(run_card, backends=(), param_card_path=None,
             if i not in need:
                 matrix_elements.append(None)
                 continue
-            lib = me_path.format(device=backend)
+            lib = me_path.format(
+                device=backend, fptype=fptype_of(run_card["run"]["precision"])
+            )
             if not os.path.exists(lib):
                 print("WARNING systematics: %s not found, mu_R variations of the "
                       "mixed-order subprocesses are dropped" % lib)
diff --git a/madgraph/iolibs/template_files/mg7/launch.py b/madgraph/iolibs/template_files/mg7/launch.py
index 6c70fbdbeb..7ba0ed3641 100644
--- a/madgraph/iolibs/template_files/mg7/launch.py
+++ b/madgraph/iolibs/template_files/mg7/launch.py
@@ -1,4 +1,5 @@
 import argparse
+import contextlib
 import gc
 import math
 import os
@@ -26,6 +27,8 @@
     MG_ROOT as _MG_ROOT,
     MADSPACE_DIR as _MADSPACE_DIR,
     INSTALL_DIR as _INSTALL_DIR,
+    MADTREX_DIR as _MADTREX_DIR,
+    MADTREX_INSTALL_DIR as _MADTREX_INSTALL_DIR,
 )
 
 _ensure_madspace()
@@ -305,6 +308,11 @@ class HistItem:
 
 
 class MadgraphProcess:
+    # write each event's squared matrix element into the events (the
+    # "matrix_element" column of events.npy), for MadtRex to normalise its
+    # reweighting with instead of evaluating it again; set before survey()
+    store_matrix_elements = False
+
     def __init__(self, extra_make_args=(), split_flavors=False):
         # appended to the make command that builds the matrix elements
         self.extra_make_args = list(extra_make_args)
@@ -1317,7 +1325,8 @@ def build_systematics_matrix_elements(self, args, context):
                 matrix_elements.append(None)
                 continue
             api = context.load_matrix_element(
-                meta["me_path"].format(device=backend), self.param_card_path,
+                meta["me_path"].format(device=backend, fptype=self.fptype),
+                self.param_card_path,
                 me_parameters(self.run_card))
             matrix_elements.append(ms.MatrixElement(
                 api,
@@ -1597,14 +1606,14 @@ def compile_matrix_elements(self) -> list[str]:
             device_names = [device_names]
         cpu_mode = self.run_card["run"]["cpu_mode"]
         backends = [backend_of(name, cpu_mode) for name in device_names]
+        # FPTYPE is part of the library names ({fptype} in me_path)
+        fptype = self.fptype = fptype_of(self.run_card["run"]["precision"])
         if not self.subprocess_data:
             return backends
 
         first_proc_path = self.subprocess_data[0]["path"]
         subproc_path = os.path.dirname(first_proc_path)
 
-        fptype = fptype_of(self.run_card["run"]["precision"])
-
         # Resolve cpu_mode='auto' once (the build rules pick the best SIMD
         # backend available here), so all subprocesses agree on the library names.
         auto_backend = None
@@ -1625,7 +1634,9 @@ def compile_matrix_elements(self) -> list[str]:
             # by library: the flavours of a split subprocess share one
             missing = list({
                 meta["me_path"]: meta for meta in self.subprocess_data
-                if not os.path.isfile(meta["me_path"].format(device=backend))
+                if not os.path.isfile(
+                    meta["me_path"].format(device=backend, fptype=fptype)
+                )
             }.values())
             if not missing:
                 continue
@@ -2003,26 +2014,33 @@ def finish_without_events(self) -> None:
         self.build_systematics()
         self.save_gridpack()
 
-    def generate_events(self) -> None:
+    def generate_events(self, output_format=None) -> None:
+        """Generate the events and write them in output_format, by default the
+        run_card's. run_single() passes "lhe_npy" when MadtRex reweights an
+        "lhe" run: the LHE file is then made from the reweighted npy events
+        (see lhe_via_binary())."""
         if self.gridpack_run_mode != "regular":
             self.finish_without_events()
             return
         start_time = get_start_time()
         self.event_generator.generate()
-        output_format = self.run_card["run"]["output_format"]
+        if output_format is None:
+            output_format = self.run_card["run"]["output_format"]
         systematics = self.build_systematics()
         histograms = self.build_event_histograms()
         if output_format == "compact_npy":
             self.lhe_completer = None
+            self.events_path = os.path.join(self.run_path, "events.npy")
             self.event_generator.combine_to_compact_npy(
-                os.path.join(self.run_path, "events.npy"), systematics, histograms
+                self.events_path, systematics, histograms
             )
             self.write_lhe_header(systematics)
             self.save_lhe_completer()
         elif output_format == "lhe_npy":
             self.lhe_completer = self.build_lhe_completer()
+            self.events_path = os.path.join(self.run_path, "events.npy")
             self.event_generator.combine_to_lhe_npy(
-                os.path.join(self.run_path, "events.npy"), self.lhe_completer,
+                self.events_path, self.lhe_completer,
                 systematics, histograms
             )
             self.write_lhe_header(systematics)
@@ -2040,6 +2058,7 @@ def generate_events(self) -> None:
             # events.lhe.gz, and switches to an external multithreaded tool
             # above 256 MB.
             misc.gzip(lhe_path)
+            self.events_path = _find_event_file(self.run_path)
         else:
             raise ValueError("Unknown output format")
         self.write_hwu(histograms)
@@ -2359,6 +2378,57 @@ def save_gridpack(self) -> None:
                 ignore=shutil.ignore_patterns("include"),
             )
 
+        # MadtRex (event reweighting) is optional and installed separately on
+        # top of madspace (madtrex/install.py), so a missing install is
+        # skipped rather than an error.
+        if self.run_card["gridpack"]["include_madtrex_source"]:
+            shutil.copytree(
+                _MADTREX_DIR,
+                os.path.join(gridpack_path, "madtrex"),
+                ignore=shutil.ignore_patterns("build", "install", "__pycache__"),
+            )
+
+        # madspace/madtrex sources build with their pyproject.toml, which takes
+        # the version from the MadGraph VERSION file one level above them
+        if (self.run_card["gridpack"]["include_madspace_source"]
+                or self.run_card["gridpack"]["include_madtrex_source"]):
+            shutil.copy(_MG_ROOT / "VERSION", os.path.join(gridpack_path, "VERSION"))
+
+        include_madtrex = (
+            self.run_card["gridpack"]["include_madtrex"]
+            and (_MADTREX_INSTALL_DIR / "madtrex").is_dir()
+        )
+        if include_madtrex:
+            # only the runtime pieces: madtrex itself, the bundled Rex fork
+            # (pyrex, plus librex/libtearex in lib/), not the build tools
+            # install.py also put there
+            for name in ("madtrex", "pyrex", "lib"):
+                if (_MADTREX_INSTALL_DIR / name).is_dir():
+                    shutil.copytree(
+                        _MADTREX_INSTALL_DIR / name,
+                        os.path.join(gridpack_path, "madtrex", "install", name),
+                        ignore=shutil.ignore_patterns("__pycache__"),
+                    )
+        elif self.run_card["gridpack"]["include_madtrex"]:
+            logger.info(
+                "madtrex is not installed (python madtrex/install.py), so it is "
+                "not included in the gridpack%s",
+                "; bin/reweight builds it from the included sources on first use"
+                if self.run_card["gridpack"]["include_madtrex_source"] else "")
+        # bin/reweight (below) with an installed madtrex, or with its sources,
+        # which bin/reweight builds when the gridpack has no (usable) install
+        with_reweight = include_madtrex or self.run_card["gridpack"]["include_madtrex_source"]
+        if with_reweight:
+            # the standalone reweighting driver finds the matrix elements
+            # through subprocesses.json, which the gridpack otherwise only has
+            # with include_source
+            gridpack_subprocesses_json = os.path.join(
+                gridpack_path, "SubProcesses", "subprocesses.json")
+            if not os.path.exists(gridpack_subprocesses_json):
+                os.makedirs(os.path.dirname(gridpack_subprocesses_json), exist_ok=True)
+                shutil.copy(os.path.join("SubProcesses", "subprocesses.json"),
+                            gridpack_subprocesses_json)
+
         matrix_elements = []
         for subproc in self.subprocess_data:
             me_path = subproc["me_path"]
@@ -2425,6 +2495,18 @@ def save_gridpack(self) -> None:
             os.path.join(bin_path, "custom_function_loader.py"),
         )
         os.chmod(gen_events_file, 0o755)
+        if with_reweight:
+            # MadtRex reweighting of the gridpack's events, with the modules
+            # parsing its reweight cards
+            reweight_file = os.path.join(bin_path, "reweight")
+            shutil.copy(
+                os.path.join(os.path.dirname(__file__), "madtrex_driver.py"), reweight_file
+            )
+            os.chmod(reweight_file, 0o755)
+            for name in ("param_reweight_card.py", "pdf_reweight_card.py",
+                         "generic_reweight_card.py"):
+                shutil.copy(os.path.join(os.path.dirname(__file__), name),
+                            os.path.join(bin_path, name))
 
         data = {
             "matrix_elements": matrix_elements,
@@ -2848,7 +2930,10 @@ def __init__(
 
         # The libraries were all built up front by MadgraphProcess.compile_matrix_elements
         all_api_paths = [
-            [api_path_format.format(device=backend) for backend in self.process.backends]
+            [
+                api_path_format.format(device=backend, fptype=self.process.fptype)
+                for backend in self.process.backends
+            ]
             for api_path_format in api_path_formats
         ]
 
@@ -3721,6 +3806,7 @@ def build_integrands(
                     pdf2=pdf_arg,
                     input_momentum_fraction=True,
                     decay=self.process.is_decay,
+                    matrix_element_output=self.process.store_matrix_elements,
                 )
             )
         partial_weights = self.process.systematics_enabled()
@@ -3964,31 +4050,92 @@ def do_compute_widths(self, line):
     # running
     # ------------------------------------------------------------------
     def do_generate_events(self, line):
-        """generate_events [-f] [-n NAME] [--laststep=parton]
+        """generate_events [-f] [-n NAME] [--laststep=parton] [--reweight=CARD]
+                           [--reweight-backend=classic|madtrex|auto]
 
-        Run the mg7 generation and the selected post-processing tools."""
+        Run the mg7 generation and the selected post-processing tools.
+        --reweight=CARD is a shortcut for switching reweighting on with CARD
+        as Cards/reweight_card.dat; --reweight-backend chooses who reweights
+        (see madgraph/various/reweight_backend.py)."""
 
         opts = self._parse_run_options(line)
         cwd = os.getcwd()
         target = self.me_dir if self.me_dir != '.' else cwd
+        if opts['reweight_card']:
+            opts['reweight_card'] = os.path.abspath(opts['reweight_card'])
         try:
             os.chdir(target)
             _setup_logging()
             if opts['name']:
                 self._set_run_name(opts['name'])
+            if opts['reweight_card']:
+                install_reweight_card(opts['reweight_card'])
             switch = {}
             if not opts['force']:
                 switch = ask_edit_cards(mother=self)
             switch = self._apply_laststep(switch, opts)
-            force_lhe_output_if_needed(switch)
+            if opts['reweight_card'] and _off(switch.get('reweight')):
+                switch = dict(switch, reweight='ON')
+            reweight = prepare_reweighting(switch, opts['reweight_backend'],
+                                           self.options)
+            force_lhe_output_if_needed(switch, reweight)
             _raise_open_file_limit()
-            run_generation(switch)
+            run_generation(switch, reweight)
         finally:
             os.chdir(cwd)
 
     # ``launch`` is the name MG5 users type; keep it working here too.
     do_launch = do_generate_events
 
+    def do_reweight(self, line):
+        """reweight RUN_NAME [--reweight=CARD] [--reweight-backend=classic|madtrex|auto]
+
+        Reweight the events of an earlier run with Cards/reweight_card.dat
+        (or CARD, which then becomes Cards/reweight_card.dat), as
+        generate_events does after the generation: the weights are added to
+        the run's event file, after the weights already there."""
+        args = self.split_arg(line)
+        run_name, card, backend = None, None, None
+        for arg in args:
+            if arg.startswith('--reweight='):
+                card = os.path.abspath(arg.split('=', 1)[1])
+            elif arg.startswith('--reweight-backend='):
+                backend = arg.split('=', 1)[1]
+            elif arg.startswith('-'):
+                logger.warning("'%s' is not an option of reweight and is ignored", arg)
+            else:
+                run_name = arg
+        cwd = os.getcwd()
+        target = self.me_dir if self.me_dir != '.' else cwd
+        try:
+            os.chdir(target)
+            _setup_logging()
+            if run_name is None:
+                runs = list_runs()
+                raise self.InvalidCmd(
+                    'reweight: give the run to reweight%s'
+                    % (' (%s)' % ', '.join(runs) if runs else ': there is no run yet'))
+            try:
+                run = ExistingRun(run_name)
+            except ValueError as error:
+                raise self.InvalidCmd('reweight: %s' % error)
+            if card:
+                install_reweight_card(card)
+            if not os.path.exists(REWEIGHT_CARD):
+                raise self.InvalidCmd('reweight: %s is missing' % REWEIGHT_CARD)
+            decision = prepare_reweighting({'reweight': 'ON'}, backend, self.options)
+            if decision is None:
+                return
+            run_reweighting(run, REWEIGHT_CARD, decision=decision)
+        finally:
+            os.chdir(cwd)
+
+    def complete_reweight(self, text, line, begidx, endidx):
+        me_dir = self.me_dir if self.me_dir != '.' else os.getcwd()
+        options = ['--reweight=', '--reweight-backend=']
+        return self.list_completion(text, list_runs(os.path.join(me_dir, 'Events'))
+                                    + options, line)
+
     def do_npy_to_lhe(self, line):
         """npy_to_lhe [RUN_NAME] [-o OUTPUT] [--seed N] [--no-gzip]
 
@@ -4014,7 +4161,8 @@ def do_quit(self, line):
     def _parse_run_options(self, line):
         """Understand the subset of MG5's `launch` options that mean something
         for an mg7 run, and say so out loud for the ones that do not."""
-        out = {'force': False, 'name': '', 'laststep': ''}
+        out = {'force': False, 'name': '', 'laststep': '', 'reweight_card': None,
+               'reweight_backend': None}
         args = self.split_arg(line)
         index = 0
         while index < len(args):
@@ -4032,6 +4180,16 @@ def _parse_run_options(self, line):
                 index += 1
             elif arg.startswith(('-s=', '--laststep=')):
                 out['laststep'] = arg.split('=', 1)[1]
+            elif arg == '--reweight' and index < len(args):
+                out['reweight_card'] = args[index]
+                index += 1
+            elif arg.startswith('--reweight='):
+                out['reweight_card'] = arg.split('=', 1)[1]
+            elif arg == '--reweight-backend' and index < len(args):
+                out['reweight_backend'] = args[index]
+                index += 1
+            elif arg.startswith('--reweight-backend='):
+                out['reweight_backend'] = arg.split('=', 1)[1]
             elif arg.startswith('-'):
                 logger.warning(
                     "'%s' is not supported by the mg7 output and is ignored; "
@@ -4496,8 +4654,8 @@ def selected_tools(switch) -> list:
     # analysis == 'MadAnalysis5' and shower != 'OFF').
     ma5 = switch.get("analysis") == "MadAnalysis5"
     showered = not _off(switch.get("shower"))
+    # reweighting is not one of them: run_single runs it (run_reweighting)
     return [t for t, on in (
-        ("reweighting", not _off(switch.get("reweight"))),
         ("MadSpin", not _off(switch.get("madspin"))),
         ("MadAnalysis5 (parton level)", ma5),
         ("Pythia8 shower", switch.get("shower") == "Pythia8"),
@@ -4516,9 +4674,10 @@ def run_selected_tools(switch, process) -> None:
     output reuses the *exact* madevent tool drivers -- in particular Pythia8 is
     showered in parallel over the cluster/multicore backend, identically to a
     madevent run -- instead of re-implementing each tool here. The command
-    sequence mirrors ``MadEventCmd.do_launch`` (reweight, MadSpin, MA5 parton,
-    shower, Delphes, MA5 hadron, Rivet); each tool checks its own card, so a
-    tool is invoked only when its switch is on.
+    sequence mirrors ``MadEventCmd.do_launch`` (MadSpin, MA5 parton, shower,
+    Delphes, MA5 hadron, Rivet); each tool checks its own card, so a tool is
+    invoked only when its switch is on. Reweighting, which comes first in
+    that sequence, already ran (run_reweighting, from run_single).
     """
     log = logging.getLogger("madevent")
 
@@ -4570,12 +4729,10 @@ def run(tool, command):
             log.info("  -> %s done (%.1fs)", tool, time.time() - start)
 
     # Same order as MadEventCmd.do_launch. The commands take no run name and so
-    # act on cmd.run_name: reweight runs on the generated events, then
-    # decay_events (MadSpin) repoints cmd.run_name to the "<run>_decayed_i" run,
-    # so the shower and the analyses that follow act on the decayed events --
-    # exactly as a madevent run does. Each driver still guards on its own card.
-    if not _off(switch.get("reweight")):
-        run("reweight", "reweight -from_cards")
+    # act on cmd.run_name: decay_events (MadSpin) repoints cmd.run_name to the
+    # "<run>_decayed_i" run, so the shower and the analyses that follow act on
+    # the decayed events -- exactly as a madevent run does. Each driver still
+    # guards on its own card.
     if not _off(switch.get("madspin")):
         run("MadSpin", "decay_events -from_cards")
     if ma5:
@@ -5055,18 +5212,348 @@ def _release_channel_generators(process) -> None:
         shutil.rmtree(process.temp_path, ignore_errors=True)
 
 
-def run_single(switch=None) -> "MadgraphProcess":
-    """Run a single generation and return the process (for its result)."""
+REWEIGHT_CARD = os.path.join("Cards", "reweight_card.dat")
+
+
+class ExistingRun:
+    """The run directory Events/<run_name> of an earlier generation, with
+    what run_reweighting() needs of a MadgraphProcess: run_path, events_path
+    (the file the generation wrote: events.npy for the npy formats, else
+    events.lhe[.gz]) and run_card["run"]["output_format"] (from the files)."""
+
+    def __init__(self, run_name, events_dir="Events"):
+        self.run_path = os.path.abspath(os.path.join(events_dir, run_name))
+        if not os.path.isdir(self.run_path):
+            runs = list_runs(events_dir)
+            raise ValueError("no run '%s' in %s%s" % (
+                run_name, events_dir,
+                "; runs: %s" % ", ".join(runs) if runs else ""))
+        npy = os.path.join(self.run_path, npy_to_lhe.EVENT_FILE)
+        if os.path.isfile(npy):
+            self.events_path = npy
+            compact = os.path.isfile(os.path.join(self.run_path,
+                                                  npy_to_lhe.COMPLETER_FILE))
+            output_format = "compact_npy" if compact else "lhe_npy"
+        else:
+            self.events_path = _find_event_file(self.run_path)
+            output_format = "lhe"
+        if self.events_path is None:
+            raise ValueError("no event file (events.npy, events.lhe[.gz]) in %s"
+                             % self.run_path)
+        self.run_card = {"run": {"output_format": output_format}}
+
+
+def list_runs(events_dir="Events") -> list:
+    """The runs of events_dir with an event file, sorted by name."""
+    if not os.path.isdir(events_dir):
+        return []
+    return sorted(
+        name for name in os.listdir(events_dir)
+        if any(os.path.isfile(os.path.join(events_dir, name, f)) for f in
+               (npy_to_lhe.EVENT_FILE, "events.lhe", "events.lhe.gz")))
+
+
+def install_reweight_card(path) -> None:
+    """``--reweight=CARD``: CARD becomes Cards/reweight_card.dat. A different
+    card already there is kept as reweight_card.dat.bak."""
+    if os.path.exists(REWEIGHT_CARD):
+        if os.path.samefile(path, REWEIGHT_CARD):
+            return
+        with open(path) as new, open(REWEIGHT_CARD) as old:
+            if new.read() == old.read():
+                return
+        shutil.copy(REWEIGHT_CARD, REWEIGHT_CARD + ".bak")
+        logger.info("--reweight: %s saved as %s.bak", REWEIGHT_CARD, REWEIGHT_CARD)
+    shutil.copy(path, REWEIGHT_CARD)
+    logger.info("--reweight: using %s as %s", path, REWEIGHT_CARD)
+
+
+# reweight switch values naming a backend (ON is auto)
+_SWITCH_BACKENDS = {"classic": "classic", "density": "classic", "madtrex": "madtrex"}
+
+
+def prepare_reweighting(switch, cli_backend=None, options=None):
+    """The reweight_backend.Decision for this run, or None when it does not
+    reweight: the reweight switch is off, or Cards/reweight_card.dat is
+    missing or has no "launch". Chosen before the generation, so that a card
+    that cannot run fails before the events are generated, and the event
+    format the backend needs is known (force_lhe_output_if_needed).
+
+    Precedence: cli_backend (--reweight-backend), then a switch value naming
+    a backend (classic/density/madtrex; ON is auto), then the card's
+    "change backend" line; auto otherwise."""
+    from madgraph.various import reweight_backend as rb
+
+    value = (switch or {}).get("reweight")
+    if _off(value) or not os.path.exists(REWEIGHT_CARD):
+        return None
+    with open(REWEIGHT_CARD) as f:
+        traits = rb.classify_card(f.read())
+    if not traits.launches:
+        logger.info("reweighting is on but %s has no 'launch': nothing to reweight",
+                    REWEIGHT_CARD)
+        return None
+    plugin = value if value not in ("ON", "classic", "madtrex", "density") else None
+    # a plugin is a classic reweighting module
+    switch_backend = "classic" if plugin else _SWITCH_BACKENDS.get(value)
+    requested = [("--reweight-backend", cli_backend),
+                 ("the reweight switch (%s)" % value, switch_backend)]
+    if value == "density":
+        traits.density = True
+    if options is None:
+        options = load_mg5_options()
+    env = rb.probe_environment(os.getcwd(), options)
+    decision = rb.select_backend(requested, traits, env, plugin=plugin)
+    for text in decision.warnings:
+        logger.warning(text)
+    return decision
+
+
+@contextlib.contextmanager
+def _stripped_reweight_card(path):
+    """path without its "change backend" line for the duration (the classic
+    module reads the card itself); restored afterwards."""
+    from madgraph.various import reweight_backend as rb
+    with open(path) as f:
+        text = f.read()
+    stripped = rb.strip_backend_line(text)
+    if stripped == text:
+        yield
+        return
+    with open(path, "w") as f:
+        f.write(stripped)
+    try:
+        yield
+    finally:
+        with open(path, "w") as f:
+            f.write(text)
+
+
+# the <weightgroup name=...> each backend writes into the LHE header
+REWEIGHT_WEIGHT_GROUPS = {"classic": "mg_reweighting", "madtrex": "tearex_reweighting"}
+
+
+def write_reweighting_record(run_path, decision, card_path, events=None) -> str:
+    """Record who reweighted the run, and why, in <run>/reweighting.json:
+    backend, how it was chosen (source, reasons, warnings), the card kind,
+    the event file the weights went into (events), the LHE weight group it
+    wrote and the card itself. Returns the path."""
+    with open(card_path) as f:
+        card = f.read()
+    record = {
+        "backend": decision.backend,
+        "source": decision.source,
+        "reasons": [r.replace("{card}", os.path.basename(card_path))
+                    for r in decision.reasons],
+        "warnings": list(decision.warnings),
+        "card_kind": decision.card_kind,
+        "plugin": decision.plugin,
+        "weight_group": REWEIGHT_WEIGHT_GROUPS[decision.backend],
+        "events": events,
+        "card": card,
+    }
+    path = os.path.join(run_path, "reweighting.json")
+    with open(path, "w") as f:
+        json.dump(record, f, indent=1)
+    return path
+
+
+def drop_stored_matrix_elements(events_path) -> None:
+    """Removes the "matrix_element" column (MadgraphProcess.
+    store_matrix_elements) from the npy events at events_path, if they have
+    one: it is only there for MadtRex's normalisation, not part of the
+    run's output."""
+    import numpy as np
+    from numpy.lib import format as npy_format
+
+    if not events_path.endswith(".npy") or not os.path.isfile(events_path):
+        return
+    events = np.load(events_path, mmap_mode="r", max_header_size=1 << 30)
+    if "matrix_element" not in (events.dtype.names or ()):
+        return
+    names = [name for name in events.dtype.names if name != "matrix_element"]
+    descr = [(name, events.dtype.fields[name][0]) for name in names]
+    tmp = events_path + ".tmp"
+    out = npy_format.open_memmap(tmp, mode="w+", dtype=np.dtype(descr), shape=events.shape)
+    chunk = 100000
+    for start in range(0, len(events), chunk):
+        stop = min(start + chunk, len(events))
+        for name in names:
+            out[name][start:stop] = events[name][start:stop]
+    out.flush()
+    del out, events
+    os.replace(tmp, events_path)
+
+
+def run_reweighting(process, card_path=REWEIGHT_CARD, requested=None, decision=None):
+    """Reweight the events of process with card_path, once, by the backend
+    of decision (prepare_reweighting()), or of select_backend(requested)
+    when no decision is given. The classic module runs through MG7RunCmd
+    ("reweight -from_cards", on the LHE file); MadtRex through
+    madtrex.run_reweighting(), on the events as written (lhe or npy)."""
+    from madgraph.various import reweight_backend as rb
+
+    log = logging.getLogger("madevent")
+    if decision is None:
+        with open(card_path) as f:
+            traits = rb.classify_card(f.read())
+        env = rb.probe_environment(os.getcwd(), load_mg5_options())
+        decision = rb.select_backend(requested, traits, env)
+    log.info(decision.describe(os.path.basename(card_path)))
+    run_dir = process.run_path
+    # both backends add their weights to the run's own event file (in place,
+    # after the systematics weights): events.npy or events.lhe[.gz]
+
+    if decision.backend == "madtrex":
+        if process.run_card["run"]["output_format"] == "compact_npy":
+            raise ValueError(
+                "MadtRex reweighting requires run_card [run] output_format to be "
+                "'lhe' or 'lhe_npy': 'compact_npy' discards the per-event flavor/"
+                "helicity information MadtRex needs to match events to the "
+                "matrix element library")
+        from madgraph.iolibs.template_files.mg7.madtrex import (
+            run_reweighting as madtrex_reweighting)
+        try:
+            with tempfile.TemporaryDirectory() as tmp:
+                card = os.path.join(tmp, "reweight_card.dat")
+                with open(card_path) as f, open(card, "w") as out:
+                    out.write(rb.strip_backend_line(f.read()))
+                # the stored matrix elements leave events.npy with the
+                # reweighting's last weight merge
+                madtrex_reweighting(card, events_path=process.events_path,
+                                    card_kind=decision.card_kind,
+                                    process_directory=os.getcwd(),
+                                    drop_stored_matrix_elements=True)
+        finally:
+            # and here only when that merge did not happen (a failure)
+            drop_stored_matrix_elements(process.events_path)
+        tool_copy = os.path.join(run_dir, "unweighted_events.lhe.gz")
+        if not process.events_path.endswith(".npy") and os.path.exists(tool_copy):
+            from madgraph.iolibs.template_files.mg7.run_interface import sync_lhe_copy
+            sync_lhe_copy(process.events_path, tool_copy)
+        log.info("reweighting weights written to %s", process.events_path)
+        write_reweighting_record(run_dir, decision, card_path, process.events_path)
+        return decision
+
+    lhe_path = _find_or_convert_event_file(run_dir, log)
+    if lhe_path is None:
+        raise RuntimeError("No LHE event file in %s: the classic reweighting "
+                           "module needs one" % run_dir)
+    run_name = os.path.basename(os.path.dirname(os.path.abspath(lhe_path)))
+    from madgraph.iolibs.template_files.mg7.run_interface import MG7RunCmd
+    cmd = MG7RunCmd(os.getcwd(), load_mg5_options(), run_name, lhe_path)
+    cmd.set_run_name(run_name, None, "parton")
+    command = "reweight --backend=classic --mode=%s -from_cards" % (
+        "density" if decision.card_kind == "density" else "ON")
+    if decision.plugin:
+        command += " --plugin=%s" % decision.plugin
+    with _stripped_reweight_card(card_path):
+        cmd.exec_cmd(command, postcmd=False, printcmd=False)
+    try:
+        cmd.store_result()
+    except Exception as error:
+        _report_failure(log, "store_result", error, run_dir)
+    # the classic module's weights into the run's event file (lhe_path)
+    cmd.sync_reweighted_events()
+    log.info("reweighting weights written to %s", lhe_path)
+    write_reweighting_record(run_dir, decision, card_path, lhe_path)
+    return decision
+
+
+def lhe_via_binary(process, reweight) -> bool:
+    """Whether an output_format = "lhe" run that MadtRex reweights is
+    generated as npy ("lhe_npy") and turned into the LHE file it asked for
+    only after the reweighting (deliver_lhe_from_binary()): MadtRex reads
+    and extends the npy events much faster than an LHE file, which it would
+    decompress, parse and rewrite in full. Not when an LHE-only step runs
+    before the reweighting: the run_card [postprocessing] time_of_flight or
+    legacy systematics (see run_lhe_postprocessing())."""
+    if getattr(reweight, "backend", None) != "madtrex":
+        return False
+    if process.gridpack_run_mode != "regular":
+        return False
+    if process.run_card["run"]["output_format"] != "lhe":
+        return False
+    cfg = process.run_card["postprocessing"]
+    try:
+        if float(cfg.get("time_of_flight", -1.0)) >= 0:
+            return False
+    except (TypeError, ValueError):
+        return False
+    if cfg.get("systematics") and not process.run_card["systematics"]["enable"]:
+        return False
+    return True
+
+
+def deliver_lhe_from_binary(process) -> None:
+    """The LHE file of a run lhe_via_binary() generated as npy: npy_to_lhe
+    writes Events/<run>/events.lhe.gz from events.npy and header.lhe, with
+    their weights (the systematics', MadtRex's), and it replaces them as the
+    run's event file (process.events_path, and reweighting.json's events).
+    When the conversion fails the npy events are kept, and reported."""
+    log = logging.getLogger("madevent")
+    npy_path = process.events_path
+    process.lhe_from_binary = False
+    try:
+        lhe_path = npy_to_lhe.convert(npy_path)
+    except Exception as error:
+        _report_failure(log, "npy->LHE conversion", error, process.run_path)
+        log.warning("the events of %s are kept as %s", os.path.basename(process.run_path),
+                    npy_path)
+        return
+    os.remove(npy_path)
+    os.remove(os.path.join(process.run_path, npy_to_lhe.HEADER_FILE))
+    process.events_path = lhe_path
+    record = os.path.join(process.run_path, "reweighting.json")
+    if os.path.isfile(record):
+        with open(record) as f:
+            content = json.load(f)
+        if content.get("events") == npy_path:
+            content["events"] = lhe_path
+            with open(record, "w") as f:
+                json.dump(content, f, indent=1)
+
+
+def run_single(switch=None, reweight=None) -> "MadgraphProcess":
+    """Run a single generation and return the process (for its result).
+    reweight is the reweight_backend.Decision of prepare_reweighting(), or
+    None: reweighting runs once, after the run_card post-processing
+    (systematics) and before the post-processing tools."""
     compute_auto_widths()
     process = MadgraphProcess()
+    process.store_matrix_elements = getattr(reweight, "backend", None) == "madtrex"
     process.survey()
     process.train_madnis()
-    process.generate_events()
+    process.lhe_from_binary = lhe_via_binary(process, reweight)
+    if process.lhe_from_binary:
+        logging.getLogger("madevent").info(
+            "MadtRex reweights this run: the events are written as npy, and as "
+            "the requested LHE file once reweighted")
+    process.generate_events("lhe_npy" if process.lhe_from_binary else None)
     if process.gridpack_run_mode != "regular":
         return process
     # run_card-driven LHE post-processing (displaced vertex + systematics)
     run_lhe_postprocessing(process)
-    # run the post-processing tools (Pythia8/Delphes/MadSpin/reweight/analysis)
+    # reweighting, after the systematics, by the backend chosen beforehand.
+    # Like the post-processing tools, a failure is reported without losing
+    # the rest of the run (MadSpin, shower, analyses): the event file is only
+    # ever replaced atomically, so the events are kept, without the weights.
+    if reweight is not None:
+        try:
+            run_reweighting(process, REWEIGHT_CARD, decision=reweight)
+        except Exception as error:
+            log = logging.getLogger("madevent")
+            _report_failure(log, "reweighting", error, process.run_path)
+            log.warning(
+                "the events of %s are kept without the reweighting weights; to "
+                "reweight them once the problem is fixed: 'launch %s -i' in "
+                "MadGraph, then 'reweight %s'", os.path.basename(process.run_path),
+                os.getcwd(), os.path.basename(process.run_path))
+    # the LHE file of an "lhe" run MadtRex reweighted as npy, with the
+    # reweighting weights, or without them when the reweighting failed
+    if process.lhe_from_binary:
+        deliver_lhe_from_binary(process)
+    # run the post-processing tools (Pythia8/Delphes/MadSpin/analysis)
     # selected in the merged question above on the generated events.
     if switch:
         run_selected_tools(switch, process)
@@ -5099,12 +5586,13 @@ def detect_param_scan(param_card_path):
     return None
 
 
-def run_scan(iterator, card_path, switch=None) -> None:
+def run_scan(iterator, card_path, switch=None, reweight=None) -> None:
     """Iterate over all scan points, running a full generation for each and
     accumulating the results, then write the scan summary. Works for both the
     run_card (RunCardIterator) and the param_card (ParamCardIterator); their
     interface (__iter__/write/store_entry/get_next_name/write_summary) is the
-    same. The scan card is restored afterwards."""
+    same. The scan card is restored afterwards. reweight, if given, is
+    forwarded to run_single() so every scan point is reweighted in turn."""
     import tomllib
     with open(os.path.join("Cards", "run_card.toml"), "rb") as f:
         run_name = tomllib.load(f).get("run", {}).get("run_name", "run")
@@ -5118,7 +5606,7 @@ def run_scan(iterator, card_path, switch=None) -> None:
         for i, point in enumerate(iterator):
             point.write(card_path)
             logger.info("=== scan point %d ===", i + 1)
-            process = run_single(switch)
+            process = run_single(switch, reweight)
             # use the run directory the process actually created, so the
             # per-point params.dat written by write_summary has a home
             name = os.path.basename(process.run_path)
@@ -5138,9 +5626,11 @@ def run_scan(iterator, card_path, switch=None) -> None:
         shutil.move(backup, card_path)
 
 
-def run_generation(switch=None) -> None:
+def run_generation(switch=None, reweight=None) -> None:
     """Run the generation, expanding a scan over the run_card or the param_card
-    when one is present (scanning both simultaneously is not allowed)."""
+    when one is present (scanning both simultaneously is not allowed).
+    reweight (see prepare_reweighting()), if given, is forwarded to
+    run_single()/run_scan() to reweight after each generation."""
     run_card_path = os.path.join("Cards", "run_card.toml")
     param_card_path = os.path.join("Cards", "param_card.dat")
     run_iter = detect_run_scan(run_card_path)
@@ -5150,18 +5640,21 @@ def run_generation(switch=None) -> None:
             "Scanning simultaneously over the run_card and the param_card is "
             "not allowed. Please keep the scan:[...] entries in only one card.")
     if run_iter:
-        run_scan(run_iter, run_card_path, switch)
+        run_scan(run_iter, run_card_path, switch, reweight)
     elif param_iter:
-        run_scan(param_iter, param_card_path, switch)
+        run_scan(param_iter, param_card_path, switch, reweight)
     else:
-        _release_channel_generators(run_single(switch))
+        _release_channel_generators(run_single(switch, reweight))
 
 
-def _lhe_needed_by(switch, card) -> "str | None":
+def _lhe_needed_by(switch, card, reweight=None) -> "str | None":
     """Why the run needs an LHE event file, or None when it does not: every
-    post-processing tool run_selected_tools drives and the run_card
-    [postprocessing] steps read the LHE file. ``card`` is the raw TOML content
-    of the run_card."""
+    post-processing tool run_selected_tools drives, the classic reweighting
+    module and the run_card [postprocessing] steps read the LHE file. MadtRex
+    also reads the npy events. ``card`` is the raw TOML content of the
+    run_card, ``reweight`` the prepare_reweighting() decision."""
+    if reweight is not None and reweight.backend == "classic":
+        return "reweighting (classic)"
     tools = selected_tools(switch) if switch else []
     if tools:
         return tools[0]
@@ -5179,7 +5672,7 @@ def _lhe_needed_by(switch, card) -> "str | None":
     return None
 
 
-def force_lhe_output_if_needed(switch) -> None:
+def force_lhe_output_if_needed(switch, reweight=None) -> None:
     """The events are written as npy by default; switch the run_card to the
     LHE format when a selected post-processing needs an LHE file. The card is
     rewritten rather than refused, so that asking for a shower in the launch
@@ -5199,7 +5692,7 @@ def force_lhe_output_if_needed(switch) -> None:
         "output_format", RunCardMG7()["run"]["output_format"])
     if output_format == "lhe":
         return
-    reason = _lhe_needed_by(switch, card)
+    reason = _lhe_needed_by(switch, card, reweight)
     if reason is None:
         return
     with open(path) as f:
@@ -5250,6 +5743,17 @@ def main() -> None:
     parser.add_argument(
         "--make-arg", action="append", default=[],
         help="extra argument of the matrix element build (--decay-widths)")
+    parser.add_argument(
+        "--reweight", dest="reweight_card", default=None,
+        help="reweight the events after generation with this reweight card "
+             "(it becomes Cards/reweight_card.dat and switches reweighting on)",
+    )
+    parser.add_argument(
+        "--reweight-backend", dest="reweight_backend", default=None,
+        choices=("classic", "madtrex", "auto"),
+        help="who reweights: the classic module, MadtRex, or auto (the default: "
+             "MadtRex when it can run the card)",
+    )
     args = parser.parse_args()
 
     if args.decay_widths:
@@ -5260,4 +5764,9 @@ def main() -> None:
             json.dump(results, f)
         return
 
-    MG7Cmd().run_cmd("generate_events" if args.ask_edit_cards else "generate_events -f")
+    line = "generate_events" if args.ask_edit_cards else "generate_events -f"
+    if args.reweight_card:
+        line += " --reweight=%s" % args.reweight_card
+    if args.reweight_backend:
+        line += " --reweight-backend=%s" % args.reweight_backend
+    MG7Cmd().run_cmd(line)
diff --git a/madgraph/iolibs/template_files/mg7/madtrex.py b/madgraph/iolibs/template_files/mg7/madtrex.py
new file mode 100644
index 0000000000..6b154f4fe1
--- /dev/null
+++ b/madgraph/iolibs/template_files/mg7/madtrex.py
@@ -0,0 +1,154 @@
+"""MadtRex reweighting from within MadGraph.
+
+The reweighting itself lives in madtrex_driver.py, which runs without
+MadGraph (and is what gridpacks ship as bin/reweight). This module is the
+MadGraph side of it: it adds what does need MadGraph -- regenerating the
+process in another model for a "change model" reweight card, via
+bin/madgraph -- and is what launch.py's --reweight uses.
+
+Run with: python -m madgraph.iolibs.template_files.mg7.madtrex --help
+"""
+
+import json
+import logging
+import os
+import shutil
+import subprocess
+import tempfile
+from pathlib import Path
+
+import madgraph as _mg_pkg
+
+from . import param_reweight_card as param_rwgt_card
+from . import madtrex_driver as _driver
+from .madtrex_driver import (  # noqa: F401 -- re-exported
+    ms,
+    mtx,
+    find_latest_run_dir,
+    find_run_events,
+    prepare_events,
+    read_init_node,
+    resolve_pdf_set_files,
+    default_output_path,
+    run_pdf_reweighting,
+)
+
+_MG_ROOT = Path(_mg_pkg.__file__).parents[1]
+
+logger = logging.getLogger("madgraph7")
+
+
+def regenerate_process_in_model(
+    process_directory: str,
+    model_spec: str,
+    process_definition: str,
+    rw_subdir: str = "rw_me",
+    fptype: str = param_rwgt_card.DEFAULT_FPTYPE,
+) -> str:
+    """Regenerates process_definition in model_spec (a UFO model name, plus
+    any further "import model" arguments, eg a restriction card) into
+    process_directory/rw_subdir via a headless bin/madgraph run -- same
+    subprocess-with-a-command-file pattern as compute_auto_widths() in
+    launch.py -- then compiles every resulting subprocess, at fptype, for the
+    devices process_directory's own run_card.toml requests (see
+    madtrex_driver._compile_subprocesses()).
+
+    Skips both steps if rw_subdir was already regenerated for this exact
+    (model_spec, process_definition) pair (tracked via a marker file), so
+    repeated reweighting runs against the same alternate model don't pay the
+    regenerate cost every time -- note this means fptype only actually
+    drives a compile on whichever call first regenerates a given rw_subdir;
+    a later call requesting a different fptype against an already-cached
+    rw_subdir gets the directory back unchanged, uncompiled for that fptype
+    (run_parameter_reweighting() accounts for this itself by always calling
+    _compile_subprocesses() again for the fptype it actually needs,
+    regardless of whether this function regenerated anything). Returns
+    rw_subdir's full path, ready to pass to Driver.load_process_directory()."""
+    rw_dir = os.path.join(process_directory, rw_subdir)
+    marker_path = os.path.join(rw_dir, ".madtrex_source.json")
+    # The libraries are built with a LIBTAG of their own (see madmatrix.mk):
+    # they are loaded next to the original process's, for the
+    # normalisation, and the dynamic loader would take one of the same name
+    # (the same model's common library, say) for the original's.
+    libtag = "_madtrex_" + _driver._sanitize_dirname(rw_subdir)
+    marker = {"model": model_spec, "process": process_definition, "libtag": libtag}
+    if os.path.isfile(marker_path):
+        with open(marker_path) as f:
+            if json.load(f) == marker:
+                return rw_dir
+        shutil.rmtree(rw_dir)
+
+    mg5 = str(_MG_ROOT / "bin" / "madgraph")
+    if not os.path.exists(mg5):
+        raise RuntimeError(
+            f"cannot find bin/madgraph at {mg5}; needed to regenerate "
+            f"'{process_definition}' in model '{model_spec}'"
+        )
+    cmds = (
+        f"import model {model_spec}\n"
+        f"generate {process_definition}\n"
+        f"output mg7 {os.path.abspath(rw_dir)} -f\n"
+    )
+    with tempfile.NamedTemporaryFile("w", suffix=".mg5", delete=False) as fh:
+        fh.write(cmds)
+        cmdfile = fh.name
+    try:
+        logger.info(
+            "MadtRex: regenerating '%s' in model '%s' -> %s",
+            process_definition, model_spec, rw_dir,
+        )
+        proc = subprocess.run([mg5, cmdfile], capture_output=False, text=False, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
+        if proc.returncode != 0:
+            raise RuntimeError(
+                f"regenerating '{process_definition}' in model '{model_spec}' failed "
+                f"(bin/madgraph exited with {proc.returncode})"
+            )
+    finally:
+        os.remove(cmdfile)
+
+    # the libraries' names with the LIBTAG, for whoever loads them
+    json_path = os.path.join(rw_dir, "SubProcesses", "subprocesses.json")
+    with open(json_path) as f:
+        subprocesses = json.load(f)
+    for entry in subprocesses:
+        directory, name = os.path.split(entry["me_path"])
+        if not name.startswith("libmadmatrix_"):
+            raise RuntimeError(f"unexpected matrix element library name {entry['me_path']!r}")
+        entry["me_path"] = os.path.join(directory, "libmadmatrix" + libtag + name[len("libmadmatrix"):])
+    with open(json_path, "w") as f:
+        json.dump(subprocesses, f)
+    # written before compiling: _compile_subprocesses() reads the LIBTAG from it
+    with open(marker_path, "w") as f:
+        json.dump(marker, f)
+    _driver._compile_subprocesses(rw_dir, _driver._read_devices(process_directory), fptype)
+    return rw_dir
+
+
+def run_parameter_reweighting(*args, regenerate=regenerate_process_in_model, **kwargs):
+    """madtrex_driver.run_parameter_reweighting(), with "change model"
+    support (see regenerate_process_in_model())."""
+    return _driver.run_parameter_reweighting(*args, regenerate=regenerate, **kwargs)
+
+
+def run_generic_reweighting(*args, regenerate=regenerate_process_in_model, **kwargs):
+    """madtrex_driver.run_generic_reweighting(), with "change model"
+    support (see regenerate_process_in_model())."""
+    return _driver.run_generic_reweighting(*args, regenerate=regenerate, **kwargs)
+
+
+def run_reweighting(card, events_path=None, card_kind=None, regenerate=regenerate_process_in_model,
+                    **kwargs):
+    """madtrex_driver.run_reweighting(): the reweight card through the
+    parameter, PDF or generic runner its card_kind ("param", "pdf",
+    "generic"; read from the card when None) needs, with "change model"
+    support."""
+    return _driver.run_reweighting(card, events_path=events_path, card_kind=card_kind,
+                                   regenerate=regenerate, **kwargs)
+
+
+def main(argv: list[str] | None = None) -> None:
+    _driver.main(argv, regenerate=regenerate_process_in_model)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/madgraph/iolibs/template_files/mg7/madtrex_driver.py b/madgraph/iolibs/template_files/mg7/madtrex_driver.py
new file mode 100644
index 0000000000..7a28df97de
--- /dev/null
+++ b/madgraph/iolibs/template_files/mg7/madtrex_driver.py
@@ -0,0 +1,1640 @@
+#! /usr/bin/env python3
+"""Standalone MadtRex reweighting driver.
+
+Reweights the events of an MG7 process directory -- or of a gridpack, which
+ships this file as bin/reweight together with the three *_reweight_card.py
+modules it imports -- using only the madspace and madtrex packages: nothing
+here imports MadGraph. mg7/madtrex.py is the MadGraph-side wrapper around
+this module, which adds what does need MadGraph (regenerating the process
+in another model for a "change model" reweight card).
+
+Usage (from a process directory or gridpack, or see --process-dir):
+    python madtrex_driver.py --param-rwgt-card param_reweight_card.dat
+    bin/reweight --rwgt-card reweight_card.dat            # in a gridpack
+"""
+
+# annotations stay strings: some name madtrex classes (mtx.Driver), which
+# Python < 3.14 would otherwise evaluate when this module is imported
+from __future__ import annotations
+
+import argparse
+import glob
+import gzip
+import json
+import logging
+import os
+import re
+import shutil
+import subprocess
+import sys
+import tempfile
+import tomllib
+from pathlib import Path
+from typing import Callable
+
+# The madspace and madtrex installations sit next to each other in both
+# layouts this file runs from: <MG>/{madspace,madtrex}/install, with this file
+# in <MG>/madgraph/iolibs/template_files/mg7/, and <gridpack>/{madspace,
+# madtrex}/install, with this file in <gridpack>/bin/. Use the first
+# directory above this file that has either; failing that, whatever madspace/
+# madtrex are already importable. Nothing here installs them.
+_HERE = Path(__file__).resolve().parent
+
+
+def _find_install_root() -> Path | None:
+    for parent in _HERE.parents:
+        if (parent / "madspace" / "install" / "madspace").is_dir() or (
+            parent / "madtrex" / "install" / "madtrex"
+        ).is_dir():
+            return parent
+    return None
+
+
+def _ensure_gridpack_madtrex() -> None:
+    """In a gridpack made with include_madtrex_source (this file is its
+    bin/reweight, next to madtrex/install.py): build madtrex from those
+    sources when the gridpack has no madtrex install, or rebuild it when the
+    install was compiled against another madspace (e.g. a madspace the
+    gridpack rebuilt on this machine). Elsewhere, nothing."""
+    root = _HERE.parent
+    installer = root / "madtrex" / "install.py"
+    if _HERE.name != "bin" or not installer.is_file():
+        return
+    package = root / "madtrex" / "install" / "madtrex"
+    madspace_package = root / "madspace" / "install" / "madspace"
+    clean = False
+    if not package.is_dir():
+        reason = "this gridpack has no madtrex install"
+    else:
+        import importlib.util
+        spec = importlib.util.spec_from_file_location("_madtrex_compat",
+                                                      package / "_compat.py")
+        if spec is None or not madspace_package.is_dir():
+            return
+        compat = importlib.util.module_from_spec(spec)
+        spec.loader.exec_module(compat)
+        if not compat.build_problem(str(package), str(madspace_package)):
+            return
+        reason = "this gridpack's madtrex was built against another madspace"
+        clean = True
+    if not (madspace_package / "include").is_dir():
+        raise RuntimeError(
+            f"{reason}, and it cannot be built from the included sources: its "
+            "madspace has no headers (a precompiled copy). Make the gridpack "
+            "with include_madspace = false and include_madspace_source = true, "
+            "so that madspace is built here too.")
+    print(f"{reason}: building it from the included sources (madtrex/install.py)")
+    command = [sys.executable, str(installer), "--yes"] + (["--clean"] if clean else [])
+    if subprocess.run(command).returncode != 0:
+        raise RuntimeError("madtrex installation failed: see the output above")
+
+
+_ensure_gridpack_madtrex()
+_INSTALL_ROOT = _find_install_root()
+if _INSTALL_ROOT is not None:
+    for _dir, _pkg in (
+        (_INSTALL_ROOT / "madtrex" / "install", "madtrex"),
+        (_INSTALL_ROOT / "madspace" / "install", "madspace"),
+    ):
+        if (_dir / _pkg).is_dir() and str(_dir) not in sys.path:
+            sys.path.insert(0, str(_dir))
+
+_MISSING_MESSAGE = (
+    "MadtRex reweighting needs the madspace and madtrex packages (madtrex "
+    "is installed separately, on top of madspace: python madtrex/install.py "
+    "in MadGraph, and include_madtrex in the run_card for a gridpack), or "
+    "both on PYTHONPATH. ({})"
+)
+
+
+class _MissingMadtrex:
+    """Stands for the madtrex module when it is not installed: this module
+    still imports (MadGraph imports it to choose and drive the reweighting),
+    and using madtrex raises the ImportError saying how to install it."""
+
+    def __init__(self, reason: str):
+        self._reason = reason
+
+    def __getattr__(self, name):
+        raise ImportError(_MISSING_MESSAGE.format(self._reason))
+
+
+try:
+    import madspace as ms
+except ImportError as _err:
+    raise ImportError(_MISSING_MESSAGE.format(_err)) from _err
+try:
+    import madtrex as mtx
+    # with the MadGraph directory on sys.path, the madtrex/ source directory
+    # imports as an empty namespace package when madtrex is not installed
+    if not hasattr(mtx, "Driver"):
+        raise ImportError("madtrex is not installed (no madtrex.Driver in %s)"
+                          % list(getattr(mtx, "__path__", [])))
+except ImportError as _err:
+    mtx = _MissingMadtrex(str(_err))
+
+
+def _require_madtrex() -> None:
+    """Raises the ImportError of a missing madtrex, before any work."""
+    if isinstance(mtx, _MissingMadtrex):
+        mtx.Driver  # raises
+
+if __package__:
+    from . import param_reweight_card as param_rwgt_card
+    from . import pdf_reweight_card as pdf_rwgt_card
+    from . import generic_reweight_card as generic_rwgt_card
+else:
+    if str(_HERE) not in sys.path:
+        sys.path.append(str(_HERE))
+    import param_reweight_card as param_rwgt_card
+    import pdf_reweight_card as pdf_rwgt_card
+    import generic_reweight_card as generic_rwgt_card
+
+# Regenerates a process in another model; see run_parameter_reweighting()'s
+# regenerate argument and mg7/madtrex.py's regenerate_process_in_model().
+Regenerator = Callable[..., str]
+
+logger = logging.getLogger("madgraph7")
+
+
+def find_latest_run_dir(events_dir: str = "Events") -> str:
+    """Returns the most recently modified Events/<run_name>_NN run directory."""
+    run_dirs = glob.glob(os.path.join(events_dir, "*_[0-9][0-9]"))
+    if not run_dirs:
+        raise FileNotFoundError(f"no run directories found under {events_dir}")
+    return max(run_dirs, key=os.path.getmtime)
+
+
+def find_run_events(run_dir: str) -> str:
+    """Returns run_dir's generated events file, preferring LHEF XML over the
+    internal binary format (see launch.py's generate_events, which writes
+    events.lhe.gz for output_format="lhe" and events.npy for "lhe_npy")."""
+    for name in ("events.lhe", "events.lhe.gz", "events.npy"):
+        path = os.path.join(run_dir, name)
+        if os.path.isfile(path):
+            return path
+    raise FileNotFoundError(f"no events.lhe(.gz) or events.npy found in {run_dir}")
+
+
+def prepare_events(events_path: str) -> str:
+    """Returns a path MadtRex can read events_path from: the MadtRex readers
+    don't handle gzip, so a gzipped LHE file (launch.py gzips events.lhe by
+    default) is decompressed once into a sibling file without the .gz suffix,
+    reused on later calls. Any other path is returned unchanged."""
+    if not events_path.endswith(".gz"):
+        return events_path
+    plain_path = events_path[: -len(".gz")]
+    if not os.path.isfile(plain_path) or os.path.getmtime(plain_path) < os.path.getmtime(events_path):
+        logger.info("MadtRex: decompressing %s -> %s", events_path, plain_path)
+        with gzip.open(events_path, "rb") as src, open(plain_path, "wb") as dst:
+            shutil.copyfileobj(src, dst)
+    return plain_path
+
+
+def read_init_node(events_path: str) -> "ms.LHEMeta | None":
+    """Reads the <init> block launch.py's MadgraphProcess.write_lhe_header()
+    writes as header.lhe next to events_path (or, for runs made before that
+    existed, the init.lhe the old madevent.py wrote), if present, into an
+    LHEMeta that
+    Driver.load_events() can use to seed the reweighted sample's baseline
+    cross section and error. Needed for events.npy input (compact_npy and
+    lhe_npy), which -- unlike events.lhe -- carries no <init> block of its
+    own; without this, MadtRex reweighting has no baseline cross section to
+    normalise or propagate errors against (see reweightor::calc_xErrs() /
+    calc_norm() in Rex/src/teaRex.cc). Returns None if neither file is
+    found, in which case load_events() falls back to its own empty-LHEMeta
+    default."""
+    run_dir = os.path.dirname(events_path)
+    for name in ("header.lhe", "init.lhe"):
+        init_path = os.path.join(run_dir, name)
+        if os.path.isfile(init_path):
+            break
+    else:
+        return None
+    with open(init_path) as f:
+        match = re.search(r"<init>\s*\n(.*?)</init>", f.read(), re.S)
+    if not match:
+        return None
+    lines = [line.split() for line in match.group(1).strip().splitlines()]
+    header, proc_lines = lines[0], lines[1:]
+
+    def process(fields: list[str]) -> ms.LHEProcess:
+        proc = ms.LHEProcess(
+            cross_section=float(fields[0]), cross_section_error=float(fields[1]),
+            process_id=int(fields[3]),
+        )
+        proc.max_weight = float(fields[2])  # constructor kwarg is misspelled in the binding
+        return proc
+
+    return ms.LHEMeta(
+        beam1_pdg_id=int(header[0]), beam2_pdg_id=int(header[1]),
+        beam1_energy=float(header[2]), beam2_energy=float(header[3]),
+        beam1_pdf_authors=int(header[4]), beam2_pdf_authors=int(header[5]),
+        beam1_pdf_id=int(header[6]), beam2_pdf_id=int(header[7]),
+        weight_mode=int(header[8]),
+        processes=[process(fields) for fields in proc_lines],
+    )
+
+
+def _pdf_data_path() -> str | None:
+    """Resolves the LHAPDF data directory: $LHAPDF_DATA_PATH if set, else
+    lhapdf.paths()[0] (the configured lhapdf's own data dir, eg lhapdf6 in
+    HEPTools) if the lhapdf Python module is importable, else None -- same
+    convention as launch.py's own PDF_PATH."""
+    if "LHAPDF_DATA_PATH" in os.environ:
+        return os.environ["LHAPDF_DATA_PATH"]
+    try:
+        import lhapdf
+        lhapdf.setVerbosity(0)
+        return lhapdf.paths()[0]
+    except ImportError:
+        return None
+
+
+def resolve_pdf_set_files(pdf_set_name: str) -> tuple[str, str]:
+    """Resolves pdf_set_name (an LHAPDF set -- the only kind PdfDriver
+    currently supports) to its (grid_file, alpha_s_grid_file) paths, per the
+    same $LHAPDF_DATA_PATH/lhapdf.paths()[0] convention launch.py's own
+    init_beam()/ensure_pdf_set() use for the run_card's own [beam] pdf.
+    Unlike ensure_pdf_set(), this never attempts to download a missing set:
+    standalone reweighting has no run_card.toml/MADGRAPH_LHAPDF_CONFIG to
+    drive a download from, so it only resolves paths, raising a clear error
+    if pdf_set_name isn't already present on disk."""
+    data_path = _pdf_data_path()
+    if data_path is None:
+        raise RuntimeError(
+            f"cannot resolve PDF set {pdf_set_name!r}: no LHAPDF_DATA_PATH is set and the "
+            f"lhapdf Python module is not importable"
+        )
+    set_dir = os.path.join(data_path, pdf_set_name)
+    if not os.path.isdir(set_dir):
+        raise RuntimeError(
+            f"PDF set {pdf_set_name!r} not found under {data_path!r}; install it (eg via "
+            f"MadGraph's `install` command, or lhapdf's own `lhapdf install`) before "
+            f"reweighting against it"
+        )
+    return (
+        os.path.join(set_dir, f"{pdf_set_name}_0000.dat"),
+        os.path.join(set_dir, f"{pdf_set_name}.info"),
+    )
+
+
+def default_output_path(events_path: str) -> str:
+    """A path next to events_path for writing the reweighted events
+    elsewhere (the runners themselves write in place by default)."""
+    root, ext = os.path.splitext(events_path)
+    if ext == ".npy":
+        # write_weights() targets save_weights_binary() for binary input,
+        # which treats its path argument as a prefix, writing one
+        # "<prefix>_<weight_id>.npy" file per recorded weight -- so the
+        # default must not carry the .npy extension itself.
+        return f"{root}_reweighted"
+    return f"{root}_reweighted{ext}"
+
+
+# The [run]/[beam] run_card.toml defaults this module relies on, as declared
+# by RunCardMG7 in madgraph/various/banner.py -- read with tomllib here so
+# reweighting doesn't need MadGraph's card machinery. MadGraph always writes
+# every entry out, so these only matter for hand-written cards.
+_RUN_CARD_DEFAULTS = {
+    "run": {"device": ["cpu"], "cpu_mode": "auto", "precision": "color32"},
+    "beam": {
+        "leptonic": False,
+        "pdf1": "NNPDF40_lo_as_01180",
+        "pdf2": "NNPDF40_lo_as_01180",
+    },
+}
+
+
+def _run_card_path(process_directory: str) -> str:
+    """The card the matrix elements' [run] settings are read from: in a
+    gridpack, Cards/grid_run_card.toml, which (unlike the full run_card.toml
+    shipped next to it) names the cpu backend the shipped libraries were
+    actually built for rather than 'auto'; else Cards/run_card.toml."""
+    grid_card = os.path.join(process_directory, "Cards", "grid_run_card.toml")
+    if os.path.isfile(grid_card):
+        return grid_card
+    return os.path.join(process_directory, "Cards", "run_card.toml")
+
+
+def _read_run_card(path: str) -> dict:
+    """The [run] and [beam] sections of the run_card.toml at path, with
+    _RUN_CARD_DEFAULTS filled in for missing entries."""
+    with open(path, "rb") as f:
+        return _fill_run_card(tomllib.load(f))
+
+
+def _fill_run_card(card: dict) -> dict:
+    return {
+        section: {**defaults, **card.get(section, {})}
+        for section, defaults in _RUN_CARD_DEFAULTS.items()
+    }
+
+
+_RUN_CARD_BLOCK_RE = re.compile(r"<MG7RunCard>(.*?)</MG7RunCard>", re.S)
+
+
+def _read_events_header_block(events_path: str, block_re: re.Pattern) -> str | None:
+    """The content of the first block_re match in the LHE header of the
+    events of events_path -- header.lhe next to an events.npy, else the LHE
+    file's own header -- CDATA-unwrapped or unescaped, or None."""
+    if events_path.endswith(".npy"):
+        header_path = os.path.join(os.path.dirname(events_path), "header.lhe")
+    else:
+        header_path = events_path
+    if not os.path.isfile(header_path):
+        return None
+    opener = gzip.open if header_path.endswith(".gz") else open
+    lines = []
+    try:
+        with opener(header_path, "rt", errors="replace") as f:
+            for line in f:  # the header only: stop at <init> or the first event
+                if "<init" in line or "<event" in line:
+                    break
+                lines.append(line)
+    except OSError:
+        return None
+    match = block_re.search("".join(lines))
+    if not match:
+        return None
+    content = match.group(1).strip()
+    if content.startswith("<![CDATA[") and content.endswith("]]>"):
+        return content[len("<![CDATA["):-len("]]>")]
+    import html
+    return html.unescape(content)
+
+
+def _read_events_run_card(events_path: str) -> dict | None:
+    """The run_card.toml the events of events_path were generated with --
+    the <MG7RunCard> block launch.py (and a gridpack) write into the LHE
+    header (see _read_events_header_block()) -- parsed as _read_run_card()
+    does, or None when there is none. This, not the process directory's
+    current Cards/run_card.toml (which may have been edited since, e.g. for
+    another run), is what the sample was made with."""
+    content = _read_events_header_block(events_path, _RUN_CARD_BLOCK_RE)
+    if content is None:
+        return None
+    try:
+        return _fill_run_card(tomllib.loads(content))
+    except tomllib.TOMLDecodeError:
+        return None
+
+
+_SLHA_BLOCK_RE = re.compile(r"<slha>(.*?)</slha>", re.S)
+
+
+def _read_events_param_card(events_path: str) -> str | None:
+    """The param_card the events of events_path were generated with: the
+    <slha> block of their LHE header (see _read_events_header_block()), as
+    launch.py's write_lhe_header() and madevent write it -- with the widths
+    "auto" resolved to the values the generation used. None without one."""
+    content = _read_events_header_block(events_path, _SLHA_BLOCK_RE)
+    if content is None or not content.strip():
+        return None
+    return content.strip() + "\n"
+
+
+def _param_card_values(text: str) -> dict[tuple, float | None]:
+    """The numbers of an SLHA param_card, keyed (block, ids...): "DECAY pid
+    width" as ("decay", pid). Comments, case and formatting do not matter;
+    branching ratios are left out (the matrix elements do not depend on
+    them), a value that is no number ("auto") is None."""
+    values: dict[tuple, float | None] = {}
+    block = None
+    for raw in text.splitlines():
+        words = raw.split("#", 1)[0].split()
+        if not words:
+            continue
+        head = words[0].lower()
+        if head == "block":
+            block = words[1].lower() if len(words) > 1 else None
+            continue
+        if head == "decay":
+            block = None
+            try:
+                key = ("decay", int(words[1]))
+            except (IndexError, ValueError):
+                continue
+            try:
+                values[key] = float(words[2])
+            except (IndexError, ValueError):
+                values[key] = None
+            continue
+        if block is None:
+            continue
+        try:
+            ids = tuple(int(word) for word in words[:-1])
+        except ValueError:
+            continue
+        try:
+            values[(block,) + ids] = float(words[-1])
+        except ValueError:
+            values[(block,) + ids] = None
+    return values
+
+
+def _param_card_differences(generation_text: str, other_text: str,
+                            rel_tol: float = 1e-6) -> list[str]:
+    """The parameters other_text gives another value than generation_text
+    (numerically, to rel_tol; a value that is no number, eg an "auto"
+    width, is not compared), as "BLOCK IDS: generation -> other"."""
+    import math
+    generation, other = _param_card_values(generation_text), _param_card_values(other_text)
+    differences = []
+    for key in sorted(set(generation) | set(other), key=str):
+        a, b = generation.get(key, "missing"), other.get(key, "missing")
+        if a is None or b is None:
+            continue
+        if isinstance(a, float) and isinstance(b, float) and math.isclose(a, b, rel_tol=rel_tol, abs_tol=0.0):
+            continue
+        name = " ".join(str(part) for part in key)
+        differences.append(f"{name}: {a} -> {b}")
+    return differences
+
+
+def _generation_param_card(process_directory: str, events_path: str, param_card: str,
+                           card_dir: str) -> str:
+    """The param_card to normalise the events of events_path with: the one
+    they were generated with, from their LHE header, written into card_dir;
+    a warning lists the parameters in which param_card (default
+    process_directory/Cards/param_card.dat) differs from it. Without one in
+    the header, param_card, with a warning that the normalisation then
+    cannot be checked against the generation."""
+    current = param_card or os.path.join(process_directory, "Cards", "param_card.dat")
+    text = _read_events_param_card(events_path)
+    if text is None:
+        logger.warning(
+            "MadtRex: the events of %s carry no param_card (no <slha> block in their "
+            "header): they are normalised with %s, which cannot be checked against the "
+            "one they were generated with", events_path, current)
+        return current
+    path = os.path.join(card_dir, "generation_param_card.dat")
+    with open(path, "w") as f:
+        f.write(text)
+    if os.path.isfile(current):
+        with open(current) as f:
+            differences = _param_card_differences(text, f.read())
+        if differences:
+            shown = differences[:10] + ([f"... {len(differences) - 10} more"]
+                                        if len(differences) > 10 else [])
+            logger.warning(
+                "MadtRex: %s differs from the param_card the events of %s were generated "
+                "with (%s); the reweighting starts from the latter, as the events were "
+                "generated, not from %s", current, events_path, "; ".join(shown), current)
+    return path
+
+
+def _check_stored_matrix_elements(events_path: str, init_meta, process_directory: str,
+                                  param_card: str, devices: list[str], fptype: str,
+                                  sample: int = 32, rel_tol: float = 1e-6,
+                                  card_label: str = "") -> None:
+    """Re-evaluates the original matrix element (helicity-summed, with
+    param_card, at fptype) for a sample of the events of events_path whose
+    generation-time matrix elements are stored (STORED_MATRIX_ELEMENT_COLUMN)
+    and warns if they differ by more than rel_tol: the stored ones are then
+    not those of this process directory's libraries and card. A launch
+    changing nothing normalised by the stored ones gives the ratio of the
+    two as its weight over the event weight. rel_tol leaves room for an MLM
+    run's stored matrix elements, which are the first evaluation rescaled to
+    the event's alpha_s rather than evaluated again (about 1e-7 apart at the
+    mixed fptype)."""
+    import numpy as np
+    if not events_path.endswith(".npy"):
+        return
+    events = np.load(events_path, mmap_mode="r", max_header_size=1 << 30)
+    if STORED_MATRIX_ELEMENT_COLUMN not in (events.dtype.names or ()) or len(events) == 0:
+        return
+    picked = np.unique(np.linspace(0, len(events) - 1, min(sample, len(events))).astype(int))
+    with tempfile.TemporaryDirectory() as tmp:
+        sample_path = os.path.join(tmp, "events.npy")
+        np.save(sample_path, np.ascontiguousarray(events[picked]))
+        card = os.path.join(tmp, "check.rwgt")
+        with open(card, "w") as f:
+            f.write("launch --rwgt_name=check\n")
+        try:
+            driver = mtx.Driver()
+            driver.set_helicity_reweighting(False)
+            driver.set_device_priority(devices)
+            driver.set_fptype_priority([fptype])
+            driver.load_process_directory(process_directory, param_card)
+            driver.load_param_reweighting(card)
+            if init_meta is not None:
+                driver.load_events(sample_path, init_meta)
+            else:
+                driver.load_events(sample_path)
+            driver.warehouse().get().run()
+            driver.write_weights(os.path.join(tmp, "check"))
+            weights = np.load(os.path.join(tmp, "check_check.npy"))["weight"]
+        except Exception as error:  # a check only: the reweighting itself reports
+            logger.warning("MadtRex: could not check the stored matrix elements of %s "
+                           "against %s: %s", events_path, process_directory, error)
+            return
+    original = np.asarray(events["weight"][picked])
+    nonzero = original != 0
+    deviation = np.abs(weights[nonzero] / original[nonzero] - 1)
+    if deviation.size and deviation.max() > rel_tol:
+        logger.warning(
+            "MadtRex: the matrix elements stored with %s differ from the ones the "
+            "libraries of %s give with %s (up to %.3g relative, %d of %d sampled events "
+            "beyond %g): helicity-summed launches are normalised with the stored ones, "
+            "the others with re-evaluated ones", events_path, process_directory,
+            card_label or param_card, deviation.max(), int(np.sum(deviation > rel_tol)),
+            deviation.size, rel_tol)
+
+
+def _generation_run_card(process_directory: str, events_path: str | None,
+                         card_path: str) -> tuple[dict | None, str]:
+    """(run card, where it was read) for the sample of events_path: the run
+    card in the events' header (see _read_events_run_card()), else the card
+    at card_path in the process directory, else (None, '')."""
+    if events_path is not None:
+        card = _read_events_run_card(events_path)
+        if card is not None:
+            return card, "the header of " + events_path
+    if os.path.isfile(card_path):
+        try:
+            return _read_run_card(card_path), card_path
+        except tomllib.TOMLDecodeError:
+            pass
+    return None, ""
+
+
+def reweighting_threads(process_directory: str, events_path: str | None = None) -> int:
+    """The worker threads MadtRex reweights with unless the card says
+    otherwise ("change nb_core"): as many as the event generation uses, the
+    [run] cpu_thread_pool_size of the process directory's run card (else of
+    the run card in the header of events_path), all available cores when it
+    is -1 or 0 (as launch.py and madspace resolve it) or cannot be read."""
+    card, _ = _generation_run_card(process_directory, None, _run_card_path(process_directory))
+    if card is None and events_path is not None:
+        card = _read_events_run_card(events_path)
+    try:
+        threads = int(card["run"].get("cpu_thread_pool_size", -1)) if card else -1
+    except (TypeError, ValueError):
+        threads = -1
+    if threads > 0:
+        return threads
+    try:
+        return len(os.sched_getaffinity(0))
+    except AttributeError:
+        return os.cpu_count() or 1
+
+
+def resolve_auto_backend(build_path: str) -> str:
+    """Ask the matrix-element Makefile to resolve the ``auto`` cpu_mode (as
+    resolve_auto_backend() in launch.py, which this module can't import
+    without MadGraph)."""
+    try:
+        result = subprocess.run(
+            ["make", "-n", "BACKEND=auto", "detect-backend"],
+            cwd=build_path,
+            capture_output=True,
+            text=True,
+            check=True,
+        )
+    except (OSError, subprocess.CalledProcessError) as exc:
+        raise RuntimeError(
+            f"Could not resolve cpu_mode='auto' in '{build_path}': {exc}"
+        ) from exc
+    match = re.search(r"^BACKEND=(\S+) \(was auto\)$", result.stdout, re.MULTILINE)
+    if match is None or match.group(1) == "auto":
+        raise RuntimeError(
+            f"Could not resolve cpu_mode='auto' in '{build_path}': "
+            f"make failed to report a backend.\n{result.stdout}{result.stderr}"
+        )
+    return match.group(1)
+
+
+def _make(args: list[str], cwd: str) -> None:
+    """Runs make with args in cwd, raising with its output on failure."""
+    result = subprocess.run(
+        ["make", *args], cwd=cwd, stdout=subprocess.PIPE, stderr=subprocess.STDOUT
+    )
+    if result.returncode != 0:
+        output = result.stdout.decode("utf-8", errors="replace")
+        raise RuntimeError(
+            f"'make {' '.join(args)}' failed in {cwd}:\n{output}"
+        )
+
+
+def _read_devices(process_directory: str) -> list[str]:
+    """The matrix-element build backends (the BACKEND= / {device} values) of
+    process_directory's run_card.toml: [run] device mapped through cpu_mode
+    as launch.py's backend_of() does, with cpu_mode 'auto' resolved to the
+    concrete SIMD backend so it matches the library names on disk."""
+    run_card = _read_run_card(_run_card_path(process_directory))
+    device_names = run_card["run"]["device"]
+    if not isinstance(device_names, list):
+        device_names = [device_names]
+    cpu_mode = run_card["run"]["cpu_mode"]
+    backends = []
+    for name in device_names:
+        device_type = name.split(":", 1)[0]
+        backends.append(cpu_mode if device_type == "cpu" else device_type)
+    if "auto" in backends:
+        with open(os.path.join(process_directory, "SubProcesses", "subprocesses.json")) as f:
+            first_proc_path = json.load(f)[0]["path"]
+        if not os.path.isabs(first_proc_path):
+            first_proc_path = os.path.join(process_directory, first_proc_path)
+        auto_backend = resolve_auto_backend(first_proc_path)
+        backends = [auto_backend if b == "auto" else b for b in backends]
+    return list(dict.fromkeys(backends))
+
+
+def _read_default_pdf_set(process_directory: str, events_path: str | None = None
+                          ) -> tuple[str | None, str]:
+    """(PDF set, where it was read): [beam] pdf1 of the run card the events
+    of events_path were generated with -- the <MG7RunCard> of their LHE
+    header (see _read_events_run_card()), else process_directory/Cards/
+    run_card.toml. This is run_pdf_reweighting()'s baseline PDF set when the
+    PDF reweight card names none: the PDF set the sample's own weights
+    correspond to. The header comes first because the current run_card may
+    have been edited since the sample was made (another run with another
+    PDF), which would scale every weight by the wrong ratio.
+
+    PdfDriver needs the same set on both beams: a run with pdf1 != pdf2, a
+    leptonic run, or no run card at all gives (None, ...), and the card then
+    has to name its baseline PDF_SET explicitly."""
+    card, source = _generation_run_card(
+        process_directory, events_path,
+        os.path.join(process_directory, "Cards", "run_card.toml"))
+    if card is None:
+        return None, ""
+    try:
+        beam = card["beam"]
+        if beam["leptonic"] or beam["pdf1"] != beam["pdf2"]:
+            return None, source
+        pdf_set = beam["pdf1"]
+    except (KeyError, TypeError):
+        return None, source
+    return (pdf_set or None), source
+
+
+# [run] precision -> FPTYPE, as launch.py's fptype_of()
+_PRECISION_2_FPTYPE = {"all64": "d", "all32": "f", "color32": "m", "denom64": "v"}
+
+
+def _read_run_fptype(process_directory: str, events_path: str | None = None) -> str | None:
+    """The FPTYPE the sample of events_path was generated (and its matrix
+    elements compiled) at: [run] precision of the run card in the events'
+    LHE header (see _read_events_run_card()), else of the process
+    directory's run card (Cards/grid_run_card.toml in a gridpack), mapped as
+    launch.py's fptype_of() does. Used by run_parameter_reweighting() as
+
+    - the default *target* fptype (param_rwgt_card.group_by_fptype()'s
+      default_fptype) when a reweight card sets no "change fptype", so a
+      plain reweight stays at the sample's own precision;
+    - the *normaliser* fptype (Driver.set_normaliser_fptype_priority()),
+      always: a launch is normalised against the sample's own configuration,
+      never against the fptype it is reweighted to.
+
+    None when there is no run card to read; callers then fall back to
+    param_rwgt_card.DEFAULT_FPTYPE."""
+    card, _ = _generation_run_card(process_directory, events_path,
+                                   _run_card_path(process_directory))
+    if card is None:
+        return None
+    try:
+        precision = card["run"]["precision"]
+    except (KeyError, TypeError):
+        return None
+    if not precision:
+        return None
+    precision = str(precision).lower()
+    return _PRECISION_2_FPTYPE.get(precision, precision)
+
+
+def _read_libtag(directory: str) -> str:
+    """The madmatrix LIBTAG directory's libraries are built with: "" but for
+    a process MadtRex regenerated in another model (mg7/madtrex.py's
+    regenerate_process_in_model() records it in its marker file)."""
+    marker = os.path.join(directory, ".madtrex_source.json")
+    if not os.path.isfile(marker):
+        return ""
+    with open(marker) as f:
+        return json.load(f).get("libtag", "")
+
+
+def _check_distinct_libraries(directory: str, original_directory: str) -> None:
+    """Raises unless no madmatrix library of directory (a process
+    regenerated in another model) has the name of one of
+    original_directory's: the dynamic loader identifies a library by its
+    name, so one of the same name, loaded next to the original's for the
+    normalisation, would silently be the original's (its parameters and
+    all). See LIBTAG in madmatrix.mk."""
+    def names(d):
+        lib = os.path.join(d, "lib")
+        return {n for n in (os.listdir(lib) if os.path.isdir(lib) else [])
+                if n.startswith("libmadmatrix") and n.endswith(".so")}
+    shared = names(directory) & names(original_directory)
+    if shared:
+        raise RuntimeError(
+            f"the matrix element libraries of {directory} have the names of the original "
+            f"process's ({', '.join(sorted(shared))}), so they cannot be loaded next to them "
+            f"for the normalisation: remove {directory} to have it regenerated"
+        )
+
+
+def _compile_subprocesses(directory: str, devices: list[str], fptype: str) -> None:
+    """Compiles whichever subprocesses under
+    directory/SubProcesses/subprocesses.json aren't already compiled for
+    fptype, for each of devices (BACKEND=<device> FPTYPE=<fptype>) --
+    mirrors the existence-checked per-device compile in launch.py's
+    MadgraphProcess.compile_matrix_elements, generalized to whichever fptype a reweight
+    card's "set FPTYPE ..." (see reweight_card.py's group_by_fptype()) or a
+    caller actually asked for. Safe (a fast no-op) to call for a
+    device/fptype combination that's already built, so callers needn't
+    track that themselves."""
+    with open(os.path.join(directory, "SubProcesses", "subprocesses.json")) as f:
+        subprocesses = json.load(f)
+    for entry in subprocesses:
+        for device in devices:
+            expected_path = os.path.join(directory, entry["me_path"].format(device=device, fptype=fptype))
+            if os.path.isfile(expected_path):
+                continue
+            source_path = os.path.join(directory, entry["path"])
+            if not os.path.isfile(os.path.join(source_path, "makefile")) and not os.path.isfile(
+                os.path.join(source_path, "Makefile")
+            ):
+                raise RuntimeError(
+                    f"{expected_path} is not available and there is no source to compile "
+                    f"it from in {source_path} (a gridpack ships the sources only with "
+                    f"include_source = true): reweight at the precision the libraries "
+                    f"were built with, or remake the gridpack"
+                )
+            logger.info(
+                "MadtRex: compiling %s for device '%s', fptype '%s'",
+                entry["path"], device, fptype,
+            )
+            # madmatrix.mk's own USEBUILDDIR=1 build-directory tag
+            # (BACKEND+FPTYPE+HELINL+HRDCOD) refuses to build a *different*
+            # tag while another tag's now-stale intermediate .o files are
+            # still sitting in build.<device>/ from an earlier compile (eg
+            # the original generation's own FPTYPE, or a previous reweight
+            # card's) -- "Cannot build for tag=... as old builds exist for
+            # other tags ...: Please run 'make clean' first". Harmless to
+            # always do so before compiling here: this branch only runs
+            # when expected_path doesn't exist yet, ie a real compile is
+            # needed regardless, and the *already-built* .so for whatever
+            # other tag was there previously lives in lib/, untouched by
+            # 'clean' removing only build.<device>/'s own intermediates.
+            # 'clean' here (madmatrix.mk) also recurses into the shared
+            # directory/src/ common-library build (fixed to do so under
+            # USEBUILDDIR=1 too -- it used to only clean src/ that way for
+            # the non-USEBUILDDIR path), so this one call covers both.
+            variables = [f"BACKEND={device}", f"FPTYPE={fptype}", "USEBUILDDIR=1"]
+            libtag = _read_libtag(directory)
+            if libtag:
+                variables.append(f"LIBTAG={libtag}")
+            _make(["clean"] + variables, cwd=source_path)
+            _make(variables, cwd=source_path)
+
+
+def _sanitize_dirname(name: str) -> str:
+    return re.sub(r"[^A-Za-z0-9_.-]+", "_", name.strip()).strip("_") or "model"
+
+
+def _replace_atomically(new_path: str, target: str) -> None:
+    """Moves new_path onto target, gzipping it first when target is a .gz
+    file; new_path is next to target, so the final rename is atomic."""
+    if target.endswith(".gz"):
+        tmp = target + ".tmp"
+        with open(new_path, "rb") as src, gzip.open(tmp, "wb") as dst:
+            shutil.copyfileobj(src, dst)
+        os.remove(new_path)
+        os.replace(tmp, target)
+    else:
+        os.replace(new_path, target)
+
+
+def _npy_weight_column(weight_id: str) -> str:
+    return f"rwgt_{weight_id}"
+
+
+def _merge_npy_weights(events_path: str, prefix: str, weight_ids: list[str],
+                       weight_contexts: list[str], drop_columns=()) -> None:
+    """Adds the weight files "<prefix>_<id>.npy" of the new weight_ids to
+    events_path as "rwgt_<id>" columns, after its existing rwgt_ columns
+    (the systematics' rwgt_<n>, a previous reweighting's), and lists them
+    in a weight group of the header.lhe next to it, which npy_to_lhe and the
+    MadtRex readers take the weight names from. The files are removed. The
+    drop_columns of events_path are left out of the rewritten file."""
+    import numpy as np
+    from numpy.lib import format as npy_format
+
+    events = np.load(events_path, mmap_mode="r", max_header_size=1 << 30)
+    names = [name for name in events.dtype.names if name not in drop_columns]
+    columns = {}
+    for weight_id in weight_ids:
+        column = _npy_weight_column(weight_id)
+        if column in names:
+            raise ValueError(
+                f"{events_path} already has a weight {weight_id!r}: give the reweight "
+                f"card's launches other names (launch --rwgt_name=...)"
+            )
+        data = np.load(f"{prefix}_{weight_id}.npy")
+        columns[column] = data["weight"] if data.dtype.names else data
+        if len(columns[column]) != len(events):
+            raise ValueError(f"{prefix}_{weight_id}.npy does not have one weight per event")
+    # the new columns go where madspace puts weight columns: after the last
+    # rwgt_ column, else before the bias and the particles
+    position = max((i + 1 for i, name in enumerate(names) if name.startswith("rwgt_")),
+                   default=None)
+    if position is None:
+        position = next((i for i, name in enumerate(names)
+                         if name == "bias" or name.startswith("part")), len(names))
+    descr = [(name, events.dtype.fields[name][0]) for name in names]
+    descr[position:position] = [(column, "<f8") for column in columns]
+    tmp = events_path + ".tmp"
+    out = npy_format.open_memmap(tmp, mode="w+", dtype=np.dtype(descr), shape=events.shape)
+    chunk = 100000
+    for start in range(0, len(events), chunk):
+        stop = min(start + chunk, len(events))
+        for name in names:
+            out[name][start:stop] = events[name][start:stop]
+        for column, data in columns.items():
+            out[column][start:stop] = data[start:stop]
+    out.flush()
+    del out, events
+    os.replace(tmp, events_path)
+    for weight_id in weight_ids:
+        os.remove(f"{prefix}_{weight_id}.npy")
+
+    header = os.path.join(os.path.dirname(events_path), "header.lhe")
+    if os.path.isfile(header):
+        with open(header) as f:
+            text = f.read()
+        group = "<weightgroup name='tearex_reweighting' weight_name_strategy='includeIdInWeightName'>\n"
+        group += "".join(f"<weight id='{weight_id}'>{context}</weight>\n"
+                         for weight_id, context in zip(weight_ids, weight_contexts))
+        group += "</weightgroup>\n"
+        if "</initrwgt>" in text:
+            text = text.replace("</initrwgt>", group + "</initrwgt>", 1)
+        else:
+            text = text.replace("</header>", f"<initrwgt>\n{group}</initrwgt>\n</header>", 1)
+        with open(header + ".tmp", "w") as f:
+            f.write(text)
+        os.replace(header + ".tmp", header)
+
+
+# the generation-time matrix elements MadGraph stores in events.npy for
+# MadtRex's normalisation (launch.py's MadgraphProcess.store_matrix_elements)
+STORED_MATRIX_ELEMENT_COLUMN = "matrix_element"
+
+
+def write_weights_in_place(driver, result, source_path: str, events_path: str,
+                           drop_stored_matrix_elements: bool = False) -> str:
+    """Writes the weights driver (a Driver or PdfDriver, whose reweighted
+    sample is result) computed into the event file they were computed for,
+    and returns its path: source_path, the run's event file, as the other
+    tools (MadSpin, the shower, npy_to_lhe) and the classic reweighting
+    module use it. An LHE file is rewritten through a temporary file next
+    to it (gzipped again when it was) and replaced atomically, the
+    decompressed copy events_path read from removed; an npy file gets the
+    new weights as rwgt_<id> columns (see _merge_npy_weights()). Either
+    way the existing weights (systematics, earlier reweightings) are kept,
+    and the new ones follow them. drop_stored_matrix_elements removes the
+    stored matrix elements (STORED_MATRIX_ELEMENT_COLUMN) from an npy file in
+    the same rewrite: for the last reweighting pass over it."""
+    run_dir = os.path.dirname(os.path.abspath(source_path))
+    if source_path.endswith(".npy"):
+        prefix = os.path.join(run_dir, ".madtrex_weights")
+        existing = set(_npy_existing_weight_ids(source_path))
+        driver.write_weights(prefix)
+        ids = list(result.weight_ids)
+        contexts = list(getattr(result, "weight_context", [])) or [""] * len(ids)
+        new = [(i, c) for i, c in zip(ids, contexts + [""] * len(ids)) if i not in existing]
+        for weight_id in ids:
+            if weight_id in existing and os.path.isfile(f"{prefix}_{weight_id}.npy"):
+                os.remove(f"{prefix}_{weight_id}.npy")
+        _merge_npy_weights(source_path, prefix, [i for i, _ in new], [c for _, c in new],
+                           drop_columns=(STORED_MATRIX_ELEMENT_COLUMN,)
+                           if drop_stored_matrix_elements else ())
+        return source_path
+    tmp = os.path.join(run_dir, ".madtrex_reweighted.lhe")
+    driver.write_weights(tmp)
+    _replace_atomically(tmp, source_path)
+    if os.path.abspath(events_path) != os.path.abspath(source_path) and os.path.isfile(events_path):
+        os.remove(events_path)  # the decompressed copy, stale now
+    return source_path
+
+
+def _npy_existing_weight_ids(events_path: str) -> list[str]:
+    """The weight ids events_path already has columns for (rwgt_<id>)."""
+    import numpy as np
+    events = np.load(events_path, mmap_mode="r", max_header_size=1 << 30)
+    return [name[len("rwgt_"):] for name in events.dtype.names if name.startswith("rwgt_")]
+
+
+def _tag_output_path(output_path: str, tag: str) -> str:
+    root, ext = os.path.splitext(output_path)
+    return f"{root}_{tag}{ext}"
+
+
+def run_parameter_reweighting(
+    rwgt_path: str,
+    events_path: str | None = None,
+    output_path: str | None = None,
+    process_directory: str = ".",
+    param_card: str = "",
+    regenerate: Regenerator | None = None,
+    helicity_reweighting: bool | None = None,
+    drop_stored_matrix_elements: bool = False,
+) -> dict[tuple[str | None, str], mtx.Driver]:
+    """Runs MadtRex reweighting end to end against an existing MadGraph7 run.
+
+    Parses rwgt_path (see reweight_card.py) into groups of launches by which
+    UFO model *and* which floating-point precision (fptype) they target (a
+    bare card with no "change model"/"set FPTYPE" lines is one group
+    targeting the original process_directory as-is, at its baseline
+    precision -- see reweight_card.group_by_fptype()). Each group with at
+    least one launch runs as its own independent reweighting pass, since
+    each needs its own MatrixElementApi loaded (a different model needs
+    different physics code entirely; a different fptype needs a
+    differently-compiled library for the *same* code -- see
+    Driver.set_fptype_priority()): any group with a "change model" is
+    regenerated/compiled into its own subdirectory of process_directory
+    first by regenerate (mg7/madtrex.py's regenerate_process_in_model();
+    without one, such a card is an error) -- named "rw_me" if the card
+    only ever requests that one model, else "rw_me_<model>" per distinct
+    model (shared across that model's fptype groups, since regeneration
+    itself doesn't depend on fptype) -- and every group, regenerated
+    directory or not, has its own fptype compiled on demand (see
+    _compile_subprocesses()) before it's loaded. By default (output_path
+    None) every pass adds its weights to events_path itself, after the
+    weights already there (see write_weights_in_place()), so the groups'
+    weights accumulate in that one file. With output_path, each pass writes
+    its own output file next to output_path instead: untagged if it is the
+    card's only group, else suffixed "_<model>_<fptype>" (the original-model
+    group, if not alone, uses "original" for <model>). events_path is
+    auto-discovered from the latest run under process_directory/Events if
+    not given.
+
+    The card's preamble may also set any of reweight_card.DRIVER_OPTION_NAMES
+    (helicity reweighting, warning verbosity, event-batch splitting, deferred-
+    iteration flushing, streaming -- see reweight_card.DriverOptions), applied
+    to every Driver this builds, regardless of group. helicity_reweighting,
+    unless None, overrides the card's "set helicity_reweighting": True
+    reweights each event at its own helicity (Driver's default), False uses
+    helicity-summed matrix elements.
+
+    Returns a dict of the Driver built for each group, keyed by its
+    (model, fptype) (model is None for the original-model group), so
+    callers can inspect driver.warehouse().get() afterwards.
+
+    drop_stored_matrix_elements: the last group writing its weights into an
+    npy events_path in place also removes the stored matrix elements from
+    it (see write_weights_in_place()).
+
+    Every launch is normalised by the matrix element the events were
+    generated with: the original process directory's, with the param_card
+    of the events' own header (see _generation_param_card(); a warning
+    lists how process_directory's Cards/param_card.dat differs from it,
+    another one says when the header has none), at the sample's fptype.
+    The original model's launches start from that same card, unless
+    param_card names another one. When the events store their matrix
+    elements, a sample of them is first compared with re-evaluated ones
+    (see _check_stored_matrix_elements())."""
+    with tempfile.TemporaryDirectory() as card_dir:
+        return _run_parameter_reweighting(
+            rwgt_path, events_path, output_path, process_directory, param_card,
+            regenerate, helicity_reweighting, drop_stored_matrix_elements, card_dir)
+
+
+def _run_parameter_reweighting(
+    rwgt_path, events_path, output_path, process_directory, param_card,
+    regenerate, helicity_reweighting, drop_stored_matrix_elements, card_dir,
+):
+    """run_parameter_reweighting(), writing the generation's param_card
+    into card_dir."""
+    if events_path is None:
+        events_path = find_run_events(
+            find_latest_run_dir(os.path.join(process_directory, "Events"))
+        )
+    source_path = events_path
+    events_path = prepare_events(source_path)
+    in_place = output_path is None
+    init_meta = read_init_node(events_path)
+
+    with open(rwgt_path) as f:
+        parsed = param_rwgt_card.expand_scans(param_rwgt_card.parse_reweight_card(f.read()))
+    # Card-wide, not per-group -- see DriverOptions -- so resolved once here
+    # from the still-ungrouped ParsedCard and applied to every Driver below.
+    driver_options = parsed.driver_options
+    # The sample's own, actual generation-time fptype -- see
+    # _read_run_fptype()'s own doc comment. Used both as group_by_fptype()'s
+    # default *target* fptype (a card that sets no FPTYPE at all should stay
+    # at the sample's own precision, not drift to a hardcoded default) and,
+    # unconditionally below, as every group.model_spec-is-None group's
+    # *normaliser* fptype -- a launch must always be normalised against this
+    # original configuration, never against whichever fptype it targets.
+    original_fptype = (_read_run_fptype(process_directory, source_path)
+                       or param_rwgt_card.DEFAULT_FPTYPE)
+    # The param_card the events were generated with (from their header), not
+    # whatever Cards/param_card.dat is now: the normaliser of every group,
+    # and where the original model's launches start from unless param_card
+    # names another one.
+    generation_card = _generation_param_card(process_directory, source_path, param_card,
+                                             card_dir)
+    groups = [
+        group for group in param_rwgt_card.group_by_fptype(parsed, default_fptype=original_fptype)
+        if group.launches
+    ]
+    if not groups:
+        raise ValueError(f"reweight card {rwgt_path!r} has no 'launch' blocks")
+    multi = len(groups) > 1
+
+    devices = _read_devices(process_directory)
+    threads = reweighting_threads(process_directory, source_path)
+    _compile_subprocesses(process_directory, devices, original_fptype)
+    _check_stored_matrix_elements(
+        events_path, init_meta, process_directory, generation_card, devices, original_fptype,
+        card_label=("the param_card they were generated with"
+                    if generation_card.startswith(card_dir) else ""))
+    drivers: dict[tuple[str | None, str], mtx.Driver] = {}
+    for group_index, group in enumerate(groups):
+        model_tag = _sanitize_dirname(group.model_spec) if group.model_spec is not None else "original"
+        tag = f"{model_tag}_{group.fptype}" if multi else None
+
+        if group.model_spec is None:
+            load_directory, group_param_card = process_directory, param_card or generation_card
+        else:
+            if regenerate is None:
+                raise RuntimeError(
+                    f"reweight card {rwgt_path!r} asks for 'change model {group.model_spec}', "
+                    f"which needs a full MadGraph installation to regenerate the process in "
+                    f"that model: run it through MadGraph (mg7/madtrex.py) rather than this "
+                    f"standalone driver, or reweight within the process's own model"
+                )
+            rw_subdir = "rw_me" if not multi else f"rw_me_{model_tag}"
+            load_directory = regenerate(
+                process_directory, group.model_spec, group.process_definition, rw_subdir,
+                fptype=group.fptype,
+            )
+            group_param_card = ""
+
+        # Guarantees group.fptype is actually compiled for load_directory,
+        # regardless of whether regenerate() just did so
+        # (only true on the first group to regenerate a given rw_subdir --
+        # see its own doc comment) or load_directory is process_directory
+        # itself (never compiled for anything but launch.py's own
+        # hardcoded default fptype at generation time).
+        _compile_subprocesses(load_directory, devices, group.fptype)
+        # The normaliser needs original_fptype's own compiled library too,
+        # whenever it differs from the target fptype just compiled above
+        # (see the driver.set_normaliser_fptype_priority() call below). A
+        # "change model" group is normalised against the original process
+        # directory's own matrix elements, at original_fptype, like every
+        # other group (see Driver.set_normaliser_process_directory()).
+        if group.model_spec is not None:
+            _compile_subprocesses(process_directory, devices, original_fptype)
+        elif original_fptype != group.fptype:
+            _compile_subprocesses(load_directory, devices, original_fptype)
+
+        if in_place:
+            # each group adds its weights to the file the previous one wrote
+            events_path = prepare_events(source_path)
+        else:
+            group_output_path = output_path if tag is None else _tag_output_path(output_path, tag)
+
+        launches = group.launches
+        if param_rwgt_card.has_parameter_names(launches):
+            # "set MT 175": named by the param_card this group runs with
+            card_path = group_param_card or os.path.join(load_directory, "Cards", "param_card.dat")
+            with open(card_path) as f:
+                launches = param_rwgt_card.resolve_parameter_names(launches, f.read())
+
+        with tempfile.NamedTemporaryFile("w", suffix=".rwgt", delete=False) as fh:
+            fh.write(param_rwgt_card.render_launches(launches))
+            resolved_rwgt_path = fh.name
+        try:
+            logger.info(
+                "MadtRex: reweighting %s against model '%s' at fptype '%s' with %s",
+                events_path, group.model_spec or "<original>", group.fptype, rwgt_path,
+            )
+            driver = mtx.Driver()
+            # threads and batch splitting by default as the generation, then
+            # whatever the card sets
+            driver.set_threads(threads)
+            driver.set_min_batch(param_rwgt_card.DEFAULT_MIN_BATCH)
+            driver_options.apply(driver)
+            logger.info(
+                "MadtRex: %d thread(s), subprocesses split into batches of at least "
+                "%d events", driver.threads(), driver.min_batch())
+            if helicity_reweighting is not None:
+                driver.set_helicity_reweighting(helicity_reweighting)
+            driver.set_device_priority(devices)
+            driver.set_fptype_priority([group.fptype])
+            # The normaliser is always the matrix element the events were
+            # generated with: the original process directory's, with its
+            # param_card, at the sample's own generation-time fptype
+            # (original_fptype), never group.fptype -- see
+            # Driver.set_normaliser_fptype_priority()'s own doc comment. A
+            # no-op for the original model (normaliser_priority ==
+            # fptype_priority, one api shared for both roles) whenever the
+            # card doesn't actually change fptype, since group.fptype then
+            # also resolves to original_fptype (see group_by_fptype()'s
+            # default_fptype). A "change model" group's weights are the
+            # other model's matrix element over that original one
+            # (MadGraph's reweighting convention), never over the other
+            # model's own.
+            driver.set_normaliser_fptype_priority([original_fptype])
+            if group.model_spec is not None:
+                _check_distinct_libraries(load_directory, process_directory)
+                driver.set_normaliser_process_directory(process_directory, generation_card)
+            else:
+                driver.set_normaliser_param_card(generation_card)
+            driver.load_process_directory(load_directory, group_param_card)
+            driver.load_param_reweighting(resolved_rwgt_path)
+            if init_meta is not None:
+                driver.load_events(events_path, init_meta)
+            else:
+                driver.load_events(events_path)
+            driver.warehouse().get().run()
+            if in_place:
+                group_output_path = write_weights_in_place(
+                    driver, driver.warehouse().get(), source_path, events_path,
+                    drop_stored_matrix_elements=(drop_stored_matrix_elements
+                                                 and group_index == len(groups) - 1))
+            else:
+                driver.write_weights(group_output_path)
+        finally:
+            os.remove(resolved_rwgt_path)
+        logger.info("MadtRex: wrote %s", group_output_path)
+        drivers[(group.model_spec, group.fptype)] = driver
+
+    return drivers
+
+
+def run_pdf_reweighting(
+    rwgt_path: str,
+    events_path: str | None = None,
+    output_path: str | None = None,
+    process_directory: str = ".",
+    drop_stored_matrix_elements: bool = False,
+) -> mtx.PdfDriver:
+    """Runs MadtRex PDF-only reweighting end to end against an existing
+    MadGraph7 run, via PdfDriver -- entirely independent of run_parameter_reweighting()
+    (matrix-element/parameter reweighting): see pdf_reweight_card.py's module
+    docstring for why the two are split into separate cards/flows (PDFs and
+    matrix elements factorise), and main() for running both together against
+    the same sample.
+
+    Unlike run_parameter_reweighting(), no model regeneration or subprocess compilation
+    is needed: a PDF reweighting hypothesis is just a different (PDF set,
+    scale) pair fed to the same PdfSet/PdfDriver machinery, not a different
+    physics library, so this is a single pass regardless of how many "launch"
+    blocks or distinct PDF sets rwgt_path uses.
+
+    Parses rwgt_path (see pdf_reweight_card.py), first expanding any
+    scan:[...] MUF/MUR/MUF_FACTOR/MUR_FACTOR values into explicit per-point
+    launches (pdf_reweight_card.expand_scans(), the same mechanism
+    run_parameter_reweighting() uses for physics parameter scans), then resolving the
+    result into a ResolvedPdfCard: a required baseline PDF set -- unlike
+    run_parameter_reweighting()'s FPTYPE, which falls back to a hard-set default when
+    unspecified, MadtRex does not
+    currently record which PDF set an existing sample was generated with
+    anywhere *on the sample itself*. If the card's preamble never sets
+    PDF_SET/BEAM1_PDF_SET/BEAM2_PDF_SET at all, this looks it up instead,
+    from process_directory's own run_card.toml [beam] pdf (see
+    _read_default_pdf_set()) -- a real fact about the run, not a guessed
+    default, available whenever process_directory still has (or ever had)
+    the Cards/ that generated events_path, which covers both "reweighting
+    immediately after generation" and any later run against that same
+    process directory. Only once that lookup also comes up empty (no
+    run_card.toml, or one predating [beam] pdf) does pdf_rwgt_card.resolve()
+    raise, requiring PDF_SET explicitly -- eg for a bare/relocated sample,
+    or to deliberately reweight against a baseline other than the one the
+    run actually used. A further hook could still infer the baseline PDF
+    set straight from the LHE file's own metadata instead of the run_card
+    -- not implemented, since MadGraph7's LHE output carries no such field
+    today. Every distinct PDF set name the card references (baseline and
+    launches alike) is load_pdf_set()'d once, resolved to its LHAPDF
+    grid/.info files via resolve_pdf_set_files().
+
+    Beam energies are taken from events_path's sibling header.lhe/init.lhe
+    (see read_init_node()) if present, else from its LHE <init> block directly
+    (PdfDriver.load_events() sniffs this itself). events_path is auto-
+    discovered from the latest run under process_directory/Events if not
+    given. The weights are added to events_path itself unless output_path
+    is given (see write_weights_in_place()).
+
+    Returns the built PdfDriver, so callers can inspect driver.get()
+    afterwards."""
+    if events_path is None:
+        events_path = find_run_events(
+            find_latest_run_dir(os.path.join(process_directory, "Events"))
+        )
+    source_path = events_path
+    events_path = prepare_events(source_path)
+    init_meta = read_init_node(events_path)
+
+    with open(rwgt_path) as f:
+        parsed = pdf_rwgt_card.expand_scans(pdf_rwgt_card.parse_pdf_reweight_card(f.read()))
+    baseline = parsed.baseline
+    if baseline.pdf_set_name is None and baseline.beam1_pdf_set_name is None and baseline.beam2_pdf_set_name is None:
+        default_pdf_set, source = _read_default_pdf_set(process_directory, source_path)
+        if default_pdf_set is not None:
+            logger.info(
+                "MadtRex: PDF reweight card %s sets no baseline PDF_SET; using '%s', "
+                "the PDF set the events were generated with (from %s)",
+                rwgt_path, default_pdf_set, source,
+            )
+            baseline.pdf_set_name = default_pdf_set
+    card = pdf_rwgt_card.resolve(parsed)
+    if not card.launches:
+        raise ValueError(f"PDF reweight card {rwgt_path!r} has no 'launch' blocks")
+
+    driver = mtx.PdfDriver()
+    if card.verbose is not None:
+        driver.set_verbose(card.verbose)
+    if card.alpha_s_power is not None:
+        driver.set_alpha_s_power(card.alpha_s_power)
+    if card.scale_variation_output is not None:
+        driver.set_scale_variation_output(
+            getattr(mtx.ScaleVariationOutput, card.scale_variation_output.upper())
+        )
+
+    def as_scale_choice(scale: pdf_rwgt_card.PdfScale) -> mtx.ScaleChoice:
+        return mtx.ScaleChoice(scale.value, scale.absolute)
+
+    loaded_pdf_sets: set[str] = set()
+
+    def ensure_loaded(name: str) -> None:
+        if name in loaded_pdf_sets:
+            return
+        grid_file, alpha_s_grid_file = resolve_pdf_set_files(name)
+        driver.load_pdf_set(name, grid_file, alpha_s_grid_file)
+        loaded_pdf_sets.add(name)
+
+    for name in card.baseline_pdf_set_names:
+        ensure_loaded(name)
+    for launch in card.launches:
+        for name in launch.pdf_set_names:
+            ensure_loaded(name)
+
+    driver.set_baseline(
+        *card.baseline_pdf_set_names,
+        as_scale_choice(card.baseline_muF_scale),
+        as_scale_choice(card.baseline_muR_scale),
+    )
+    for i, launch in enumerate(card.launches):
+        driver.add_hypothesis(
+            launch.name or f"pdf_rwgt_{i + 1}",
+            *launch.pdf_set_names,
+            as_scale_choice(launch.muF_scale),
+            as_scale_choice(launch.muR_scale),
+        )
+
+    logger.info("MadtRex: PDF-reweighting %s with %s", events_path, rwgt_path)
+    if init_meta is not None:
+        # PdfDriver.load_events() only actually consults meta's beam
+        # energies for binary input -- for XML LHE it reads the file's own
+        # <init> block instead (older MadGraph7 writers left a zeroed-out
+        # stub there, not real beam energies). set_beam_energies() is the
+        # documented way to override that, so use it explicitly whenever
+        # init_meta (from header.lhe/init.lhe) is available, regardless of
+        # events_path's format.
+        driver.set_beam_energies(init_meta.beam1_energy, init_meta.beam2_energy)
+        driver.load_events(events_path, init_meta)
+    else:
+        driver.load_events(events_path)
+    driver.get().run()
+    if output_path is None:
+        output_path = write_weights_in_place(
+            driver, driver.get(), source_path, events_path,
+            drop_stored_matrix_elements=drop_stored_matrix_elements)
+    else:
+        driver.write_weights(output_path)
+    logger.info("MadtRex: wrote %s", output_path)
+    return driver
+
+
+def _merge_reweighting_results(target, source) -> None:
+    """Merges source's weight columns onto target, in place: appends every
+    one of source's weight_ids after target's own, and correspondingly
+    extends every target event's own wgts with the matching source event's.
+    target and source must come from reweighting the exact same events, in
+    the same order (true for run_generic_reweighting()'s own dispatch,
+    since both passes load the same events_path) -- raises if their event
+    counts differ, as a cheap sanity check of that assumption.
+
+    Uses add_weight_id() (in-place push_back through the *existing*
+    weight_ids shared_ptr) rather than reassigning target.weight_ids
+    directly -- the bound weight_ids property's setter replaces the whole
+    shared_ptr, which would desync every event's own (separately
+    shared_ptr-aliased) weight_ids from it and corrupt the write. See
+    madtrex::PdfDriver::keep_weights()'s own doc comment (pdfware.hpp) for
+    the same gotcha encountered the other way round (dropping columns
+    rather than adding them)."""
+    target_events = target.events
+    source_events = source.events
+    if len(target_events) != len(source_events):
+        raise ValueError(
+            "cannot coalesce reweighting results: event counts differ "
+            f"({len(target_events)} vs {len(source_events)}) -- were they reweighted "
+            f"against the same events file?"
+        )
+    for name in source.weight_ids:
+        target.add_weight_id(name)
+    for t_ev, s_ev in zip(target_events, source_events):
+        t_ev.wgts = t_ev.wgts + s_ev.wgts
+
+
+def run_generic_reweighting(
+    rwgt_path: str,
+    events_path: str | None = None,
+    output_path: str | None = None,
+    process_directory: str = ".",
+    param_card: str = "",
+    regenerate: Regenerator | None = None,
+    helicity_reweighting: bool | None = None,
+    drop_stored_matrix_elements: bool = False,
+) -> dict[str, object]:
+    """Runs MadtRex's generic reweight card end to end: splits rwgt_path
+    (see generic_reweight_card.py) into its "parameter"/"pdf" sections --
+    both, or whichever one is present -- runs each through its own real
+    routine (run_parameter_reweighting()/run_pdf_reweighting()) exactly as
+    if it had been handed a standalone card of that kind, then, if *both*
+    modes ran, coalesces the results back into a single sample per
+    parameter-reweighting group: every parameter-reweighting output
+    additionally gets the PDF-reweighting weights merged onto it (in place,
+    then rewritten to that same group's own output path), so a caller only
+    looking at that one "primary" output file sees every requested weight,
+    from both modes, on one set of events -- not two separate files to
+    reconcile themselves. If only one mode is present, nothing needs
+    coalescing -- that mode's own output, written by its own routine, is
+    already the whole answer.
+
+    events_path/output_path/process_directory/param_card are all forwarded
+    to whichever of run_parameter_reweighting()/run_pdf_reweighting()
+    actually runs; param_card is unused if the card has no "parameter"
+    section. Without output_path, each pass adds its weights to events_path
+    in turn and nothing needs coalescing. With output_path, it anchors the
+    *parameter* pass's own naming (its usual untagged-unless-multiple-groups
+    scheme, same as run_parameter_reweighting() alone); the PDF pass's own
+    intermediate output is separately tagged "_pdf" whenever a parameter
+    section is also present, mirroring main()'s own two-flag convention, so
+    the two never collide, and the PDF weights are merged onto the
+    parameter output(s).
+
+    Returns {"parameter": dict[(model, fptype), Driver] | None, "pdf":
+    PdfDriver | None} -- one real entry per mode that actually ran
+    (mirroring each routine's own native return value), None for a mode
+    the card didn't request. Callers after the *merged* result should read
+    the output file(s) written, not these driver objects -- which of them
+    ends up with its in-memory weight_ids/events mutated by the merge (the
+    parameter side's, when both modes ran) is an implementation detail, not
+    part of the return contract."""
+    with open(rwgt_path) as f:
+        sections = generic_rwgt_card.split_by_mode(f.read())
+    if not sections:
+        raise ValueError(f"generic reweight card {rwgt_path!r} has no 'set mode' sections")
+
+    if events_path is None:
+        events_path = find_run_events(
+            find_latest_run_dir(os.path.join(process_directory, "Events"))
+        )
+    # output_path None: each pass adds its weights to events_path in turn
+    in_place = output_path is None
+
+    result: dict[str, object] = {"parameter": None, "pdf": None}
+    with tempfile.TemporaryDirectory() as tmp_dir:
+        if "parameter" in sections:
+            param_path = os.path.join(tmp_dir, "parameter.rwgt")
+            with open(param_path, "w") as f:
+                f.write(sections["parameter"])
+            result["parameter"] = run_parameter_reweighting(
+                param_path,
+                events_path=events_path,
+                output_path=output_path,
+                process_directory=process_directory,
+                param_card=param_card,
+                regenerate=regenerate,
+                helicity_reweighting=helicity_reweighting,
+                drop_stored_matrix_elements=(drop_stored_matrix_elements
+                                             and "pdf" not in sections),
+            )
+        if "pdf" in sections:
+            pdf_path = os.path.join(tmp_dir, "pdf.rwgt")
+            with open(pdf_path, "w") as f:
+                f.write(sections["pdf"])
+            pdf_output_path = output_path
+            if not in_place and "parameter" in sections:
+                pdf_output_path = _tag_output_path(output_path, "pdf")
+            result["pdf"] = run_pdf_reweighting(
+                pdf_path,
+                events_path=events_path,
+                output_path=pdf_output_path,
+                process_directory=process_directory,
+                drop_stored_matrix_elements=drop_stored_matrix_elements,
+            )
+
+    param_drivers, pdf_driver = result["parameter"], result["pdf"]
+    if not in_place and param_drivers is not None and pdf_driver is not None:
+        multi = len(param_drivers) > 1
+        for (model_spec, fptype), driver in param_drivers.items():
+            model_tag = _sanitize_dirname(model_spec) if model_spec is not None else "original"
+            tag = f"{model_tag}_{fptype}" if multi else None
+            group_output_path = output_path if tag is None else _tag_output_path(output_path, tag)
+            _merge_reweighting_results(driver.warehouse().get(), pdf_driver.get())
+            driver.write_weights(group_output_path)
+            logger.info(
+                "MadtRex: coalesced PDF reweighting into parameter reweighting result for "
+                "model '%s' at fptype '%s' -> %s",
+                model_spec or "<original>", fptype, group_output_path,
+            )
+
+    return result
+
+
+def run_reweighting(
+    rwgt_path: str,
+    events_path: str | None = None,
+    card_kind: str | None = None,
+    output_path: str | None = None,
+    process_directory: str = ".",
+    param_card: str = "",
+    regenerate: Regenerator | None = None,
+    helicity_reweighting: bool | None = None,
+    drop_stored_matrix_elements: bool = False,
+):
+    """Runs the reweight card rwgt_path with the runner its card_kind needs:
+    "param" -> run_parameter_reweighting(), "pdf" -> run_pdf_reweighting(),
+    "generic" -> run_generic_reweighting(). card_kind defaults to
+    generic_reweight_card.card_kind() of the card. Returns that runner's
+    result."""
+    _require_madtrex()
+    if card_kind is None:
+        with open(rwgt_path) as f:
+            card_kind = generic_rwgt_card.card_kind(f.read())
+    if card_kind not in generic_rwgt_card.CARD_KINDS:
+        raise ValueError(
+            f"unknown reweight card kind {card_kind!r}; expected one of "
+            f"{list(generic_rwgt_card.CARD_KINDS)}"
+        )
+    common = dict(events_path=events_path, output_path=output_path,
+                  process_directory=process_directory,
+                  drop_stored_matrix_elements=drop_stored_matrix_elements)
+    if card_kind == "pdf":
+        return run_pdf_reweighting(rwgt_path, **common)
+    runner = run_parameter_reweighting if card_kind == "param" else run_generic_reweighting
+    return runner(rwgt_path, param_card=param_card, regenerate=regenerate,
+                  helicity_reweighting=helicity_reweighting, **common)
+
+
+def check_card_for_standalone(rwgt_path: str) -> None:
+    """Raises when the reweight card at rwgt_path needs the classic
+    reweighting module ("change backend classic", or a directive only the
+    classic module understands): that module is part of MadGraph, which
+    this standalone driver (a gridpack's bin/reweight) runs without.
+    "change backend madtrex|auto" is fine."""
+    problems = []
+    with open(rwgt_path) as f:
+        for number, raw_line in enumerate(f, 1):
+            line = generic_rwgt_card.strip_comment(raw_line)
+            words = line.split()
+            if len(words) >= 2 and words[0].lower() == "change" and words[1].lower() == "backend":
+                value = words[2].lower() if len(words) > 2 else ""
+                if value == "classic":
+                    problems.append(f"line {number}: {line.strip()!r}")
+                elif value not in generic_rwgt_card.BACKENDS:
+                    raise ValueError(
+                        f"{rwgt_path} line {number}: unknown backend {value!r}; expected "
+                        f"one of {sorted(generic_rwgt_card.BACKENDS)}"
+                    )
+                continue
+            reason = generic_rwgt_card.classic_only_reason(line)
+            if reason is not None:
+                problems.append(f"line {number}: {line.strip()!r} ({reason})")
+    if problems:
+        raise SystemExit(
+            f"{rwgt_path} needs the classic reweighting module, and classic reweighting "
+            f"needs MadGraph (this standalone MadtRex driver runs without it): "
+            + "; ".join(problems)
+        )
+
+
+def _default_process_directory() -> str:
+    """The process directory this file belongs to when it is shipped inside
+    one (a gridpack's bin/reweight: <gridpack>/SubProcesses/subprocesses.json
+    next to <gridpack>/bin/), else the current directory."""
+    root = _HERE.parent
+    if (root / "SubProcesses" / "subprocesses.json").is_file() and (root / "Cards").is_dir():
+        return str(root)
+    return "."
+
+
+def main(argv: list[str] | None = None, regenerate: Regenerator | None = None) -> None:
+    """Command line entry point. regenerate is passed on to
+    run_parameter_reweighting(); without it a "change model" card is an
+    error (see mg7/madtrex.py for the MadGraph-side entry point)."""
+    _require_madtrex()
+    parser = argparse.ArgumentParser(
+        description="Run MadtRex event reweighting against an existing MadGraph7 run. "
+                     "Parameter (matrix-element) and PDF reweighting are independent "
+                     "passes -- see run_parameter_reweighting()/run_pdf_reweighting(). Pass "
+                     "--rwgt-card alone for a generic card that dispatches ('change mode "
+                     "LO'/'change mode pdf') to one or both automatically and coalesces "
+                     "the result -- see run_generic_reweighting() -- or --param-rwgt-card/"
+                     "--pdf-rwgt-card directly (either, or both, for two separate outputs)."
+    )
+    parser.add_argument(
+        "--rwgt-card", dest="generic_rwgt_path", default=None,
+        help="generic reweight card dispatching to parameter and/or PDF reweighting (see "
+             "generic_reweight_card.py), eg Cards/reweight_card.dat; mutually exclusive "
+             "with --param-rwgt-card/--pdf-rwgt-card",
+    )
+    parser.add_argument(
+        "--param-rwgt-card", dest="param_rwgt_path", default=None,
+        help="parameter (SLHA/matrix-element) reweight card, eg "
+             "Cards/param_reweight_card.dat",
+    )
+    parser.add_argument(
+        "--pdf-rwgt-card", dest="pdf_rwgt_path", default=None,
+        help="PDF reweight card, eg Cards/pdf_reweight_card.dat",
+    )
+    parser.add_argument(
+        "--events", dest="events_path", default=None,
+        help="events file to reweight (events.lhe or events.npy); "
+             "defaults to the latest run under Events/",
+    )
+    parser.add_argument(
+        "--output", dest="output_path", default=None,
+        help="write the reweighted events to this path (or binary weight-file prefix for "
+             "events.npy input) instead of adding the weights to --events itself, the "
+             "default (--pdf-rwgt-card's own output, when both flags are given, is further "
+             "tagged '_pdf' so the two don't collide)",
+    )
+    parser.add_argument(
+        "--process-dir", dest="process_directory", default=None,
+        help="process directory (or gridpack) containing SubProcesses/ and Cards/ "
+             "(default: the gridpack this script is in, else .)",
+    )
+    parser.add_argument(
+        "--helicity-reweighting", dest="helicity_reweighting",
+        action=argparse.BooleanOptionalAction, default=None,
+        help="reweight each event at its own helicity (--helicity-reweighting, the "
+             "default) or with helicity-summed matrix elements "
+             "(--no-helicity-reweighting); overrides the reweight card's 'change "
+             "helicity'",
+    )
+    parser.add_argument(
+        "--param-card", dest="param_card", default="",
+        help="baseline param_card.dat (--param-rwgt-card/--rwgt-card only); defaults to "
+             "<process-dir>/Cards/param_card.dat",
+    )
+    args = parser.parse_args(argv)
+    if args.process_directory is None:
+        args.process_directory = _default_process_directory()
+    if not logger.handlers and not logging.getLogger().handlers:
+        # standalone run: nothing has set up MadGraph's logging
+        logging.basicConfig(level=logging.INFO, format="%(message)s")
+    if regenerate is None:
+        # without MadGraph (see mg7/madtrex.py for the MadGraph side)
+        for path in (args.generic_rwgt_path, args.param_rwgt_path, args.pdf_rwgt_path):
+            if path is not None:
+                check_card_for_standalone(path)
+    if args.generic_rwgt_path is not None:
+        if args.param_rwgt_path is not None or args.pdf_rwgt_path is not None:
+            parser.error("--rwgt-card is mutually exclusive with --param-rwgt-card/--pdf-rwgt-card")
+        run_generic_reweighting(
+            args.generic_rwgt_path,
+            events_path=args.events_path,
+            output_path=args.output_path,
+            process_directory=args.process_directory,
+            param_card=args.param_card,
+            regenerate=regenerate,
+            helicity_reweighting=args.helicity_reweighting,
+        )
+        return
+    if args.param_rwgt_path is None and args.pdf_rwgt_path is None:
+        parser.error("at least one of --rwgt-card/--param-rwgt-card/--pdf-rwgt-card is required")
+    if args.param_rwgt_path is not None:
+        run_parameter_reweighting(
+            args.param_rwgt_path,
+            events_path=args.events_path,
+            output_path=args.output_path,
+            process_directory=args.process_directory,
+            param_card=args.param_card,
+            regenerate=regenerate,
+            helicity_reweighting=args.helicity_reweighting,
+        )
+    if args.pdf_rwgt_path is not None:
+        # An explicit --output is untagged by default (run_pdf_reweighting()
+        # only tags "_pdf" onto its own auto-derived default); tag it
+        # ourselves here so an explicit --output can't collide with
+        # --param-rwgt-card's own write when both flags are given.
+        pdf_output_path = args.output_path
+        if pdf_output_path is not None and args.param_rwgt_path is not None:
+            pdf_output_path = _tag_output_path(pdf_output_path, "pdf")
+        run_pdf_reweighting(
+            args.pdf_rwgt_path,
+            events_path=args.events_path,
+            output_path=pdf_output_path,
+            process_directory=args.process_directory,
+        )
+
+
+if __name__ == "__main__":
+    main()
diff --git a/madgraph/iolibs/template_files/mg7/npy_to_lhe.py b/madgraph/iolibs/template_files/mg7/npy_to_lhe.py
index df9c62c2f5..7978ae0b07 100644
--- a/madgraph/iolibs/template_files/mg7/npy_to_lhe.py
+++ b/madgraph/iolibs/template_files/mg7/npy_to_lhe.py
@@ -19,7 +19,8 @@
 In both cases the LHE header and ``<init>`` block (with the systematics
 ``<initrwgt>``) are taken verbatim from the ``header.lhe`` file the run writes
 next to its npy events; the ``rwgt_<id>`` columns become the per-event
-``<rwgt>`` block. The ``<mgrwt>`` reweighting-input block (``[systematics]
+``<rwgt>`` block: the systematics' numbered ones, then the named ones MadtRex
+reweighting adds (``rwgt_<name>``, listed in header.lhe's ``<initrwgt>``). The ``<mgrwt>`` reweighting-input block (``[systematics]
 write_inputs``) is not reconstructed.
 
 Usage::
@@ -147,6 +148,19 @@ def _lhe_npy_events(ms, events, columns):
         yield event, [float(w[n]) for w in rwgt]
 
 
+def _add_named_weights(text, names, values):
+    """The formatted LHE event text with the named weights (ids madspace's
+    integer-id ``rwgt`` cannot carry) appended to its ``<rwgt>`` block, which
+    is created when the event has none."""
+    if not names:
+        return text
+    extra = "".join("<wgt id='%s'> %+.7e </wgt>\n" % (name, value)
+                    for name, value in zip(names, values))
+    if "</rwgt>" in text:
+        return text.replace("</rwgt>", extra + "</rwgt>", 1)
+    return text.replace("</event>", "<rwgt>\n" + extra + "</rwgt>\n</event>", 1)
+
+
 def convert(path, output=None, seed=None, compress=True, me_dir=None) -> str:
     """Write the LHE file for the npy events at ``path`` (npy file, run
     directory or run name) and return its path. By default it is written as
@@ -170,10 +184,13 @@ def convert(path, output=None, seed=None, compress=True, me_dir=None) -> str:
             "and cannot be converted." % os.path.join(run_path, HEADER_FILE))
     header = _read_header(run_path)
 
-    events = np.load(npy_path, mmap_mode="r")
+    events = np.load(npy_path, mmap_mode="r", max_header_size=1 << 30)
     names = events.dtype.names
     columns = [n for n in names if re.fullmatch(r"rwgt_\d+", n)]
     ids = [int(c.split("_", 1)[1]) for c in columns]
+    # MadtRex reweighting's named weights (rwgt_<name>)
+    named = [n for n in names if n.startswith("rwgt_") and n not in columns]
+    named_ids = [n[len("rwgt_"):] for n in named]
     if "diagram_index" in names:
         completer_path = os.path.join(run_path, COMPLETER_FILE)
         if not os.path.isfile(completer_path):
@@ -196,11 +213,13 @@ def convert(path, output=None, seed=None, compress=True, me_dir=None) -> str:
         for start in range(0, len(events), _CHUNK):
             chunk = np.asarray(events[start:start + _CHUNK])
             text = []
-            for event, weights in stream(chunk):
+            named_values = [chunk[n] for n in named]
+            for i, (event, weights) in enumerate(stream(chunk)):
                 if ids:
                     event.rwgt_ids = ids
                     event.rwgt = weights
-                text.append(event.format())
+                text.append(_add_named_weights(event.format(), named_ids,
+                                               [float(v[i]) for v in named_values]))
             out.write("".join(text))
         out.write("</LesHouchesEvents>\n")
     if compress:
diff --git a/madgraph/iolibs/template_files/mg7/param_reweight_card.py b/madgraph/iolibs/template_files/mg7/param_reweight_card.py
new file mode 100644
index 0000000000..d9a1316db3
--- /dev/null
+++ b/madgraph/iolibs/template_files/mg7/param_reweight_card.py
@@ -0,0 +1,642 @@
+"""Parsing/preprocessing for MadtRex parameter (matrix-element) reweight
+cards -- see pdf_reweight_card.py for the separate, independent PDF
+reweighting card, and generic_reweight_card.py for the dispatcher that lets
+one card mix both.
+
+madtrex.ParamHandler.read_rwgt_card() (madtrex/include/madtrex/
+param_handler.hpp) only understands the literal "# launch / # set BLOCK_NAME
+PARAM_ID PARAM_VALUE" grammar parsed by REX::tea::rwgt_slha::parse_rwgt_card
+(Rex/src/teaRex.cc) -- unrecognized lines (eg "change model ...") are silently
+ignored there, and scan:[...] value expressions fail to parse as a plain
+double. This module extracts the extra directives MadtRex needs to act on
+before handing a card to that low-level parser:
+
+- any number of "change model <spec>" lines, each paired with a "change
+  process <definition>" line, requesting the launches that follow be run
+  against a different UFO model, regenerated and compiled on demand (see
+  madtrex.py's regenerate_process_in_model()) -- MG7 does not persist the
+  original "generate <process>" command anywhere on disk, so it must be given
+  again explicitly whenever the model changes. Launches are grouped by their
+  (model, process) target regardless of how they're interleaved in the card,
+  so "change model" may appear more than once (switching back and forth, or
+  requesting several distinct alternate models) without duplicating work.
+  Such launches are still normalised against the *original* model's matrix
+  element (the one the events were generated with): their weights are the
+  new model's |M|^2 over the original |M|^2 (see Driver.
+  set_normaliser_process_directory() in madtrex.hpp);
+- "change fptype <value>" lines (formerly "set FPTYPE <value>", still
+  accepted with a deprecation warning), requesting the launch they belong to be
+  evaluated against a differently-precision-compiled matrix element library
+  (see MADMATRIX_LIB in madmatrix.mk / Driver.set_fptype_priority() in
+  madtrex.hpp) -- settable per launch, exactly like a physics "set BLOCK ID
+  VALUE" line (nothing carries over between launches). This is always the
+  *target* fptype to reweight TO -- normalisation is a separate, always-on
+  concern: every launch must be normalised against the sample's own actual
+  generation-time fptype, never against whichever fptype it's being
+  reweighted to (this is a fundamental assumption of how reweighting works
+  here -- see Driver.set_normaliser_fptype_priority() in madtrex.hpp/.cpp,
+  and madtrex.py's run_parameter_reweighting()). A "set FPTYPE ..."
+  appearing before the card's very first "launch" is special: it is taken as
+  the *default target* fptype, used by every launch that doesn't specify its
+  own "set FPTYPE" (see ParsedCard.baseline_fptype/group_by_fptype()). When
+  a card sets no FPTYPE at all (neither a baseline nor any per-launch
+  override), the default target -- like the always-on normaliser -- resolves
+  to the process directory's own run_card.toml [run] fptype (madtrex.py's
+  run_parameter_reweighting(), via group_by_fptype()'s default_fptype
+  parameter), falling back further to DEFAULT_FPTYPE only if that can't be
+  read either; this means a plain, non-FPTYPE-changing reweight (eg a bare
+  "set MT 200") correctly stays at whatever precision the sample was already
+  generated at, rather than silently drifting to a hardcoded default;
+- "change <option> <value>" lines for "helicity" and DRIVER_OPTION_NAMES,
+  configuring one of Driver's options (helicity reweighting, warning
+  verbosity, event-batch splitting, deferred-iteration flushing, streaming)
+  for every Driver madtrex.py builds from this card. "change helicity
+  <bool>" -- the classic reweighting module's option, with its default --
+  chooses between reweighting each event at its own helicity (True, the
+  default: the matrix elements are evaluated for the helicities in the
+  event's LHE SPINUP column) and helicity-summed matrix elements (False).
+  Helicity reweighting needs every event to carry helicity information
+  (SPINUP -1/0/+1 for every particle) and is an error otherwise; set it
+  False for a sample without it. The former spellings "set
+  helicity_reweighting"/"set use_helicity" and "set <option> <value>" are
+  still accepted, with a deprecation warning (see DEPRECATED_SET_NAMES).
+  Unlike
+  "change fptype", these are *not* per-launch (a Driver's own options are all
+  fixed before it ever loads/runs anything, uniformly for whatever launches
+  end up sharing that Driver -- see run_reweighting()), so they may only
+  appear before the card's very first "launch", exactly like "change
+  model"/the FPTYPE baseline; see ParsedCard.driver_options. Sizes/counts
+  (MAX_BATCH, MAX_DEFERRED_ITERATIONS, STREAM_BATCH_SIZE) are REX::npos
+  ("off"/"no cap") by default on the C++ side; any positive integer here
+  both sets the value *and* implies the option is wanted "on" (see
+  DriverOptions.streaming's own doc comment for STREAM_BATCH_SIZE's
+  particular case: it has no boolean of its own on the C++ side, only
+  STREAMING does, so setting just a batch size still needs to flip that on);
+- "set <NAME> <value>" lines naming a parameter by its param_card name (eg
+  "set MT 175", as the classic reweighting module allows), resolved to
+  "set BLOCK ID VALUE" against the param_card the launch runs with (see
+  resolve_parameter_names());
+- scan:[...] values in "set" lines, expanded into explicit per-value launch
+  blocks via the same eval-based mechanism and group-index semantics as
+  models.check_param_card.ParamCardIterator.iterate().
+
+The card shares its vocabulary with the classic reweighting module (see
+generic_reweight_card.py): "change backend madtrex|auto" and "change mode
+LO" are accepted, and the classic-only directives ("change output", "change
+mode NLO", density-matrix options, ...) are errors here.
+"""
+
+import itertools
+import re
+from dataclasses import dataclass, field
+
+try:
+    from . import generic_reweight_card as _vocab
+except ImportError:  # loaded as a plain module (gridpack bin/, tests)
+    import os as _os
+    import sys as _sys
+    _sys.path.append(_os.path.dirname(_os.path.abspath(__file__)))
+    import generic_reweight_card as _vocab
+
+# Hard-set for now (see the module docstring); should become a modifiable
+# global default once MadtRex/madevent.py grow one of their own.
+DEFAULT_FPTYPE = "m"
+
+# Recognized "change <NAME> <value>" preamble Driver options (see
+# DriverOptions) besides the shared "change helicity", lowercase, matched
+# case-insensitively against the card. MadtRex-only, as is "change fptype".
+# The threading ones (THREAD_OPTION_NAMES) are newer than the "set" spellings
+# and have none:
+# - "change nb_core <N|auto>": MadtRex's worker threads. auto (the default)
+#   uses as many as the event generation: the [run] cpu_thread_pool_size of
+#   the run card the sample was generated with, all available cores when it
+#   is -1 (see madtrex_driver.reweighting_threads());
+# - "change min_batch <N>": the fewest events per batch when the sample is
+#   split into batches across the threads (DEFAULT_MIN_BATCH by default; see
+#   Driver.set_min_batch()). A subprocess with more events than
+#   max(min_batch, events / threads) is split into batches of that size, so a
+#   sample dominated by one subprocess still keeps every thread busy, while a
+#   small one is not spread so thinly that a batch costs more to set up (its
+#   own matrix element instance, its own sorting) than it saves. An explicit
+#   "change max_batch" takes precedence.
+THREAD_OPTION_NAMES = frozenset({"nb_core", "min_batch"})
+DRIVER_OPTION_NAMES = frozenset({
+    "verbose", "max_batch", "max_deferred_iterations", "streaming", "stream_batch_size",
+}) | THREAD_OPTION_NAMES
+# "change" options only MadtRex understands.
+MADTREX_CHANGE_OPTIONS = DRIVER_OPTION_NAMES | {"fptype"}
+# Former "set <NAME> <value>" spellings -> the "change" option they now are.
+DEPRECATED_SET_NAMES = {
+    "helicity_reweighting": "helicity", "use_helicity": "helicity", "fptype": "fptype",
+    **{name: name for name in DRIVER_OPTION_NAMES - THREAD_OPTION_NAMES},
+}
+# The default of "change min_batch". Splitting a batch off costs a few
+# milliseconds (a matrix element instance reloaded from its library and
+# param_card, about 2.5 ms for p p > t t~ j j on a laptop, and the batch's
+# own setup), which a batch of 1000 events amortises even for a simple
+# process (2 -> 2 at its own helicity: about a microsecond per event and
+# launch); below that, threads gain little over running the subprocesses as
+# they are.
+DEFAULT_MIN_BATCH = 1000
+# PDF-reweighting names (pdf_reweight_card.py), errors in a parameter card.
+_PDF_NAMES = frozenset({
+    "pdf_set", "beam1_pdf_set", "beam2_pdf_set", "muf", "mur", "muf_factor",
+    "mur_factor", "alpha_s_power", "scale_variation_output",
+})
+
+_MODEL_RE = re.compile(r"^\s*change\s+model\s+(?P<spec>.+?)\s*$", re.I)
+_PROCESS_RE = re.compile(r"^\s*change\s+process\s+(?P<definition>.+?)\s*$", re.I)
+_LAUNCH_RE = re.compile(r"^\s*launch\b(?P<rest>.*)$", re.I)
+_NAME_RE = re.compile(r"rwgt_name\s*=\s*(?P<name>\S+)", re.I)
+# "set [param_card] <block> <id> <value>", <id> one integer or several
+# (multi-index blocks: "set NMIX 1 2 0.5"); and "set [param_card] <name>
+# <value>" for a parameter named by its param_card name. <value> may hold
+# spaces (scan:[1, 2]).
+_SET_RE = re.compile(
+    r"^\s*set\s+(?:param_card\s+)?(?P<block>\S+)\s+(?P<id>[+-]?\d+(?:\s+[+-]?\d+)*)"
+    r"\s+(?P<value>.+?)\s*$",
+    re.I,
+)
+_NAMED_SET_RE = re.compile(r"^\s*set\s+(?:param_card\s+)?(?P<name>\S+)\s+(?P<value>.+?)\s*$", re.I)
+_SCAN_RE = re.compile(r"scan\s*(?P<group>\d*)\s*:\s*(?P<values>.*)$", re.I)
+# as the classic module (banner.ConfigFile.format_variable), plus yes/on/no/off
+_TRUE_VALUES = frozenset({"1", "true", "yes", "on", "t", ".true."})
+_FALSE_VALUES = frozenset({"0", "false", "no", "off", "f", ".false."})
+
+
+def _parse_bool(name: str, value: str) -> bool:
+    v = value.strip().lower()
+    if v in _TRUE_VALUES:
+        return True
+    if v in _FALSE_VALUES:
+        return False
+    raise ValueError(f"cannot parse '{name} {value}': expected a boolean (eg True/False)")
+
+
+def _parse_optional_size(name: str, value: str) -> int | None:
+    """Parses a MAX_BATCH/MAX_DEFERRED_ITERATIONS/STREAM_BATCH_SIZE value:
+    any positive integer sets it (and implies the option is wanted "on");
+    "0"/"none"/"off"/"false" (case-insensitively) mean "leave it at its
+    REX::npos ("off"/"no cap") default", represented here as None."""
+    v = value.strip().lower()
+    if v in _FALSE_VALUES or v == "none":
+        return None
+    try:
+        parsed = int(value)
+    except ValueError as error:
+        raise ValueError(
+            f"cannot parse '{name} {value}': expected a non-negative integer, or "
+            f"none/off to leave it unset"
+        ) from error
+    if parsed < 0:
+        raise ValueError(f"'{name} {value}': must not be negative")
+    return parsed or None
+
+
+@dataclass
+class DriverOptions:
+    """Card-wide (not per-launch) Driver options -- see DRIVER_OPTION_NAMES
+    and the module docstring. Every field is None ("don't touch, leave
+    Driver's own default") unless the card actually set it; apply() is what
+    turns a resolved instance into the matching Driver calls."""
+    helicity_reweighting: bool | None = None
+    verbose: bool | None = None
+    max_batch: int | None = None
+    max_deferred_iterations: int | None = None
+    # streaming=None with stream_batch_size set still means "on": see
+    # apply_options(). streaming=False always means "off", regardless of
+    # stream_batch_size (matches Driver.set_streaming(False)'s own C++
+    # semantics, which ignores its own batch_size argument when disabling).
+    streaming: bool | None = None
+    stream_batch_size: int | None = None
+    # None: the generation's thread count / DEFAULT_MIN_BATCH (set by
+    # madtrex_driver before apply())
+    nb_core: int | None = None
+    min_batch: int | None = None
+
+    def apply(self, driver) -> None:
+        """Applies whichever fields aren't None to driver (a
+        madtrex.Driver) via its matching set_*() method. Not typed
+        as madtrex.Driver directly so this module stays free of a
+        madspace/pybind11 import of its own; see madtrex.py's own use of
+        this for the concrete type."""
+        if self.helicity_reweighting is not None:
+            driver.set_helicity_reweighting(self.helicity_reweighting)
+        if self.verbose is not None:
+            driver.set_verbose(self.verbose)
+        if self.max_batch is not None:
+            driver.set_max_batch(self.max_batch)
+        if self.max_deferred_iterations is not None:
+            driver.set_max_deferred_iterations(self.max_deferred_iterations)
+        if self.nb_core is not None:
+            driver.set_threads(self.nb_core)
+        if self.min_batch is not None:
+            driver.set_min_batch(self.min_batch)
+        streaming = self.streaming
+        if streaming is None and self.stream_batch_size is not None:
+            streaming = True  # a batch size with no explicit STREAMING line still means "on"
+        if streaming is True:
+            if self.stream_batch_size is not None:
+                driver.set_streaming(True, self.stream_batch_size)
+            else:
+                driver.set_streaming(True)
+        elif streaming is False:
+            driver.set_streaming(False)
+
+
+@dataclass
+class LaunchBlock:
+    name: str | None
+    # (block, id, value); (name, None, value) for a parameter given by its
+    # param_card name, until resolve_parameter_names()
+    sets: list[tuple[str, str | None, str]] = field(default_factory=list)
+    # None: use the card's (or DEFAULT_FPTYPE, if that's also unset) baseline
+    # -- see group_by_fptype().
+    fptype: str | None = None
+
+
+@dataclass
+class ModelGroup:
+    # model_spec is None for launches that never followed a "change model"
+    # line, ie ones meant to run against the original process directory as-is.
+    model_spec: str | None
+    process_definition: str | None
+    launches: list[LaunchBlock] = field(default_factory=list)
+
+
+@dataclass
+class ParsedCard:
+    # One entry per distinct (model_spec, process_definition) target
+    # actually used by a launch, in order of first appearance.
+    groups: list[ModelGroup]
+    # The "set FPTYPE ..." found before the card's first "launch" line, if
+    # any; see group_by_fptype().
+    baseline_fptype: str | None = None
+    # Resolved from every "set <DRIVER_OPTION_NAME> ..." line found before
+    # the card's first "launch"; see DriverOptions.
+    driver_options: DriverOptions = field(default_factory=DriverOptions)
+
+
+@dataclass
+class FptypeGroup:
+    # Mirrors ModelGroup, further split by resolved (not None) fptype -- see
+    # group_by_fptype(). model_spec/process_definition carry the same
+    # meaning as on ModelGroup.
+    model_spec: str | None
+    process_definition: str | None
+    fptype: str
+    launches: list[LaunchBlock] = field(default_factory=list)
+
+
+def _set_option(
+    name: str,
+    value: str,
+    raw_line: str,
+    after_launch: bool,
+    current_group: "ModelGroup | None",
+    baseline_fptype: str | None,
+    driver_options: DriverOptions,
+) -> str | None:
+    """Applies "change <name> <value>" (name one of "fptype", "helicity",
+    DRIVER_OPTION_NAMES) and returns the possibly updated baseline fptype."""
+    if name == "fptype":
+        if not after_launch:
+            # Nothing has launched anywhere in the card yet: this is the
+            # baseline (see the module docstring / group_by_fptype()),
+            # not a per-launch override.
+            if baseline_fptype is not None and baseline_fptype != value:
+                raise ValueError(
+                    f"reweight card sets a baseline FPTYPE more than once with "
+                    f"different values ('{baseline_fptype}' and '{value}')"
+                )
+            return value
+        current_group.launches[-1].fptype = value
+        return baseline_fptype
+    if after_launch:
+        raise ValueError(
+            f"reweight card sets driver option '{name}' after the first "
+            f"'launch': {raw_line!r}; driver options apply to a whole "
+            f"reweighting pass (see DriverOptions) and can only be set in the "
+            f"card's preamble, before any 'launch'"
+        )
+    if name == "helicity":
+        driver_options.helicity_reweighting = _parse_bool(name, value)
+    elif name in ("verbose", "streaming"):
+        setattr(driver_options, name, _parse_bool(name, value))
+    elif name == "nb_core" and value.strip().lower() == "auto":
+        driver_options.nb_core = None
+    else:
+        setattr(driver_options, name, _parse_optional_size(name, value))
+    return baseline_fptype
+
+
+def parse_reweight_card(text: str) -> ParsedCard:
+    """Splits text into per-(model, process)-target groups of "launch"/"set"
+    blocks. Lines are matched loosely (case-insensitive, "#" strips trailing
+    comments) to stay tolerant of the same card conventions the legacy
+    reweight_card.dat format allows (eg "set param_card BLOCK ID VALUE
+    # orig: ...").
+
+    Raises if a "launch" follows a "change model" with no paired "change
+    process" yet (MG7 has no other way to learn the process definition), or if
+    the same model is later paired with a different process definition (each
+    model may be regenerated for at most one process per card), if a "change
+    fptype ..." appears before the first "launch" more than once with
+    different values, if "change helicity"/"change <DRIVER_OPTION_NAME> ..."
+    appears anywhere other than before the first "launch", if such a line's
+    value can't be parsed as whatever type that option expects, or for a
+    line only the classic reweighting module understands (see
+    generic_reweight_card.classic_only_reason()) or that belongs to PDF
+    reweighting."""
+    groups: dict[str | None, ModelGroup] = {}
+    order: list[str | None] = []
+    current_model_spec: str | None = None
+    current_process_definition: str | None = None
+    current_group: ModelGroup | None = None
+    baseline_fptype: str | None = None
+    driver_options = DriverOptions()
+
+    for raw_line in text.splitlines():
+        line = _vocab.strip_comment(raw_line)
+        if not line.strip():
+            continue
+        _vocab.check_not_classic_only(line, raw_line)
+
+        model_match = _MODEL_RE.match(line)
+        if model_match:
+            current_model_spec = model_match.group("spec").strip()
+            continue
+
+        process_match = _PROCESS_RE.match(line)
+        if process_match:
+            current_process_definition = process_match.group("definition").strip()
+            continue
+
+        launch_match = _LAUNCH_RE.match(line)
+        if launch_match:
+            if current_model_spec is not None and current_process_definition is None:
+                raise ValueError(
+                    f"reweight card requests 'change model {current_model_spec}' but "
+                    f"has no 'change process <definition>' line before the next "
+                    f"'launch': MG7 does not persist the original process definition "
+                    f"anywhere on disk, so it must be given again explicitly to "
+                    f"regenerate it in the new model"
+                )
+            group = groups.get(current_model_spec)
+            if group is None:
+                group = ModelGroup(current_model_spec, current_process_definition)
+                groups[current_model_spec] = group
+                order.append(current_model_spec)
+            elif current_model_spec is not None and group.process_definition != current_process_definition:
+                raise ValueError(
+                    f"reweight card requests model '{current_model_spec}' with two "
+                    f"different process definitions ('{group.process_definition}' and "
+                    f"'{current_process_definition}'); MadtRex only supports "
+                    f"regenerating a single process per model per card"
+                )
+            name_match = _NAME_RE.search(launch_match.group("rest"))
+            group.launches.append(LaunchBlock(name=name_match.group("name") if name_match else None))
+            current_group = group
+            continue
+
+        words = line.split()
+        name = words[1].lower() if len(words) > 1 else ""
+        after_launch = current_group is not None and bool(current_group.launches)
+
+        if words[0].lower() == "change":
+            if name == "backend":
+                _vocab.check_backend_line(words, raw_line)
+                continue
+            if name == "mode":
+                mode = words[2].lower() if len(words) > 2 else ""
+                if mode == "lo":
+                    continue
+                raise ValueError(
+                    f"{raw_line.strip()!r} in a parameter reweight card: only 'change mode "
+                    f"LO' belongs here; PDF reweighting goes in a 'change mode pdf' section "
+                    f"of the card"
+                )
+            if name in _PDF_NAMES:
+                raise ValueError(
+                    f"{raw_line.strip()!r} is a PDF reweighting option: put it in a "
+                    f"'change mode pdf' section of the card"
+                )
+            if name == "helicity" or name in MADTREX_CHANGE_OPTIONS:
+                if len(words) != 3:
+                    raise ValueError(f"cannot parse {raw_line.strip()!r}: expected 'change {name} <value>'")
+                baseline_fptype = _set_option(
+                    name, words[2], raw_line, after_launch, current_group, baseline_fptype,
+                    driver_options,
+                )
+                continue
+            _vocab.logger.warning("reweight card: ignoring unknown option %r", raw_line.strip())
+            continue
+
+        if words[0].lower() == "set":
+            if name == "mode":
+                raise ValueError(
+                    f"{raw_line.strip()!r}: card sections are read by the generic reweight "
+                    f"card reader (generic_reweight_card.py), not inside a parameter card"
+                )
+            if name in _PDF_NAMES and len(words) == 3:
+                raise ValueError(
+                    f"{raw_line.strip()!r} is a PDF reweighting setting: put it in a "
+                    f"'change mode pdf' section of the card"
+                )
+            if name in DEPRECATED_SET_NAMES and len(words) == 3:
+                _vocab.deprecated(f"set {words[1]}", f"change {DEPRECATED_SET_NAMES[name]}")
+                baseline_fptype = _set_option(
+                    DEPRECATED_SET_NAMES[name], words[2], raw_line, after_launch,
+                    current_group, baseline_fptype, driver_options,
+                )
+                continue
+            set_match = _SET_RE.match(line)
+            named_match = None if set_match else _NAMED_SET_RE.match(line)
+            if set_match or named_match:
+                if not after_launch:
+                    raise ValueError(f"reweight card 'set' line appears before any 'launch': {raw_line!r}")
+                if set_match:
+                    entry = (set_match.group("block"), " ".join(set_match.group("id").split()),
+                             set_match.group("value"))
+                else:
+                    entry = (named_match.group("name"), None, named_match.group("value"))
+                current_group.launches[-1].sets.append(entry)
+            continue
+
+    return ParsedCard(
+        groups=[groups[key] for key in order],
+        baseline_fptype=baseline_fptype,
+        driver_options=driver_options,
+    )
+
+
+def _expand_launch(block: LaunchBlock, base_index: int) -> list[LaunchBlock]:
+    """Expands one LaunchBlock's scan:[...] values (if any) into the Cartesian
+    product of concrete-valued LaunchBlocks. "set" lines sharing the same
+    scanN: group index vary together; ungrouped ("scan:") entries each vary
+    independently -- mirrors models.check_param_card.ParamCardIterator.iterate()."""
+    groups: dict[object, list[int]] = {}
+    values_by_index: dict[int, list] = {}
+    for i, (_, _, raw_value) in enumerate(block.sets):
+        value = raw_value.strip()
+        if not value.lower().startswith("scan"):
+            continue
+        scan_match = _SCAN_RE.match(value)
+        if not scan_match:
+            continue
+        try:
+            values_by_index[i] = eval(scan_match.group("values"))
+        except SyntaxError as error:
+            raise ValueError(f"invalid scan definition {raw_value!r}: {error}") from error
+        group = scan_match.group("group")
+        key = group if group else object()  # ungrouped: unique key, varies independently
+        groups.setdefault(key, []).append(i)
+
+    if not groups:
+        return [block]
+
+    keys = list(groups.keys())
+    lengths = [range(len(values_by_index[groups[key][0]])) for key in keys]
+    name = block.name or f"rwgt_{base_index}"
+    expanded = []
+    for point, positions in enumerate(itertools.product(*lengths)):
+        sets = list(block.sets)
+        for key, pos in zip(keys, positions):
+            for i in groups[key]:
+                block_name, param_id, _ = block.sets[i]
+                sets[i] = (block_name, param_id, repr(values_by_index[i][pos]))
+        expanded.append(LaunchBlock(name=f"{name}_{point + 1}", sets=sets, fptype=block.fptype))
+    return expanded
+
+
+def expand_scans(card: ParsedCard) -> ParsedCard:
+    """Returns a copy of card with every launch's scan:[...] values resolved
+    into explicit per-point launch blocks, within each model group
+    independently; a no-op for cards with no scans."""
+    groups = []
+    for group in card.groups:
+        launches = []
+        for i, block in enumerate(group.launches):
+            launches.extend(_expand_launch(block, i + 1))
+        groups.append(ModelGroup(group.model_spec, group.process_definition, launches))
+    return ParsedCard(groups=groups, baseline_fptype=card.baseline_fptype, driver_options=card.driver_options)
+
+
+def group_by_fptype(card: ParsedCard, default_fptype: str = DEFAULT_FPTYPE) -> list[FptypeGroup]:
+    """Splits each of card.groups' launches further by their own resolved
+    *target* fptype (LaunchBlock.fptype, defaulting to card.baseline_fptype,
+    itself defaulting to default_fptype if the card set neither) -- launches
+    targeting a different fptype need an independently-compiled matrix
+    element library (see Driver.set_fptype_priority() in madtrex.hpp), so
+    unlike ordinary physics "set" values (handled within one Driver/
+    WareHouse pass via SLHA iterators) they can't share their model group's
+    single pass. Groups are kept in order of first appearance of each
+    (model_spec, process_definition, fptype) combination, merging
+    non-contiguous launches for the same combination together -- the same
+    tolerance for interleaving ModelGroup's own (model, process) grouping
+    already has.
+
+    default_fptype is the module constant DEFAULT_FPTYPE unless the caller
+    overrides it -- madtrex.py's run_parameter_reweighting() always passes
+    the process directory's own run_card.toml [run] fptype instead (falling
+    back to DEFAULT_FPTYPE only if that can't be read), so a card that sets
+    no FPTYPE at all resolves every launch's target to the sample's own
+    actual generation-time precision -- a no-op fptype-wise -- rather than a
+    hardcoded value that may not match. This is purely about each launch's
+    *target*; the separate, always-on normaliser-fptype concern (see the
+    module docstring) is resolved independently in madtrex.py, not here."""
+    baseline = card.baseline_fptype or default_fptype
+    result: dict[tuple[str | None, str | None, str], FptypeGroup] = {}
+    order: list[tuple[str | None, str | None, str]] = []
+    for group in card.groups:
+        for launch in group.launches:
+            fptype = launch.fptype or baseline
+            key = (group.model_spec, group.process_definition, fptype)
+            fgroup = result.get(key)
+            if fgroup is None:
+                fgroup = FptypeGroup(group.model_spec, group.process_definition, fptype)
+                result[key] = fgroup
+                order.append(key)
+            fgroup.launches.append(launch)
+    return [result[key] for key in order]
+
+
+def read_param_card_names(text: str) -> dict[str, tuple[str, str]]:
+    """{name: (block, id)} for the parameters of the param_card text, named
+    by the comment of their line as the classic reweighting module does
+    (eg "  6 1.730000e+02 # MT" in BLOCK MASS, "DECAY 6 1.49 # WT"), names
+    lowercase. An entry with several indices (eg "  1  2 5.0e-01 # NMIX12")
+    gets them all, space separated ("1 2")."""
+    names: dict[str, tuple[str, str]] = {}
+    block = None
+    for line in text.splitlines():
+        content, _, comment = line.partition("#")
+        words = content.split()
+        if not words:
+            continue
+        keyword = words[0].lower()
+        if keyword == "block":
+            block = words[1] if len(words) > 1 else None
+            continue
+        if keyword == "decay":
+            block = None
+            if len(words) >= 3 and comment.split():
+                names.setdefault(comment.split()[0].lower(), ("DECAY", words[1]))
+            continue
+        if block is not None and len(words) >= 2 and comment.split():
+            names.setdefault(comment.split()[0].lower(), (block, " ".join(words[:-1])))
+    return names
+
+
+def has_parameter_names(launches: list[LaunchBlock]) -> bool:
+    """Some launch sets a parameter by its param_card name."""
+    return any(param_id is None for block in launches for _, param_id, _ in block.sets)
+
+
+def resolve_parameter_names(
+    launches: list[LaunchBlock], param_card_text: str
+) -> list[LaunchBlock]:
+    """launches with every parameter given by its param_card name ("set MT
+    175") replaced by its "set BLOCK ID VALUE" form, looked up in the
+    param_card text (see read_param_card_names()). Raises for a name the
+    param_card does not have."""
+    names = read_param_card_names(param_card_text)
+    resolved = []
+    for block in launches:
+        sets = []
+        for name, param_id, value in block.sets:
+            if param_id is None:
+                found = names.get(name.lower())
+                if found is None:
+                    found = names.get(f"mdl_{name.lower()}")
+                if found is None:
+                    raise ValueError(
+                        f"reweight card: no parameter named {name!r} in the param_card; "
+                        f"use 'set BLOCK ID VALUE'"
+                    )
+                name, param_id = found
+            sets.append((name, param_id, value))
+        resolved.append(LaunchBlock(name=block.name, sets=sets, fptype=block.fptype))
+    return resolved
+
+
+def render_launches(launches: list[LaunchBlock]) -> str:
+    """Renders one group's launches back into the literal "launch"/
+    "set BLOCK ID VALUE" text ParamHandler.read_rwgt_card() understands -- no
+    "change model"/"change process" lines, no "change fptype ..." lines (that's
+    consumed entirely at the Python level -- see group_by_fptype() and its
+    use in madtrex.py -- and never has any meaning to REX::tea::rwgt_slha's
+    own SLHA-only grammar), and no scan:[...] values, all of which must
+    already be resolved (see expand_scans() and the group handling in
+    madtrex.py)."""
+    lines = []
+    for block in launches:
+        lines.append(f"launch --rwgt_name={block.name}" if block.name else "launch")
+        for block_name, param_id, value in block.sets:
+            if param_id is None:
+                raise ValueError(
+                    f"parameter {block_name!r} is given by name: resolve it first "
+                    f"(resolve_parameter_names())"
+                )
+            lines.append(f"set {block_name} {param_id} {value}")
+    return "\n".join(lines) + "\n"
diff --git a/madgraph/iolibs/template_files/mg7/pdf_reweight_card.py b/madgraph/iolibs/template_files/mg7/pdf_reweight_card.py
new file mode 100644
index 0000000000..9387bc4912
--- /dev/null
+++ b/madgraph/iolibs/template_files/mg7/pdf_reweight_card.py
@@ -0,0 +1,479 @@
+"""Parsing/preprocessing for MadtRex PDF reweight cards.
+
+This is a sibling of reweight_card.py: same "launch"/"set NAME VALUE" lexical
+shape, but a completely different, PDF-reweighting-specific vocabulary of
+recognized names (see PDF_OPTION_NAMES) -- kept as its own module rather than
+folded into reweight_card.py's grammar since the two modes are otherwise
+unrelated (see madtrex.py's run_reweighting()/run_pdf_reweighting(), and the
+module docstring there for why PDF and parameter reweighting are two
+independent passes to begin with).
+
+- "set PDF_SET <name>" (or "set BEAM1_PDF_SET <name>" + "set BEAM2_PDF_SET
+  <name>" together, for asymmetric beams -- eg ep colliders, or proton-ion
+  running where one beam uses a nuclear PDF set) chooses which
+  Driver.load_pdf_set()'d set a launch reweights to. Before the card's first
+  "launch", this is the *baseline*: which PDF set (and, by extension, scale
+  choice -- see below) the input sample's own weight is taken to already
+  correspond to. Unlike reweight_card.py's FPTYPE baseline, there is no
+  hard-set fallback default here: MadtRex does not currently record which
+  PDF set an existing sample was generated with anywhere on the sample
+  itself (a future hook might sniff the LHE file or other metadata for
+  this). This module's own resolve() raises if the baseline PDF set is
+  still unset by the time it's called -- but madtrex.py's
+  run_pdf_reweighting() fills it in first from the originating run's own
+  run_card.toml ([beam] pdf) when one is found next to the events being
+  reweighted, so an explicit PDF_SET is only actually required when no such
+  run_card is available (a bare card handed to a standalone/relocated
+  sample, or one deliberately reweighting to a *different* baseline than
+  the run actually used). A launch that doesn't set its own PDF_SET/
+  BEAM1_PDF_SET/BEAM2_PDF_SET inherits the baseline's (see resolve());
+- "set MUF <value>"/"set MUR <value>" (an absolute scale, in GeV, the same
+  for every event) and "set MUF_FACTOR <value>"/"set MUR_FACTOR <value>" (a
+  multiplier of each event's own recorded scale -- the default, factor 1.0,
+  ie "the event's own nominal scale") -- see madtrex.ScaleChoice,
+  which these map onto directly. Settable per launch (defaulting to factor
+  1.0, *not* inherited from the baseline's own scale choice -- a launch with
+  no MUF/MUR line reweights at the nominal scale, the common case);
+- "change alpha_s_power <int>", "change verbose <bool>", "change
+  scale_variation_output <none|envelope|all>": card-wide PdfDriver options,
+  exactly analogous to reweight_card.py's DRIVER_OPTION_NAMES -- preamble-
+  only (before the first "launch"), since they configure the whole
+  PdfDriver, not any one hypothesis. Their former spellings "set
+  ALPHA_S_POWER ..." etc. are still accepted, with a deprecation warning;
+- "change helicity ..." (or its former spellings "set HELICITY_REWEIGHTING"/
+  "set USE_HELICITY") is an error here: PDF reweighting rescales by PDF
+  ratios and never evaluates a matrix element, so it belongs to a parameter
+  reweight card (or a generic card's "change mode LO" section), as do the
+  other parameter-reweighting directives (change model/process/fptype,
+  "set BLOCK ID VALUE", ...);
+- "change mode pdf" and "change backend madtrex|auto" are accepted; the
+  directives only the classic reweighting module understands (see
+  generic_reweight_card.py) are errors;
+- scan:[...] (or scanN:[...] to vary several together) values on a launch's
+  own MUF/MUR/MUF_FACTOR/MUR_FACTOR lines, expanded into explicit per-point
+  launches via the same eval-based mechanism and group-index semantics as
+  reweight_card.py's own scan:[...] (see expand_scans()/_expand_launch()).
+  Scoped to the four scale directives only -- PDF_SET/BEAM1_PDF_SET/
+  BEAM2_PDF_SET are not scannable this way (list a separate "launch" per PDF
+  set instead), matching how reweight_card.py's own "change model" is a
+  distinct, non-scannable mechanism from its scan:[...] physics values.
+"""
+
+import itertools
+import re
+from dataclasses import dataclass, field
+
+try:
+    from . import generic_reweight_card as _vocab
+except ImportError:  # loaded as a plain module (gridpack bin/, tests)
+    import os as _os
+    import sys as _sys
+    _sys.path.append(_os.path.dirname(_os.path.abspath(__file__)))
+    import generic_reweight_card as _vocab
+
+# Recognized "set <NAME> <value>" names, lowercase, matched case-
+# insensitively against the card. PDF_SET_NAMES are valid both before the
+# first "launch" (the baseline) and within one (a per-launch override);
+# SCALE_NAMES are valid within a launch (or before the first one, to set the
+# baseline's own scale choice -- rarely needed, defaults to nominal);
+# CARD_OPTION_NAMES ("change <NAME> <value>", formerly "set") are
+# preamble-only.
+PDF_SET_NAMES = frozenset({"pdf_set", "beam1_pdf_set", "beam2_pdf_set"})
+SCALE_NAMES = frozenset({"muf", "mur", "muf_factor", "mur_factor"})
+CARD_OPTION_NAMES = frozenset({"alpha_s_power", "verbose", "scale_variation_output"})
+PDF_OPTION_NAMES = PDF_SET_NAMES | SCALE_NAMES | CARD_OPTION_NAMES
+# Parameter-reweighting "change" options and former "set" names: errors here.
+_PARAMETER_CHANGE_OPTIONS = frozenset({
+    "helicity", "model", "process", "fptype", "max_batch", "max_deferred_iterations",
+    "streaming", "stream_batch_size", "nb_core", "min_batch",
+})
+_PARAMETER_SET_NAMES = frozenset({"helicity_reweighting", "use_helicity"}) | (
+    _PARAMETER_CHANGE_OPTIONS - {"nb_core", "min_batch"})
+
+_LAUNCH_RE = re.compile(r"^\s*launch\b(?P<rest>.*)$", re.I)
+_NAME_RE = re.compile(r"rwgt_name\s*=\s*(?P<name>\S+)", re.I)
+_OPTION_RE = re.compile(r"^\s*(?:set|change)\s+(?P<name>\S+)\s+(?P<value>.+?)\s*$", re.I)
+_SCAN_RE = re.compile(r"scan\s*(?P<group>\d*)\s*:\s*(?P<values>.*)$", re.I)
+_TRUE_VALUES = frozenset({"1", "true", "yes", "on", "t", ".true."})
+_FALSE_VALUES = frozenset({"0", "false", "no", "off", "f", ".false."})
+_SCALE_VARIATION_OUTPUTS = frozenset({"none", "envelope", "all"})
+
+
+def _parse_bool(name: str, value: str) -> bool:
+    v = value.strip().lower()
+    if v in _TRUE_VALUES:
+        return True
+    if v in _FALSE_VALUES:
+        return False
+    raise ValueError(f"cannot parse '{name} {value}': expected a boolean (eg true/false)")
+
+
+def _parse_int(name: str, value: str) -> int:
+    try:
+        return int(value)
+    except ValueError as error:
+        raise ValueError(f"cannot parse '{name} {value}': expected an integer") from error
+
+
+def _parse_float(name: str, value: str) -> float:
+    try:
+        return float(value)
+    except ValueError as error:
+        raise ValueError(f"cannot parse '{name} {value}': expected a number") from error
+
+
+def _resolve_scale(name: str, value: str) -> tuple[str, "PdfScale"]:
+    """Parses one already-lowercased MUF/MUR/MUF_FACTOR/MUR_FACTOR (name,
+    value) pair into which field it targets ("muF" or "muR") and the
+    PdfScale it resolves to. Shared by the baseline's eager resolution
+    (parse_pdf_reweight_card()) and a launch's post-scan-expansion
+    resolution (_finalize_scales()), so both go through identical parsing."""
+    if name == "muf":
+        return "muF", PdfScale(_parse_float(name, value), absolute=True)
+    if name == "mur":
+        return "muR", PdfScale(_parse_float(name, value), absolute=True)
+    if name == "muf_factor":
+        return "muF", PdfScale(_parse_float(name, value), absolute=False)
+    return "muR", PdfScale(_parse_float(name, value), absolute=False)  # mur_factor
+
+
+@dataclass
+class PdfScale:
+    """Mirrors madtrex.ScaleChoice (kept as plain data here so this
+    module stays usable without a madspace install -- same rationale as
+    reweight_card.DriverOptions): value, and whether it's absolute (GeV,
+    MUF/MUR) or a multiplicative factor of the event's own recorded scale
+    (MUF_FACTOR/MUR_FACTOR, the default -- factor 1.0 is "the event's own
+    nominal scale")."""
+    value: float = 1.0
+    absolute: bool = False
+
+
+@dataclass
+class PdfLaunch:
+    name: str | None
+    pdf_set_name: str | None = None
+    beam1_pdf_set_name: str | None = None
+    beam2_pdf_set_name: str | None = None
+    # Raw (name, value) pairs from this launch's own "set MUF/MUR/
+    # MUF_FACTOR/MUR_FACTOR <value>" lines, in card order -- kept unparsed
+    # (rather than resolved into PdfScale directly, unlike pdf_set_name
+    # above) so expand_scans() can substitute concrete scan:[...] values
+    # before _finalize_scales() actually parses them into muF_scale/
+    # muR_scale (see resolve()). The last entry for a given name wins,
+    # matching every other "set" directive in this card.
+    scale_sets: list[tuple[str, str]] = field(default_factory=list)
+
+
+@dataclass
+class PdfBaseline:
+    pdf_set_name: str | None = None
+    beam1_pdf_set_name: str | None = None
+    beam2_pdf_set_name: str | None = None
+    muF_scale: PdfScale = field(default_factory=PdfScale)
+    muR_scale: PdfScale = field(default_factory=PdfScale)
+
+
+@dataclass
+class ParsedPdfCard:
+    baseline: PdfBaseline = field(default_factory=PdfBaseline)
+    alpha_s_power: int | None = None
+    verbose: bool | None = None
+    # "none"/"envelope"/"all" (lowercase) if set; see
+    # madtrex.ScaleVariationOutput, which this maps onto directly.
+    scale_variation_output: str | None = None
+    launches: list[PdfLaunch] = field(default_factory=list)
+
+
+@dataclass
+class ResolvedPdfLaunch:
+    name: str | None
+    # One name (same PDF set both beams) or two (beam1, beam2) -- pass as
+    # *pdf_set_names to PdfDriver.add_hypothesis()/set_baseline(), whose
+    # matching overload is selected by the resulting argument count.
+    pdf_set_names: tuple[str] | tuple[str, str]
+    muF_scale: PdfScale
+    muR_scale: PdfScale
+
+
+@dataclass
+class ResolvedPdfCard:
+    baseline_pdf_set_names: tuple[str] | tuple[str, str]
+    baseline_muF_scale: PdfScale
+    baseline_muR_scale: PdfScale
+    alpha_s_power: int | None
+    verbose: bool | None
+    scale_variation_output: str | None
+    launches: list[ResolvedPdfLaunch]
+
+
+def parse_pdf_reweight_card(text: str) -> ParsedPdfCard:
+    """Splits text into a ParsedPdfCard: a baseline (PDF_SET/BEAM1_PDF_SET+
+    BEAM2_PDF_SET/MUF/MUR/MUF_FACTOR/MUR_FACTOR lines found before the first
+    "launch") plus a flat list of launches (unlike reweight_card.py's
+    ModelGroup/FptypeGroup, there is no further grouping here: a single
+    PdfDriver handles every launch in one pass, since -- unlike a different
+    UFO model or a different FPTYPE -- a different PDF set or scale choice
+    needs no separate library at all, just a different PdfSet/ScaleChoice
+    passed to the same already-loaded PdfDriver).
+
+    Lines are matched loosely (case-insensitive, "#" strips trailing
+    comments), same conventions as reweight_card.py.
+
+    Raises if a recognized card-wide option (ALPHA_S_POWER/VERBOSE/
+    SCALE_VARIATION_OUTPUT) appears anywhere other than before the first
+    "launch", or if a recognized line's value can't be parsed as whatever
+    type it expects. Does *not* raise here if no baseline PDF_SET was ever
+    given -- see resolve(), which is where that's actually needed and thus
+    where it's checked."""
+    baseline = PdfBaseline()
+    alpha_s_power: int | None = None
+    verbose: bool | None = None
+    scale_variation_output: str | None = None
+    launches: list[PdfLaunch] = []
+    current_launch: PdfLaunch | None = None
+
+    for raw_line in text.splitlines():
+        line = _vocab.strip_comment(raw_line)
+        if not line.strip():
+            continue
+
+        launch_match = _LAUNCH_RE.match(line)
+        if launch_match:
+            name_match = _NAME_RE.search(launch_match.group("rest"))
+            current_launch = PdfLaunch(name=name_match.group("name") if name_match else None)
+            launches.append(current_launch)
+            continue
+
+        _vocab.check_not_classic_only(line, raw_line)
+        words = line.split()
+        keyword = words[0].lower()
+        name = words[1].lower() if len(words) > 1 else ""
+        if keyword == "change":
+            if name == "backend":
+                _vocab.check_backend_line(words, raw_line)
+                continue
+            if name == "mode":
+                if len(words) > 2 and words[2].lower() == "pdf":
+                    continue
+                raise ValueError(
+                    f"{raw_line.strip()!r} in a PDF reweight card: parameter reweighting "
+                    f"goes in a 'change mode LO' section of the card"
+                )
+        if (keyword == "change" and name in _PARAMETER_CHANGE_OPTIONS) or (
+            keyword == "set" and name in _PARAMETER_SET_NAMES
+        ):
+            raise ValueError(
+                f"PDF reweight card sets {words[1]!r}: {raw_line!r}; this only applies "
+                f"to parameter (matrix-element) reweighting -- set it in a parameter "
+                f"reweight card, or in the 'change mode LO' section of a generic card"
+            )
+        if keyword == "set" and len(words) >= 4:
+            raise ValueError(
+                f"PDF reweight card sets a model parameter: {raw_line!r}; parameter "
+                f"(matrix-element) reweighting goes in a parameter reweight card, or in "
+                f"the 'change mode LO' section of a generic card"
+            )
+        option_match = _OPTION_RE.match(line) if keyword in ("set", "change") else None
+        if keyword == "change" and name not in CARD_OPTION_NAMES:
+            _vocab.logger.warning("PDF reweight card: ignoring unknown option %r", raw_line.strip())
+            continue
+        if not option_match or option_match.group("name").strip().lower() not in PDF_OPTION_NAMES:
+            continue  # not a recognized line: silently ignored, same tolerant philosophy as reweight_card.py
+        name = option_match.group("name").strip().lower()
+        value = option_match.group("value").strip()
+        if keyword == "set" and name in CARD_OPTION_NAMES:
+            _vocab.deprecated(f"set {option_match.group('name')}", f"change {name}")
+
+        if name in CARD_OPTION_NAMES:
+            if current_launch is not None:
+                raise ValueError(
+                    f"PDF reweight card sets card-wide option '{name}' after the first "
+                    f"'launch': {raw_line!r}; ALPHA_S_POWER/VERBOSE/SCALE_VARIATION_OUTPUT "
+                    f"configure the whole PdfDriver and can only be set in the card's "
+                    f"preamble, before any 'launch'"
+                )
+            if name == "alpha_s_power":
+                alpha_s_power = _parse_int(name, value)
+            elif name == "verbose":
+                verbose = _parse_bool(name, value)
+            else:  # scale_variation_output
+                v = value.strip().lower()
+                if v not in _SCALE_VARIATION_OUTPUTS:
+                    raise ValueError(
+                        f"cannot parse '{name} {value}': expected one of "
+                        f"{sorted(_SCALE_VARIATION_OUTPUTS)}"
+                    )
+                scale_variation_output = v
+            continue
+
+        target = current_launch if current_launch is not None else baseline
+        if name == "pdf_set":
+            target.pdf_set_name = value
+        elif name == "beam1_pdf_set":
+            target.beam1_pdf_set_name = value
+        elif name == "beam2_pdf_set":
+            target.beam2_pdf_set_name = value
+        elif current_launch is not None:
+            # A launch's own MUF/MUR/MUF_FACTOR/MUR_FACTOR: kept raw so
+            # expand_scans() gets a chance at any scan:[...] value first --
+            # see PdfLaunch.scale_sets and _finalize_scales() (called from
+            # resolve(), after expansion).
+            current_launch.scale_sets.append((name, value))
+        else:
+            # The baseline's own scale choice: resolved eagerly, same as
+            # pdf_set_name -- not scannable (see the module docstring).
+            field_name, scale = _resolve_scale(name, value)
+            setattr(baseline, f"{field_name}_scale", scale)
+
+    return ParsedPdfCard(
+        baseline=baseline,
+        alpha_s_power=alpha_s_power,
+        verbose=verbose,
+        scale_variation_output=scale_variation_output,
+        launches=launches,
+    )
+
+
+def _resolve_pdf_set_names(
+    pdf_set_name: str | None,
+    beam1_pdf_set_name: str | None,
+    beam2_pdf_set_name: str | None,
+    context: str,
+) -> tuple[str] | tuple[str, str]:
+    """Resolves PDF_SET/BEAM1_PDF_SET/BEAM2_PDF_SET into either a single-name
+    tuple (same set both beams) or a two-name tuple (beam1, beam2), matching
+    PdfDriver.set_baseline()/add_hypothesis()'s two overloads. context names
+    where this came from (eg "baseline", "launch 'foo'"), used only in error
+    messages."""
+    if pdf_set_name is not None and (beam1_pdf_set_name is not None or beam2_pdf_set_name is not None):
+        raise ValueError(
+            f"{context}: both PDF_SET and BEAM1_PDF_SET/BEAM2_PDF_SET are set; use one or "
+            f"the other"
+        )
+    if beam1_pdf_set_name is not None or beam2_pdf_set_name is not None:
+        if beam1_pdf_set_name is None or beam2_pdf_set_name is None:
+            raise ValueError(f"{context}: BEAM1_PDF_SET and BEAM2_PDF_SET must both be set together")
+        return (beam1_pdf_set_name, beam2_pdf_set_name)
+    if pdf_set_name is None:
+        raise ValueError(f"{context}: no PDF_SET (or BEAM1_PDF_SET/BEAM2_PDF_SET) is set")
+    return (pdf_set_name,)
+
+
+def _finalize_scales(scale_sets: list[tuple[str, str]]) -> tuple[PdfScale, PdfScale]:
+    """Resolves a launch's raw scale_sets (post scan-expansion) into its
+    final (muF_scale, muR_scale) -- both default to factor(1.0) (nominal)
+    if scale_sets sets neither. The last MUF/MUF_FACTOR entry wins for
+    muF_scale, independently for MUR/MUR_FACTOR and muR_scale."""
+    muF_scale, muR_scale = PdfScale(), PdfScale()
+    for name, value in scale_sets:
+        field_name, scale = _resolve_scale(name, value)
+        if field_name == "muF":
+            muF_scale = scale
+        else:
+            muR_scale = scale
+    return muF_scale, muR_scale
+
+
+def _expand_launch(launch: PdfLaunch, base_index: int) -> list[PdfLaunch]:
+    """Expands one PdfLaunch's scan:[...] MUF/MUR/MUF_FACTOR/MUR_FACTOR
+    values (if any) into the Cartesian product of concrete-valued
+    PdfLaunches -- same scanN:/scan: group-index semantics as
+    reweight_card._expand_launch() (scanN: entries sharing a group index
+    vary together; ungrouped scan: entries each vary independently).
+    PDF_SET/BEAM1_PDF_SET/BEAM2_PDF_SET are copied through unchanged onto
+    every expanded point -- they're resolved eagerly at parse time (see the
+    module docstring) and never scannable here."""
+    groups: dict[object, list[int]] = {}
+    values_by_index: dict[int, list] = {}
+    for i, (_, raw_value) in enumerate(launch.scale_sets):
+        value = raw_value.strip()
+        if not value.lower().startswith("scan"):
+            continue
+        scan_match = _SCAN_RE.match(value)
+        if not scan_match:
+            continue
+        try:
+            values_by_index[i] = eval(scan_match.group("values"))
+        except SyntaxError as error:
+            raise ValueError(f"invalid scan definition {raw_value!r}: {error}") from error
+        group = scan_match.group("group")
+        key = group if group else object()  # ungrouped: unique key, varies independently
+        groups.setdefault(key, []).append(i)
+
+    if not groups:
+        return [launch]
+
+    keys = list(groups.keys())
+    lengths = [range(len(values_by_index[groups[key][0]])) for key in keys]
+    name = launch.name or f"pdf_rwgt_{base_index}"
+    expanded = []
+    for point, positions in enumerate(itertools.product(*lengths)):
+        scale_sets = list(launch.scale_sets)
+        for key, pos in zip(keys, positions):
+            for i in groups[key]:
+                set_name, _ = launch.scale_sets[i]
+                scale_sets[i] = (set_name, repr(values_by_index[i][pos]))
+        expanded.append(
+            PdfLaunch(
+                name=f"{name}_{point + 1}",
+                pdf_set_name=launch.pdf_set_name,
+                beam1_pdf_set_name=launch.beam1_pdf_set_name,
+                beam2_pdf_set_name=launch.beam2_pdf_set_name,
+                scale_sets=scale_sets,
+            )
+        )
+    return expanded
+
+
+def expand_scans(card: ParsedPdfCard) -> ParsedPdfCard:
+    """Returns a copy of card with every launch's scan:[...] MUF/MUR/
+    MUF_FACTOR/MUR_FACTOR values resolved into explicit per-point launch
+    blocks (see _expand_launch()); a no-op for cards with no scans. Mirrors
+    reweight_card.expand_scans() -- call this before resolve()."""
+    launches = []
+    for i, launch in enumerate(card.launches):
+        launches.extend(_expand_launch(launch, i + 1))
+    return ParsedPdfCard(
+        baseline=card.baseline,
+        alpha_s_power=card.alpha_s_power,
+        verbose=card.verbose,
+        scale_variation_output=card.scale_variation_output,
+        launches=launches,
+    )
+
+
+def resolve(card: ParsedPdfCard) -> ResolvedPdfCard:
+    """Resolves card into a ResolvedPdfCard ready to drive PdfDriver calls:
+    the baseline's PDF set name(s) (raising if none was ever given -- see
+    the module docstring), and each launch's own, with any launch that set
+    none of PDF_SET/BEAM1_PDF_SET/BEAM2_PDF_SET inheriting the baseline's.
+    Call after expand_scans() if the card might contain scan:[...] values --
+    resolve() itself does not expand them, it just finalizes whatever
+    scale_sets each launch already has (see _finalize_scales())."""
+    baseline_names = _resolve_pdf_set_names(
+        card.baseline.pdf_set_name,
+        card.baseline.beam1_pdf_set_name,
+        card.baseline.beam2_pdf_set_name,
+        "PDF reweight card baseline",
+    )
+    resolved_launches = []
+    for i, launch in enumerate(card.launches):
+        context = f"PDF reweight card launch {launch.name!r}" if launch.name else f"PDF reweight card launch #{i + 1}"
+        if launch.pdf_set_name is None and launch.beam1_pdf_set_name is None and launch.beam2_pdf_set_name is None:
+            names = baseline_names
+        else:
+            names = _resolve_pdf_set_names(
+                launch.pdf_set_name, launch.beam1_pdf_set_name, launch.beam2_pdf_set_name, context
+            )
+        muF_scale, muR_scale = _finalize_scales(launch.scale_sets)
+        resolved_launches.append(ResolvedPdfLaunch(launch.name, names, muF_scale, muR_scale))
+    return ResolvedPdfCard(
+        baseline_pdf_set_names=baseline_names,
+        baseline_muF_scale=card.baseline.muF_scale,
+        baseline_muR_scale=card.baseline.muR_scale,
+        alpha_s_power=card.alpha_s_power,
+        verbose=card.verbose,
+        scale_variation_output=card.scale_variation_output,
+        launches=resolved_launches,
+    )
diff --git a/madgraph/iolibs/template_files/mg7/reweight_card_default.dat b/madgraph/iolibs/template_files/mg7/reweight_card_default.dat
new file mode 100644
index 0000000000..4f161423d9
--- /dev/null
+++ b/madgraph/iolibs/template_files/mg7/reweight_card_default.dat
@@ -0,0 +1,89 @@
+#*************************************************************************
+#                          Reweight Module                               *
+#            Matrix-element and PDF reweighting for MG7 outputs          *
+#*************************************************************************
+#
+# Two programs can reweight the events of an MG7 output, and both read this
+# card:
+#  - MadtRex (madtrex/, 'install madtrex' in MadGraph): matrix-element and
+#    PDF/scale reweighting on the events as written (lhe or lhe_npy)
+#  - the classic reweighting module (Mattelaer, arXiv:1607.00763): needs
+#    f2py; LO and NLO reweighting, density matrices and plugins
+# By default (change backend auto) MadtRex reweights whenever it can run the
+# card, else the classic module. The launch question's reweight switch
+# (classic/madtrex) or 'generate_events --reweight-backend=...' choose
+# explicitly, and win over the line below.
+#
+# Note:
+#   1) the value of alpha_s will be used from the event so the value in
+#      the param_card is not taken into account.
+#   2) It is (in general) dangerous/wrong to change parameters by a large
+#      amount, if this changes the shape of the matrix elements a lot.
+#      (For example, changing a particle's mass by much more than its
+#      width leads to very inaccurate result). In such a case, separate
+#      event generation runs are needed.
+#
+#************************************************************************
+# ENTER YOUR COMMANDS BELOW.
+#************************************************************************
+
+#change backend auto   # classic | madtrex | auto
+
+#change helicity True  # reweight each event at its own helicity (default),
+                       # False: helicity-summed matrix elements
+
+launch
+
+#  SPECIFY THE PARAMETERS TO CHANGE LIKE THIS:
+# set sminputs 1 130 # modify 1/alpha_EW
+# set MT 175         # a parameter by its param_card name
+
+
+#************************************************************************
+# Example of code for the computation of two weights:
+#
+# launch --rwgt_name=mt175  ! start the computation of the first weight
+#  set BLOCKNAME ID VALUE   ! rule to modify the current param_card
+#  set MT 175               ! or by the parameter's param_card name
+# launch                    ! start to compute a second weight
+#  set MT scan:[170, 180]   ! one weight per value
+#
+# Note: The commands to specify the parameters are AFTER the associated
+#       "launch".
+#
+#************************************************************************
+# Options ('change NAME VALUE', before the first 'launch'), both programs:
+#
+# change backend classic|madtrex|auto : who reweights (default auto)
+# change helicity True|False : helicity reweighting (default True). MadtRex
+#        stops on events without helicity information when True.
+# change mode LO : parameter (matrix-element) reweighting, the default
+# change model NAME / change process DEF : reweight with another model
+#        (MadtRex needs both lines). The weights are the new model's matrix
+#        element over the original one the events were generated with;
+#        MadtRex requires the regenerated subprocesses to match the original
+#        ones (same particles, flavours and helicities).
+#
+# MadtRex only:
+# change mode pdf : the lines that follow are PDF/scale reweighting:
+#        launch --rwgt_name=nnpdf
+#         set PDF_SET NNPDF40_nnlo_as_01180  ! or BEAM1_PDF_SET + BEAM2_PDF_SET
+#         set MUF_FACTOR 2                   ! or MUR_FACTOR, MUF, MUR (GeV)
+#        'change mode LO' switches back to parameter reweighting.
+# change fptype d|f|m : floating-point precision to reweight to (also per
+#        launch)
+# change verbose|max_batch|max_deferred_iterations|streaming|
+#        stream_batch_size VALUE : MadtRex driver tuning
+# change alpha_s_power|scale_variation_output VALUE : PDF reweighting options
+#
+# Classic module only:
+# change mode NLO|NLO_tree|LO+NLO : NLO reweighting (NLO samples)
+# change output default|2.0|unweight, change rwgt_dir PATH,
+# change tree_path/virtual_path PATH, change systematics, change multicore,
+# change boost, change process DEF --add, a param_card given by its path,
+# density-matrix options (change particle_in_density_matrix ...) and
+# reweighting plugins.
+#
+# Manual: https://cp3.irmp.ucl.ac.be/projects/madgraph/wiki/Reweight
+#         madtrex/README.md
+#************************************************************************
diff --git a/madgraph/iolibs/template_files/mg7/run_card.toml b/madgraph/iolibs/template_files/mg7/run_card.toml
index 42bcc709af..ca8fe56aee 100644
--- a/madgraph/iolibs/template_files/mg7/run_card.toml
+++ b/madgraph/iolibs/template_files/mg7/run_card.toml
@@ -45,6 +45,8 @@ save_gridpack = %(gridpack.save_gridpack)s
 include_source = %(gridpack.include_source)s
 include_madspace = %(gridpack.include_madspace)s
 include_madspace_source = %(gridpack.include_madspace_source)s
+include_madtrex = %(gridpack.include_madtrex)s
+include_madtrex_source = %(gridpack.include_madtrex_source)s
 compress = %(gridpack.compress)s
 # end of a run creating a gridpack; options: regular, minimal (no events), fix_max_weight (no events)
 run_mode = %(gridpack.run_mode)s
diff --git a/madgraph/iolibs/template_files/mg7/run_interface.py b/madgraph/iolibs/template_files/mg7/run_interface.py
index 24f6c8adef..2726335945 100644
--- a/madgraph/iolibs/template_files/mg7/run_interface.py
+++ b/madgraph/iolibs/template_files/mg7/run_interface.py
@@ -31,6 +31,24 @@
 logger = logging.getLogger('madevent')
 
 
+def sync_lhe_copy(source, target):
+    """Make the LHE file target hold the events of source, gzipped or not
+    as each file's name says; replaced atomically. The mg7 run's event file
+    (events.lhe[.gz]) and the copy the madevent tools read
+    (unweighted_events.lhe.gz, see MG7RunCmd._prepare_run_dir) are kept the
+    same this way after a reweighting wrote into one of them."""
+    import gzip
+    tmp = target + '.tmp'
+    opener_in = gzip.open if source.endswith('.gz') else open
+    opener_out = gzip.open if target.endswith('.gz') else open
+    if source.endswith('.gz') == target.endswith('.gz'):
+        shutil.copyfile(source, tmp)
+    else:
+        with opener_in(source, 'rb') as src, opener_out(tmp, 'wb') as dst:
+            shutil.copyfileobj(src, dst)
+    os.replace(tmp, target)
+
+
 @contextlib.contextmanager
 def _quiet_setup():
     """Silence the INFO chatter of a madevent interface being constructed.
@@ -264,6 +282,79 @@ def get_model(self):
         from madgraph.iolibs.template_files.mg7.launch import load_process_model
         return load_process_model(self.me_dir)
 
+    def do_reweight(self, line):
+        """``reweight [--backend=classic|madtrex|auto] ...``: the backend
+        option (also spelled ``--reweight-backend=``) chooses who reweights,
+        together with the reweight switch and the card's ``change backend``
+        (see madgraph/various/reweight_backend.py). MadtRex reweights this
+        run's events (an MG7 output); the classic module is the inherited
+        command, without the option."""
+        from madgraph.various import reweight_backend as rb
+        requested = None
+        words = []
+        for word in line.split():
+            if word.startswith(('--backend=', '--reweight-backend=')):
+                requested = word.split('=', 1)[1]
+            else:
+                words.append(word)
+        line = ' '.join(words)
+        card_path = pjoin(self.me_dir, 'Cards', 'reweight_card.dat')
+        if '-from_cards' in words and not os.path.exists(card_path):
+            return
+        if requested == 'classic' or not os.path.exists(card_path):
+            return self._classic_reweight(line)
+        with open(card_path) as fsock:
+            traits = rb.classify_card(fsock.read())
+        plugin = [w.split('=', 1)[1] for w in words if w.startswith('--plugin=')]
+        if '--mode=density' in words:
+            traits.density = True
+        decision = rb.select_backend(
+            [('reweight --backend', requested)], traits,
+            rb.probe_environment(self.me_dir, self.options),
+            plugin=plugin[0] if plugin else None)
+        if decision.backend == 'classic':
+            return self._classic_reweight(line)
+        logger.info(decision.describe())
+        from madgraph.iolibs.template_files.mg7 import madtrex
+        import tempfile
+        with tempfile.TemporaryDirectory() as tmp:
+            card = pjoin(tmp, 'reweight_card.dat')
+            with open(card_path) as fsock, open(card, 'w') as out:
+                out.write(rb.strip_backend_line(fsock.read()))
+            # in place: the weights go into the run's event file
+            madtrex.run_reweighting(card, events_path=self._mg7_lhe_path,
+                                    card_kind=decision.card_kind,
+                                    process_directory=self.me_dir)
+        tool_copy = self._tool_copy()
+        if os.path.exists(tool_copy):
+            sync_lhe_copy(self._mg7_lhe_path, tool_copy)
+
+    def _tool_copy(self):
+        return pjoin(os.path.dirname(self._mg7_lhe_path), 'unweighted_events.lhe.gz')
+
+    def _classic_reweight(self, line):
+        """The classic module, which writes its weights into the tools' copy
+        of the events (unweighted_events.lhe.gz); the run's own event file
+        gets them too, as with MadtRex."""
+        out = super(MG7RunCmd, self).do_reweight(line)
+        self.sync_reweighted_events()
+        return out
+
+    def sync_reweighted_events(self):
+        """Copy the tools' copy of the events (unweighted_events.lhe[.gz])
+        onto the run's event file when it is newer, i.e. when the classic
+        reweighting module wrote into it. Safe to call again (e.g. after
+        store_result(), which may gzip the copy)."""
+        run_dir = os.path.dirname(self._mg7_lhe_path)
+        copies = [pjoin(run_dir, name) for name in
+                  ('unweighted_events.lhe.gz', 'unweighted_events.lhe')]
+        copies = [path for path in copies if os.path.exists(path)]
+        if not copies:
+            return
+        newest = max(copies, key=os.path.getmtime)
+        if os.path.getmtime(newest) > os.path.getmtime(self._mg7_lhe_path):
+            sync_lhe_copy(newest, self._mg7_lhe_path)
+
     def do_set(self, line, log=True):
         """Intercept the compiler options: the base handler patches
         ``Source/make_opts`` to switch Fortran/C++ compilers, which does not
diff --git a/madgraph/various/banner.py b/madgraph/various/banner.py
index eaa49f8703..0016744ee4 100755
--- a/madgraph/various/banner.py
+++ b/madgraph/various/banner.py
@@ -6861,6 +6861,12 @@ def default_setup(self):
             comment="include the installed madspace package")
         self.add_toml_param('gridpack', 'include_madspace_source', False,
             comment="include the madspace source code")
+        self.add_toml_param('gridpack', 'include_madtrex', True,
+            comment="include the installed madtrex package (MadtRex reweighting), "
+                    "if madtrex is installed")
+        self.add_toml_param('gridpack', 'include_madtrex_source', False,
+            comment="include the madtrex source code (with its bundled Rex); "
+                    "bin/reweight builds it when the gridpack has no usable madtrex")
         self.add_toml_param('gridpack', 'compress', False,
             comment="compress the gridpack into a tarball")
         self.add_toml_param('gridpack', 'run_mode', "minimal",
diff --git a/madgraph/various/misc.py b/madgraph/various/misc.py
index 53f9cdc114..f882f55f85 100755
--- a/madgraph/various/misc.py
+++ b/madgraph/various/misc.py
@@ -388,16 +388,23 @@ def is_exe(fpath):
                 return exe_file
     return None
 
+def default_f2py():
+    """The f2py to use when none is configured (f2py_compiler): the one
+    installed with this Python (next to sys.executable, e.g. in a virtual
+    environment), else f2pyX.Y, f2pyX or f2py on the PATH -- which may
+    belong to another Python, whose modules this one cannot import. None
+    when there is none."""
+    for name in ('f2py%d.%d' % sys.version_info[:2], 'f2py%d' % sys.version_info[0], 'f2py'):
+        path = os.path.join(os.path.dirname(sys.executable), name)
+        if os.path.isfile(path) and os.access(path, os.X_OK):
+            return path
+    for name in ('f2py%d.%d' % sys.version_info[:2], 'f2py%d' % sys.version_info[0], 'f2py'):
+        if which(name):
+            return name
+    return None
+
 def has_f2py():
-    has_f2py = False
-    if which('f2py'):
-        has_f2py = True
-    elif which('f2py%d.%d' %(sys.version_info.major, sys.version_info.minor)):
-        has_f2py = True
-    elif which('f2py%d' %(sys.version_info.major)):
-        has_f2py = True
-
-    return has_f2py       
+    return default_f2py() is not None
         
 #===============================================================================
 #  Activate dependencies if possible. Mainly for tests
@@ -3419,6 +3426,18 @@ def import_python_lhapdf(lhapdfconfig):
                 os.environ['LD_LIBRARY_PATH'] = lhapdf_libdir
             else:
                 os.environ['LD_LIBRARY_PATH'] = '%s:%s' %(lhapdf_libdir,os.environ['LD_LIBRARY_PATH'])
+        # LD_LIBRARY_PATH set now only reaches child processes: the python
+        # module (built without an rpath, e.g. by HEPTools) would not find
+        # libLHAPDF in this one. Load it first, globally, so it does.
+        for libname in ('libLHAPDF.so', 'libLHAPDF.dylib'):
+            libpath = os.path.join(lhapdf_libdir, libname)
+            if os.path.exists(libpath):
+                try:
+                    import ctypes
+                    ctypes.CDLL(libpath, mode=ctypes.RTLD_GLOBAL)
+                except OSError:
+                    pass
+                break
         try:
             candidates=[dirname for dirname in os.listdir(lhapdf_libdir) \
                             if os.path.isdir(os.path.join(lhapdf_libdir,dirname))]
diff --git a/madgraph/various/reweight_backend.py b/madgraph/various/reweight_backend.py
new file mode 100644
index 0000000000..b55eb71fb2
--- /dev/null
+++ b/madgraph/various/reweight_backend.py
@@ -0,0 +1,514 @@
+################################################################################
+#
+# Copyright (c) 2026 The MadGraph7 Development team and Contributors
+#
+# This file is a part of the MadGraph7 project, an application which
+# automatically generates Feynman diagrams and matrix elements for arbitrary
+# high-energy processes in the Standard Model and beyond.
+#
+# It is subject to the MadGraph7 license which should accompany this
+# distribution.
+#
+# For more information, visit madgraph.phys.ucl.ac.be and amcatnlo.web.cern.ch
+#
+################################################################################
+"""Which program reweights the events: the classic reweighting module
+(madgraph/interface/reweight_interface.py, f2py-compiled matrix elements) or
+MadtRex (madtrex/, MG7 outputs only).
+
+Both read the same Cards/reweight_card.dat. This module looks at that card
+and at the process directory and decides which backend runs it:
+
+- classify_card() lists what the card asks for: a "change backend" line, and
+  the directives only one of the backends understands (with line numbers);
+- probe_environment() describes the output (mg7/madevent/aMC@NLO, its run
+  card's output_format) and which backends can run here;
+- select_backend() combines the two with the backend the user asked for
+  (command line, launch-question switch, card) into a Decision, or raises
+  ReweightBackendError with every reason the request cannot be met.
+
+This module is pure Python: it imports neither madspace nor madtrex (whether
+MadtRex is installed is a file-system check), so it is safe to use when
+neither is installed.
+"""
+
+import os
+import re
+from dataclasses import dataclass, field
+from pathlib import Path
+
+BACKENDS = ('classic', 'madtrex', 'auto')
+
+# The card line choosing the backend: "change backend classic|madtrex|auto".
+_BACKEND_RE = re.compile(r'^\s*change\s+backend(?:\s+(?P<value>\S+))?\s*(?:#.*)?$', re.I)
+
+# "change mode" values: LO is understood by both backends, pdf only by
+# MadtRex, the NLO ones only by the classic module.
+SHARED_MODES = frozenset(['lo'])
+MADTREX_MODES = frozenset(['pdf', 'parameter'])
+CLASSIC_MODES = frozenset(['nlo', 'nlo_tree', 'lo+nlo'])
+
+# "change" options both backends understand (with the classic name and
+# default); "mode" depends on its value, see above.
+SHARED_CHANGE_OPTIONS = frozenset(['backend', 'model', 'process', 'helicity', 'mode'])
+
+_MG_ROOT = Path(__file__).resolve().parents[2]
+
+
+class ReweightBackendError(Exception):
+    """The requested reweighting cannot run. reasons lists every reason."""
+
+    def __init__(self, message, reasons=()):
+        self.reasons = list(reasons)
+        if self.reasons:
+            message = '%s:\n%s' % (message, '\n'.join('  - %s' % r for r in self.reasons))
+        super().__init__(message)
+
+
+@dataclass
+class Directive:
+    """One reweight-card line only one of the backends understands."""
+    line: int  # 1-based line number in the card
+    text: str  # the line, without its comment
+    why: str   # what makes it backend-specific
+
+    def __str__(self):
+        return 'line %d: %r (%s)' % (self.line, self.text, self.why)
+
+
+@dataclass
+class CardTraits:
+    """What a reweight card asks for (see classify_card())."""
+    # value of the card's "change backend" line, if any, and its line number
+    backend: 'str | None' = None
+    backend_line: 'int | None' = None
+    classic_only: list = field(default_factory=list)  # list[Directive]
+    madtrex_only: list = field(default_factory=list)  # list[Directive]
+    # a density-matrix card (DensityInterface); also listed in classic_only
+    density: bool = False
+    # which MadtRex runner reads the card: "param" (matrix-element
+    # reweighting), "pdf" (PDF/scale reweighting) or "generic" (sections
+    # selected with "change mode LO|pdf")
+    madtrex_kind: str = 'param'
+    launches: int = 0
+
+
+@dataclass
+class Env:
+    """The process directory and the installed backends (see
+    probe_environment()); tests build it by hand."""
+    # "mg7", "madevent", "aMC@NLO", or None when unknown
+    output_type: 'str | None' = 'mg7'
+    # the events to reweight are NLO events (an aMC@NLO output)
+    nlo_events: bool = False
+    # [run] output_format of an mg7 run_card.toml (compact_npy, lhe_npy, lhe)
+    output_format: 'str | None' = None
+    # the classic module can run (MadGraph + f2py), and why not otherwise
+    classic_available: bool = True
+    classic_reason: str = ''
+    # madtrex/install/madtrex exists
+    madtrex_installed: bool = False
+    # why the installed madtrex cannot be used although it is there (built
+    # against another madspace), or ''
+    madtrex_problem: str = ''
+
+
+@dataclass
+class Decision:
+    backend: str            # "classic" or "madtrex"
+    card_kind: str          # MadtRex: param/pdf/generic; classic: "classic" or "density"
+    reasons: list = field(default_factory=list)   # why this backend
+    warnings: list = field(default_factory=list)  # e.g. disagreeing explicit requests
+    source: str = 'auto'    # "auto", or which request chose the backend
+    plugin: 'str | None' = None  # a classic reweighting plugin
+
+    def describe(self, card_name='reweight_card.dat'):
+        """One log line, e.g. "Reweighting with madtrex (auto: MG7 output,
+        madtrex installed, no classic-only directives in reweight_card.dat)"."""
+        reasons = ', '.join(r.replace('{card}', card_name) for r in self.reasons)
+        return 'Reweighting with %s (%s%s)' % (
+            self.backend, self.source, ': %s' % reasons if reasons else '')
+
+
+def _madtrex_card_vocabulary():
+    """The MadtRex-only reweight-card names, from the MadtRex card parsers
+    (pure Python modules, importing neither madspace nor madtrex):
+    ("change" options, "set NAME VALUE" names, the PDF-reweighting ones
+    among both)."""
+    from madgraph.iolibs.template_files.mg7 import param_reweight_card as param
+    from madgraph.iolibs.template_files.mg7 import pdf_reweight_card as pdf
+    old_driver = set(param.DRIVER_OPTION_NAMES) - {'helicity_reweighting', 'use_helicity'}
+    change = {'fptype'} | old_driver | set(pdf.CARD_OPTION_NAMES)
+    # the threading options never had a "set" spelling
+    set_names = (set(pdf.PDF_SET_NAMES) | set(pdf.SCALE_NAMES)
+                 | (change - param.THREAD_OPTION_NAMES)
+                 | {'helicity_reweighting', 'use_helicity'})
+    pdf_only = set(pdf.PDF_SET_NAMES) | set(pdf.SCALE_NAMES) | {'alpha_s_power',
+                                                                 'scale_variation_output'}
+    return frozenset(change), frozenset(set_names), frozenset(pdf_only)
+
+
+def _classic_change_options():
+    """("change" options only the classic module understands, the density
+    ones among them)."""
+    from madgraph.interface.reweight_interface import (REWEIGHT_CHANGE_OPTIONS,
+                                                       DENSITY_CHANGE_OPTIONS)
+    return (frozenset(REWEIGHT_CHANGE_OPTIONS | DENSITY_CHANGE_OPTIONS) - SHARED_CHANGE_OPTIONS,
+            frozenset(DENSITY_CHANGE_OPTIONS))
+
+
+def _check_backend_value(value, where):
+    value = (value or '').lower()
+    if value not in BACKENDS:
+        raise ReweightBackendError(
+            "%s: unknown reweighting backend %r; expected one of %s"
+            % (where, value, ', '.join(BACKENDS)))
+    return value
+
+
+def classify_card(text):
+    """CardTraits of the reweight card text: its "change backend" line, the
+    lines only the classic module or only MadtRex understands, and which
+    MadtRex card reader it needs. Lines neither backend knows are left to the
+    backend that runs the card. Raises ReweightBackendError for an unknown
+    or contradictory "change backend"."""
+    madtrex_change, madtrex_set, pdf_names = _madtrex_card_vocabulary()
+    classic_change, density_change = _classic_change_options()
+    traits = CardTraits()
+    has_mode_line = False
+    has_pdf_names = False
+
+    for number, raw in enumerate(text.splitlines(), 1):
+        line = strip_comment(raw).strip()
+        if not line:
+            continue
+        words = line.split()
+        keyword = words[0].lower()
+
+        if keyword == 'launch':
+            traits.launches += 1
+            continue
+
+        if keyword == 'change' and len(words) > 1:
+            name = words[1].lower()
+            if name == 'backend':
+                value = _check_backend_value(words[2] if len(words) > 2 else '',
+                                             'reweight card line %d' % number)
+                if traits.backend is not None and traits.backend != value:
+                    raise ReweightBackendError(
+                        "reweight card asks for two backends: '%s' (line %d) and "
+                        "'%s' (line %d)" % (traits.backend, traits.backend_line,
+                                            value, number))
+                traits.backend, traits.backend_line = value, number
+            elif name == 'mode':
+                has_mode_line = True
+                value = words[2].lower() if len(words) > 2 else ''
+                if value in CLASSIC_MODES:
+                    traits.classic_only.append(Directive(
+                        number, line, 'NLO reweighting needs the classic module'))
+                elif value in MADTREX_MODES:
+                    traits.madtrex_only.append(Directive(
+                        number, line, 'PDF reweighting needs MadtRex'))
+                    has_pdf_names = has_pdf_names or value == 'pdf'
+            elif name == 'process' and words[-1] == '--add':
+                traits.classic_only.append(Directive(
+                    number, line, 'several processes per model need the classic module'))
+            elif name in classic_change:
+                if name in density_change:
+                    traits.density = True
+                    why = 'density-matrix reweighting needs the classic module'
+                else:
+                    why = "'change %s' is a classic reweighting option" % name
+                traits.classic_only.append(Directive(number, line, why))
+            elif name in madtrex_change:
+                traits.madtrex_only.append(Directive(
+                    number, line, "'change %s' is a MadtRex option" % name))
+                has_pdf_names = has_pdf_names or name in pdf_names
+            continue
+
+        if keyword == 'set' and len(words) == 3:
+            # "set NAME VALUE"; "set BLOCK ID VALUE" has four words
+            name = words[1].lower()
+            if name == 'mode':
+                has_mode_line = True
+                traits.madtrex_only.append(Directive(
+                    number, line, "'set mode' selects a MadtRex card section"))
+            elif name in madtrex_set:
+                why = ('a PDF/scale reweighting setting (MadtRex)' if name in pdf_names
+                       else "'set %s' is a MadtRex option" % name)
+                traits.madtrex_only.append(Directive(number, line, why))
+                has_pdf_names = has_pdf_names or name in pdf_names
+            continue
+
+        if keyword not in ('set', 'change', 'launch', 'import') and \
+                (os.sep in words[0] or words[0].endswith('.dat')):
+            traits.classic_only.append(Directive(
+                number, line, 'a param_card given by path needs the classic module'))
+
+    if has_mode_line:
+        traits.madtrex_kind = 'generic'
+    elif has_pdf_names:
+        traits.madtrex_kind = 'pdf'
+    return traits
+
+
+def strip_comment(line):
+    """line without its comment ("#", or "!" at the start or after
+    whitespace), as the MadtRex card readers strip it
+    (generic_reweight_card.strip_comment)."""
+    from madgraph.iolibs.template_files.mg7 import generic_reweight_card
+    return generic_reweight_card.strip_comment(line)
+
+
+def strip_backend_line(text):
+    """text with its "change backend" lines commented out (the backends'
+    own card readers need not know it; the line numbers stay the same)."""
+    lines = text.splitlines(True)
+    for i, line in enumerate(lines):
+        if _BACKEND_RE.match(line):
+            lines[i] = '# ' + line
+    return ''.join(lines)
+
+
+def _read_output_format(me_dir):
+    """[run] output_format of Cards/run_card.toml, or None."""
+    path = os.path.join(me_dir, 'Cards', 'run_card.toml')
+    try:
+        import tomllib
+        with open(path, 'rb') as f:
+            card = tomllib.load(f)
+    except Exception:
+        # tomllib is Python 3.11+; fall back to the one line we need
+        try:
+            with open(path) as f:
+                text = f.read()
+        except OSError:
+            return None
+        match = re.search(r'''^\s*output_format\s*=\s*["']([^"']*)["']''', text, re.M)
+        return match.group(1) if match else None
+    value = card.get('run', {}).get('output_format')
+    if value is None:
+        # MadGraph always writes it out; RunCardMG7's default otherwise
+        return 'lhe_npy'
+    return value
+
+
+def output_type(me_dir):
+    """"mg7", "madevent", "aMC@NLO" or None for the process directory me_dir."""
+    if os.path.exists(os.path.join(me_dir, 'Cards', 'run_card.toml')):
+        return 'mg7'
+    if os.path.exists(os.path.join(me_dir, 'bin', 'internal', 'amcatnlo_run_interface.py')) or \
+            os.path.exists(os.path.join(me_dir, 'Cards', 'FKS_params.dat')):
+        return 'aMC@NLO'
+    if os.path.exists(os.path.join(me_dir, 'bin', 'madevent')):
+        return 'madevent'
+    return None
+
+
+def madtrex_is_installed(install_dir=None):
+    """MadtRex is installed (madtrex/install/madtrex exists); nothing is
+    imported."""
+    install_dir = Path(install_dir) if install_dir else _MG_ROOT / 'madtrex' / 'install'
+    return (install_dir / 'madtrex').is_dir()
+
+
+def madtrex_build_problem(install_dir=None, madspace_dir=None):
+    """Why the installed madtrex cannot be used with the madspace MadGraph
+    uses (it was built against other madspace headers), or ''. Reads
+    madtrex/_compat.py directly: nothing compiled is imported."""
+    install_dir = Path(install_dir) if install_dir else _MG_ROOT / 'madtrex' / 'install'
+    madspace_dir = (Path(madspace_dir) if madspace_dir
+                    else _MG_ROOT / 'madspace' / 'install' / 'madspace')
+    compat_path = install_dir / 'madtrex' / '_compat.py'
+    if not compat_path.is_file() or not madspace_dir.is_dir():
+        return ''
+    import importlib.util
+    spec = importlib.util.spec_from_file_location('_madtrex_compat', compat_path)
+    compat = importlib.util.module_from_spec(spec)
+    try:
+        spec.loader.exec_module(compat)
+        return compat.build_problem(str(install_dir / 'madtrex'), str(madspace_dir)) or ''
+    except Exception:
+        return ''
+
+
+def probe_environment(me_dir, options=None, madtrex_install_dir=None, has_f2py=None):
+    """Env of the process directory me_dir. options are the MadGraph run
+    options (mg5_path, f2py_compiler); has_f2py overrides the f2py lookup
+    (misc.has_f2py)."""
+    options = options or {}
+    kind = output_type(me_dir)
+    env = Env(output_type=kind, nlo_events=(kind == 'aMC@NLO'))
+    if kind == 'mg7':
+        env.output_format = _read_output_format(me_dir)
+
+    # as AskRun.check_available_module: MadGraph and f2py. This module is
+    # part of MadGraph, so MadGraph is there unless options say otherwise.
+    mg5_path = options.get('mg5_path', str(_MG_ROOT))
+    if has_f2py is None:
+        from madgraph.various import misc
+        has_f2py = misc.has_f2py() or bool(options.get('f2py_compiler'))
+    if not mg5_path:
+        env.classic_available = False
+        env.classic_reason = 'the classic module needs MadGraph (mg5_path is not set)'
+    elif not has_f2py:
+        env.classic_available = False
+        env.classic_reason = 'the classic module needs f2py, which was not found'
+    env.madtrex_installed = madtrex_is_installed(madtrex_install_dir)
+    if env.madtrex_installed:
+        env.madtrex_problem = madtrex_build_problem(madtrex_install_dir)
+    return env
+
+
+def _madtrex_problems(env):
+    """Why MadtRex cannot reweight these events (empty when it can)."""
+    problems = []
+    if env.output_type != 'mg7':
+        problems.append('MadtRex needs an MG7 output (this is %s)'
+                        % (env.output_type or 'not a known output'))
+    if env.nlo_events:
+        problems.append('MadtRex does not reweight NLO events')
+    if env.output_format == 'compact_npy':
+        problems.append("MadtRex needs run_card [run] output_format 'lhe' or 'lhe_npy': "
+                        "'compact_npy' drops the flavour/helicity information it matches "
+                        "events with")
+    if not env.madtrex_installed:
+        problems.append('MadtRex is not installed (python madtrex/install.py, or '
+                        "'install madtrex' in MadGraph)")
+    elif env.madtrex_problem:
+        problems.append(env.madtrex_problem)
+    return problems
+
+
+def _resolve_request(requested, traits):
+    """(backend, source, warnings) from the explicit requests, highest
+    precedence first, then the card's "change backend"; backend is "auto"
+    when nothing explicit was asked."""
+    if requested is None:
+        requested = []
+    elif isinstance(requested, str):
+        requested = [('requested', requested)]
+    explicit = []
+    for source, value in requested:
+        if value is None:
+            continue
+        value = _check_backend_value(value, source)
+        if value != 'auto':
+            explicit.append((source, value))
+    if traits.backend not in (None, 'auto'):
+        explicit.append(("'change backend' (reweight_card.dat line %d)" % traits.backend_line,
+                         traits.backend))
+    if not explicit:
+        return 'auto', 'auto', []
+    source, value = explicit[0]
+    warnings = ['%s asks for the %s backend but %s asks for %s: using %s'
+                % (other, other_value, source, value, value)
+                for other, other_value in explicit[1:] if other_value != value]
+    return value, source, warnings
+
+
+def _classic_kind(traits):
+    return 'density' if traits.density else 'classic'
+
+
+def select_backend(requested, traits, env, plugin=None):
+    """The Decision for reweighting the card described by traits in env.
+
+    requested is "classic"/"madtrex"/"auto"/None, or a list of (source,
+    value) pairs, highest precedence first (the --reweight-backend option,
+    then the launch-question switch); the card's own "change backend" comes
+    after them. When explicit requests disagree, the highest one wins and
+    Decision.warnings names both. plugin is a reweighting plugin selected in
+    the launch question (classic only).
+
+    Raises ReweightBackendError listing every reason when the card cannot
+    be run: a forced backend that cannot run it, a card mixing classic-only
+    and MadtRex-only directives, or no usable backend."""
+    decision = _select(requested, traits, env, plugin)
+    decision.plugin = plugin
+    return decision
+
+
+def _select(requested, traits, env, plugin):
+    backend, source, warnings = _resolve_request(requested, traits)
+
+    if backend == 'classic':
+        problems = []
+        if not env.classic_available:
+            problems.append(env.classic_reason or 'the classic module is not available')
+        problems += ['%s: not understood by the classic module' % d for d in traits.madtrex_only]
+        if problems:
+            raise ReweightBackendError('cannot reweight with the classic module (%s)' % source,
+                                       problems)
+        return Decision('classic', _classic_kind(traits), [], warnings, source)
+
+    if backend == 'madtrex':
+        problems = _madtrex_problems(env)
+        if plugin:
+            problems.append('reweighting plugin %r needs the classic module' % plugin)
+        problems += ['%s: needs the classic module' % d for d in traits.classic_only]
+        if problems:
+            raise ReweightBackendError('cannot reweight with MadtRex (%s)' % source, problems)
+        return Decision('madtrex', traits.madtrex_kind, [], warnings, source)
+
+    # auto
+    if traits.classic_only and traits.madtrex_only:
+        raise ReweightBackendError(
+            'the reweight card mixes directives only the classic module understands '
+            'with directives only MadtRex understands',
+            ['classic only, %s' % d for d in traits.classic_only]
+            + ['MadtRex only, %s' % d for d in traits.madtrex_only])
+
+    if traits.classic_only or plugin:
+        if plugin:
+            why = 'reweighting plugin %r' % plugin
+        elif traits.density:
+            why = 'density-matrix reweighting in {card}'
+        else:
+            why = 'classic-only directives in {card} (line %s)' % ', '.join(
+                str(d.line) for d in traits.classic_only)
+        if not env.classic_available:
+            raise ReweightBackendError('cannot reweight: %s needs the classic module'
+                                       % why.replace('{card}', 'reweight_card.dat'),
+                                       [env.classic_reason])
+        return Decision('classic', _classic_kind(traits), [why], warnings)
+
+    if traits.madtrex_only:
+        problems = _madtrex_problems(env)
+        if problems:
+            raise ReweightBackendError(
+                'cannot reweight: reweight_card.dat has MadtRex-only directives (line %s)'
+                % ', '.join(str(d.line) for d in traits.madtrex_only), problems)
+        return Decision('madtrex', traits.madtrex_kind,
+                        ['MadtRex-only directives in {card}'], warnings)
+
+    madtrex_problems = _madtrex_problems(env)
+    if env.output_type != 'mg7' or env.nlo_events:
+        reasons = [p for p in madtrex_problems if 'MG7' in p or 'NLO' in p]
+        if not env.classic_available:
+            raise ReweightBackendError('cannot reweight: no backend can run here',
+                                       [env.classic_reason] + reasons)
+        return Decision('classic', _classic_kind(traits), reasons, warnings)
+
+    if not madtrex_problems:
+        return Decision('madtrex', traits.madtrex_kind,
+                        ['MG7 output', 'madtrex installed',
+                         'no classic-only directives in {card}'], warnings)
+
+    if env.classic_available:
+        return Decision('classic', _classic_kind(traits),
+                        ['MadtRex unusable: %s' % '; '.join(madtrex_problems)], warnings)
+
+    raise ReweightBackendError('cannot reweight: no backend can run here',
+                               [env.classic_reason] + madtrex_problems)
+
+
+def available_backends(env, plugin=None):
+    """The backends that could reweight in env for a card with no
+    backend-specific directives: a subset of ("classic", "madtrex")."""
+    out = []
+    if env.classic_available:
+        out.append('classic')
+    if not _madtrex_problems(env) and not plugin:
+        out.append('madtrex')
+    return out
diff --git a/madmatrix/MadtRex/makefiles/cudacpp_driver.mk b/madmatrix/MadtRex/makefiles/cudacpp_driver.mk
deleted file mode 100644
index 1b0395f144..0000000000
--- a/madmatrix/MadtRex/makefiles/cudacpp_driver.mk
+++ /dev/null
@@ -1,895 +0,0 @@
-# Copyright (C) 2020-2026 CERN and UCLouvain.
-# Licensed under the GNU Lesser General Public License (version 3 or later).
-# Created originally by: S. Roiser (Feb 2020) for the MadGraph7 CUDACPP plugin.
-# Further modified by: S. Hageboeck, O. Mattelaer, S. Roiser, J. Teig, A. Valassi, Z. Wettersten (2020-2024).
-
-#=== Determine the name of this makefile (https://ftp.gnu.org/old-gnu/Manuals/make-3.80/html_node/make_17.html)
-#=== NB: use ':=' to ensure that the value of CUDACPP_MAKEFILE is not modified further down after including make_opts
-#=== NB: use 'override' to ensure that the value can not be modified from the outside
-override CUDACPP_MAKEFILE := $(word $(words $(MAKEFILE_LIST)),$(MAKEFILE_LIST))
-###$(info CUDACPP_MAKEFILE='$(CUDACPP_MAKEFILE)')
-
-#=== NB: different names (e.g. cudacpp.mk and cudacpp_src.mk) are used in the Subprocess and src directories
-override CUDACPP_SRC_MAKEFILE = cudacpp_src.mk
-
-#-------------------------------------------------------------------------------
-
-#=== Include cudacpp_config.mk
-
-# Check that the user-defined choices of BACKEND, FPTYPE, HELINL, HRDCOD are supported (and configure defaults if no user-defined choices exist)
-# Stop with an error if BACKEND=cuda and nvcc is missing or if BACKEND=hip and hipcc is missing.
-# Determine CUDACPP_BUILDDIR from a DIRTAG based on BACKEND, FPTYPE, HELINL, HRDCOD and from the user-defined choice of USEBUILDDIR
-include ../src/cudacpp_config.mk
-
-# Export CUDACPP_BUILDDIR (so that there is no need to check/define it again in cudacpp_src.mk)
-export CUDACPP_BUILDDIR
-
-#-------------------------------------------------------------------------------
-
-#=== Use bash in the Makefile (https://www.gnu.org/software/make/manual/html_node/Choosing-the-Shell.html)
-
-SHELL := /bin/bash
-
-#-------------------------------------------------------------------------------
-
-#=== Detect O/S and architecture (assuming uname is available, https://en.wikipedia.org/wiki/Uname)
-
-# Detect O/S kernel (Linux, Darwin...)
-UNAME_S := $(shell uname -s)
-###$(info UNAME_S='$(UNAME_S)')
-
-# Detect architecture (x86_64, ppc64le...)
-UNAME_P := $(shell uname -p)
-###$(info UNAME_P='$(UNAME_P)')
-UNAME_M := $(shell uname -m)
-
-#-------------------------------------------------------------------------------
-
-#=== Include the common MadGraph7 Makefile options
-
-# OM: including make_opts is crucial for MadGraph7 flag consistency/documentation
-# AV: disable the inclusion of make_opts if the file has not been generated (standalone cudacpp)
-ifneq ($(wildcard ../Source/make_opts),)
-  include ../Source/make_opts
-endif
-
-#-------------------------------------------------------------------------------
-
-#=== Redefine BACKEND if the current value is 'auto'
-
-# Set the BACKEND choice corresponding to 'auto' (the 'best' C++ vectorization available)
-ifeq ($(BACKEND),auto)
-  ifeq ($(UNAME_P),ppc64le)
-    override BACKEND = simd_128
-  else ifneq (,$(filter $(UNAME_M),arm64 aarch64))
-    override BACKEND = simd_128
-  else ifeq ($(wildcard /proc/cpuinfo),)
-    override BACKEND = scalar
-    ###$(warning Using BACKEND='$(BACKEND)' because host SIMD features cannot be read from /proc/cpuinfo)
-  else ifeq ($(shell grep -m1 -c avx512vl /proc/cpuinfo)$(shell $(CXX) --version | grep ^clang),1)
-    override BACKEND = avx512y
-  else ifeq ($(shell grep -m1 -c avx2 /proc/cpuinfo),1)
-    override BACKEND = simd_256
-    ###ifneq ($(shell grep -m1 -c avx512vl /proc/cpuinfo),1)
-    ###  $(warning Using BACKEND='$(BACKEND)' because host does not support avx512vl)
-    ###else
-    ###  $(warning Using BACKEND='$(BACKEND)' because this is faster than avx512vl for clang)
-    ###endif
-  else ifeq ($(shell grep -m1 -c sse4_2 /proc/cpuinfo),1)
-    override BACKEND = simd_128
-  else
-    override BACKEND = scalar
-  endif
-  $(info BACKEND=$(BACKEND) (was auto))
-else
-  $(info BACKEND='$(BACKEND)')
-endif
-
-#-------------------------------------------------------------------------------
-
-#=== Configure the C++ compiler
-
-CXXFLAGS = $(OPTFLAGS) -std=c++17 -Wall -Wshadow -Wextra
-ifeq ($(shell $(CXX) --version | grep ^nvc++),)
-  CXXFLAGS += -ffast-math # see issue #117
-endif
-###CXXFLAGS+= -Ofast # performance is not different from --fast-math
-###CXXFLAGS+= -g # FOR DEBUGGING ONLY
-
-# Optionally add debug flags to display the full list of flags (eg on Darwin)
-###CXXFLAGS+= -v
-
-# Note: AR, CXX and FC are implicitly defined if not set externally
-# See https://www.gnu.org/software/make/manual/html_node/Implicit-Variables.html
-
-# Add -mmacosx-version-min=11.3 to avoid "ld: warning: object file was built for newer macOS version than being linked"
-ifneq ($(shell $(CXX) --version | egrep '^Apple clang'),)
-  CXXFLAGS += -mmacosx-version-min=11.3
-endif
-
-# Export CXXFLAGS (so that there is no need to check/define it again in cudacpp_src.mk)
-export CXXFLAGS
-
-#-------------------------------------------------------------------------------
-
-#=== Configure the GPU compiler (CUDA or HIP)
-#=== (note, this is done also for C++, as NVTX and CURAND/ROCRAND are also needed by the C++ backends)
-
-# Set CUDA_HOME from the path to nvcc, if it exists
-#override CUDA_HOME = $(patsubst %%/bin/nvcc,%%,$(shell which nvcc 2>/dev/null))
-CUDA_HOME := $(patsubst %/bin/nvcc,%,$(shell which nvcc 2>/dev/null))
-
-# Set HIP_HOME from the path to hipcc, if it exists
-override HIP_HOME = $(patsubst %%/bin/hipcc,%%,$(shell which hipcc 2>/dev/null))
-
-# Configure CUDA_INC (for CURAND and NVTX) and NVTX if a CUDA installation exists
-# (FIXME? Is there any equivalent of NVTX FOR HIP? What should be configured if both CUDA and HIP are installed?)
-ifneq ($(CUDA_HOME),)
-  USE_NVTX ?=-DUSE_NVTX
-  CUDA_INC = -I$(CUDA_HOME)/include/
-else
-  override USE_NVTX=
-  override CUDA_INC=
-endif
-
-# NB: NEW LOGIC FOR ENABLING AND DISABLING CUDA OR HIP BUILDS (AV Feb-Mar 2024)
-# - In the old implementation, by default the C++ targets for one specific AVX were always built together with either CUDA or HIP.
-# If both CUDA and HIP were installed, then CUDA took precedence over HIP, and the only way to force HIP builds was to disable
-# CUDA builds by setting CUDA_HOME to an invalid value (as CUDA_HOME took precdence over PATH to find the installation of nvcc).
-# Similarly, C++-only builds could be forced by setting CUDA_HOME and/or HIP_HOME to invalid values. A check for an invalid nvcc
-# in CUDA_HOME or an invalid hipcc HIP_HOME was necessary to ensure this logic, and had to be performed at the very beginning.
-# - In the new implementation (PR #798), separate individual builds are performed for one specific C++/AVX mode, for CUDA or
-# for HIP. The choice of the type of build is taken depending on the value of the BACKEND variable (replacing the AVX variable).
-# Unlike what happened in the past, nvcc and hipcc must have already been added to PATH. Using 'which nvcc' and 'which hipcc',
-# their existence and their location is checked, and the variables CUDA_HOME and HIP_HOME are internally set by this makefile.
-# This must be still done before backend-specific customizations, e.g. because CURAND and NVTX are also used in C++ builds.
-# Note also that a preliminary check for nvcc and hipcc if BACKEND is cuda or hip is performed in cudacpp_config.mk.
-# - Note also that the REQUIRE_CUDA variable (which was used in the past, e.g. for CI tests on GPU #443) is now (PR #798) no
-# longer necessary, as it is now equivalent to BACKEND=cuda. Similarly, there is no need to introduce a REQUIRE_HIP variable.
-
-#=== Configure the CUDA or HIP compiler (only for the CUDA and HIP backends)
-#=== (NB: throughout all makefiles, an empty GPUCC is used to indicate that this is a C++ build, i.e. that BACKEND is neither cuda nor hip!)
-
-ifeq ($(BACKEND),cuda)
-
-  # If CXX is not a single word (example "clang++ --gcc-toolchain...") then disable CUDA builds (issue #505)
-  # This is because it is impossible to pass this to "GPUFLAGS += -ccbin <host-compiler>" below
-  ifneq ($(words $(subst ccache ,,$(CXX))),1) # allow at most "CXX=ccache <host-compiler>" from outside
-    $(error BACKEND=$(BACKEND) but CUDA builds are not supported for multi-word CXX "$(CXX)")
-  endif
-
-  # Set GPUCC as $(CUDA_HOME)/bin/nvcc (it was already checked above that this exists)
-  GPUCC = $(CUDA_HOME)/bin/nvcc
-  XCOMPILERFLAG = -Xcompiler
-  GPULANGUAGE = cu
-  GPUSUFFIX = cuda
-
-  # Basic compiler flags (optimization and includes)
-  GPUFLAGS = $(foreach opt, $(OPTFLAGS), $(XCOMPILERFLAG) $(opt))
-
-  # NVidia CUDA architecture flags
-  # See https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html
-  # See https://arnon.dk/matching-sm-architectures-arch-and-gencode-for-various-nvidia-cards/
-  # Default: use compute capability 70 for V100 (CERN lxbatch, CERN itscrd, Juwels Cluster).
-  # This will embed device code for 70, and PTX for 70+.
-  # One may pass MADGRAPH_CUDA_ARCHITECTURE (comma-separated list) to the make command to use another value or list of values (see #533).
-  # Examples: use 60 for P100 (Piz Daint), 80 for A100 (Juwels Booster, NVidia raplab/Curiosity).
-  MADGRAPH_CUDA_ARCHITECTURE ?= 70
-  ###GPUARCHFLAGS = -gencode arch=compute_$(MADGRAPH_CUDA_ARCHITECTURE),code=compute_$(MADGRAPH_CUDA_ARCHITECTURE) -gencode arch=compute_$(MADGRAPH_CUDA_ARCHITECTURE),code=sm_$(MADGRAPH_CUDA_ARCHITECTURE) # Older implementation (AV): go back to this one for multi-GPU support #533
-  ###GPUARCHFLAGS = --gpu-architecture=compute_$(MADGRAPH_CUDA_ARCHITECTURE) --gpu-code=sm_$(MADGRAPH_CUDA_ARCHITECTURE),compute_$(MADGRAPH_CUDA_ARCHITECTURE)  # Newer implementation (SH): cannot use this as-is for multi-GPU support #533
-  comma:=,
-  GPUARCHFLAGS = $(foreach arch,$(subst $(comma), ,$(MADGRAPH_CUDA_ARCHITECTURE)),-gencode arch=compute_$(arch),code=compute_$(arch) -gencode arch=compute_$(arch),code=sm_$(arch))
-  GPUFLAGS += $(GPUARCHFLAGS)
-
-  # Other NVidia-specific flags
-  CUDA_OPTFLAGS = -lineinfo
-  GPUFLAGS += $(CUDA_OPTFLAGS)
-
-  # NVCC version
-  ###GPUCC_VERSION = $(shell $(GPUCC) --version | grep 'Cuda compilation tools' | cut -d' ' -f5 | cut -d, -f1)
-
-  # Fast math
-  GPUFLAGS += -use_fast_math
-
-  # Extra build warnings
-  ###GPUFLAGS += $(XCOMPILERFLAG) -Wall $(XCOMPILERFLAG) -Wextra $(XCOMPILERFLAG) -Wshadow
-
-  # CUDA includes and NVTX
-  GPUFLAGS += $(CUDA_INC) $(USE_NVTX) 
-
-  # C++ standard
-  GPUFLAGS += -std=c++17 # need CUDA >= 11.2 (see #333): this is enforced in mgOnGpuConfig.h
-
-  # For nvcc, use -maxrregcount to control the maximum number of registries (this does not exist in hipcc)
-  # Without -maxrregcount: baseline throughput: 6.5E8 (16384 32 12) up to 7.3E8 (65536 128 12)
-  ###GPUFLAGS+= --maxrregcount 160 # improves throughput: 6.9E8 (16384 32 12) up to 7.7E8 (65536 128 12)
-  ###GPUFLAGS+= --maxrregcount 128 # improves throughput: 7.3E8 (16384 32 12) up to 7.6E8 (65536 128 12)
-  ###GPUFLAGS+= --maxrregcount 96 # degrades throughput: 4.1E8 (16384 32 12) up to 4.5E8 (65536 128 12)
-  ###GPUFLAGS+= --maxrregcount 64 # degrades throughput: 1.7E8 (16384 32 12) flat at 1.7E8 (65536 128 12)
-
-  # Set the host C++ compiler for nvcc via "-ccbin <host-compiler>"
-  # (NB issue #505: this must be a single word, "clang++ --gcc-toolchain..." is not supported)
-  GPUFLAGS += -ccbin $(shell which $(subst ccache ,,$(CXX)))
-
-  # Allow newer (unsupported) C++ compilers with older versions of CUDA if ALLOW_UNSUPPORTED_COMPILER_IN_CUDA is set (#504)
-  ifneq ($(origin ALLOW_UNSUPPORTED_COMPILER_IN_CUDA),undefined)
-    GPUFLAGS += -allow-unsupported-compiler
-  endif
-
-else ifeq ($(BACKEND),hip)
-
-  # Set GPUCC as $(HIP_HOME)/bin/hipcc (it was already checked above that this exists)
-  GPUCC = $(HIP_HOME)/bin/hipcc
-  XCOMPILERFLAG =
-  GPULANGUAGE = hip
-  GPUSUFFIX = hip
-
-  # Basic compiler flags (optimization and includes)
-  GPUFLAGS = $(foreach opt, $(OPTFLAGS), $(XCOMPILERFLAG) $(opt))
-
-  # AMD HIP architecture flags
-  GPUARCHFLAGS = --offload-arch=gfx90a
-  GPUFLAGS += $(GPUARCHFLAGS)
-
-  # Other AMD-specific flags
-  GPUFLAGS += -target x86_64-linux-gnu -DHIP_PLATFORM=amd
-
-  # Fast math (is -DHIP_FAST_MATH equivalent to -ffast-math?)
-  GPUFLAGS += -DHIP_FAST_MATH
-
-  # Extra build warnings
-  ###GPUFLAGS += $(XCOMPILERFLAG) -Wall $(XCOMPILERFLAG) -Wextra $(XCOMPILERFLAG) -Wshadow
-
-  # HIP includes
-  HIP_INC = -I$(HIP_HOME)/include/
-  GPUFLAGS += $(HIP_INC)
-
-  # C++ standard
-  GPUFLAGS += -std=c++17
-
-else
-
-  # Backend is neither cuda nor hip
-  override GPUCC=
-  override GPUFLAGS=
-
-  # Sanity check, this should never happen: if GPUCC is empty, then this is a C++ build, i.e. BACKEND is neither cuda nor hip.
-  # In practice, in the following, "ifeq ($(GPUCC),)" is equivalent to "BACKEND is one of the CPU backends"
-  # (NB: do NOT test the backend name for a 'cpu' substring: none of the CPU backend names contains one).
-  # Conversely, note that GPUFLAGS is non-empty also for C++ builds, but it is never used in that case.
-  ifneq ($(words $(filter $(BACKEND), scalar simd_128 simd_256 avx512y simd_512)),1)
-    $(error INTERNAL ERROR! Unknown backend BACKEND='$(BACKEND)': supported backends are $(foreach backend,$(SUPPORTED_BACKENDS),'$(backend)'))
-  endif
-
-endif
-
-# Export GPUCC, GPUFLAGS, GPULANGUAGE, GPUSUFFIX (these are needed by both src and rwgt_runners, but should not be overwritten there)
-export CUDA_HOME
-export GPUCC
-export GPUFLAGS
-export GPULANGUAGE
-export GPUSUFFIX
-export XCOMPILERFLAG
-
-#-------------------------------------------------------------------------------
-
-#=== Configure ccache for C++ and CUDA/HIP builds
-
-# Enable ccache if USECCACHE=1
-ifeq ($(USECCACHE)$(shell echo $(CXX) | grep ccache),1)
-  override CXX:=ccache $(CXX)
-endif
-#ifeq ($(USECCACHE)$(shell echo $(AR) | grep ccache),1)
-#  override AR:=ccache $(AR)
-#endif
-ifneq ($(GPUCC),)
-  ifeq ($(USECCACHE)$(shell echo $(GPUCC) | grep ccache),1)
-    override GPUCC:=ccache $(GPUCC)
-  endif
-endif
-
-#-------------------------------------------------------------------------------
-
-#=== Configure common compiler flags for C++ and CUDA/HIP
-
-INCFLAGS = -I.
-OPTFLAGS = -O3 # this ends up in GPUFLAGS too (should it?), cannot add -Ofast or -ffast-math here
-
-# Dependency on src directory
-ifeq ($(GPUCC),)
-MG5AMC_COMMONLIB = mg5amc_common_cpp
-else
-MG5AMC_COMMONLIB = mg5amc_common_$(GPUSUFFIX)
-endif
-LIBFLAGS = -L$(LIBDIR) -l$(MG5AMC_COMMONLIB) -lrex -ltearex
-INCFLAGS += -I../src
-
-# Compiler-specific googletest build directory (#125 and #738)
-ifneq ($(shell $(CXX) --version | grep '^Intel(R) oneAPI DPC++/C++ Compiler'),)
-  override CXXNAME = icpx$(shell $(CXX) --version | head -1 | cut -d' ' -f5)
-else ifneq ($(shell $(CXX) --version | egrep '^clang'),)
-  override CXXNAME = clang$(shell $(CXX) --version | head -1 | cut -d' ' -f3)
-else ifneq ($(shell $(CXX) --version | grep '^g++ (GCC)'),)
-  override CXXNAME = gcc$(shell $(CXX) --version | head -1 | cut -d' ' -f3)
-else
-  override CXXNAME = unknown
-endif
-###$(info CXXNAME=$(CXXNAME))
-override CXXNAMESUFFIX = _$(CXXNAME)
-
-# Export CXXNAMESUFFIX (so that there is no need to check/define it again in cudacpp_test.mk)
-export CXXNAMESUFFIX
-
-#-------------------------------------------------------------------------------
-
-#=== Configure PowerPC-specific compiler flags for C++ and CUDA/HIP
-
-# PowerPC-specific CXX compiler flags (being reviewed)
-ifeq ($(UNAME_P),ppc64le)
-  CXXFLAGS+= -mcpu=power9 -mtune=power9 # gains ~2-3%% both for scalar and simd_128
-  # Throughput references without the extra flags below: scalar=1.41-1.42E6, simd_128=2.15-2.19E6
-  ###CXXFLAGS+= -DNO_WARN_X86_INTRINSICS # no change
-  ###CXXFLAGS+= -fpeel-loops # no change
-  ###CXXFLAGS+= -funroll-loops # gains ~1%% for scalar, loses ~1%% for simd_128
-  ###CXXFLAGS+= -ftree-vectorize # no change
-  ###CXXFLAGS+= -flto # would increase to scalar=4.08-4.12E6, simd_128=4.99-5.03E6!
-else
-  ###CXXFLAGS+= -flto # also on Intel this would increase throughputs by a factor 2 to 4...
-  ######CXXFLAGS+= -fno-semantic-interposition # no benefit (neither alone, nor combined with -flto)
-endif
-
-# PowerPC-specific CUDA/HIP compiler flags (to be reviewed!)
-ifeq ($(UNAME_P),ppc64le)
-  GPUFLAGS+= $(XCOMPILERFLAG) -mno-float128
-endif
-
-#-------------------------------------------------------------------------------
-
-#=== Configure defaults for OMPFLAGS
-
-# Disable OpenMP by default: enable OpenMP only if USEOPENMP=1 (#758)
-ifeq ($(USEOPENMP),1)
-  ###$(info USEOPENMP==1: will build with OpenMP if possible)
-  ifneq ($(findstring hipcc,$(GPUCC)),)
-    override OMPFLAGS = # disable OpenMP MT when using hipcc #802
-  else ifneq ($(shell $(CXX) --version | egrep '^Intel'),)
-    override OMPFLAGS = -fopenmp
-    ###override OMPFLAGS = # disable OpenMP MT on Intel (was ok without GPUCC but not ok with GPUCC before #578)
-  else ifneq ($(shell $(CXX) --version | egrep '^clang version 16'),)
-    ###override OMPFLAGS = # disable OpenMP on clang16 #904
-    $(error OpenMP is not supported by cudacpp on clang16 - issue #904)
-  else ifneq ($(shell $(CXX) --version | egrep '^clang version 17'),)
-    ###override OMPFLAGS = # disable OpenMP on clang17 #904
-    $(error OpenMP is not supported by cudacpp on clang17 - issue #904)
-  else ifneq ($(shell $(CXX) --version | egrep '^(clang)'),)
-    override OMPFLAGS = -fopenmp
-    ###override OMPFLAGS = # disable OpenMP MT on clang (was not ok without or with nvcc before #578)
-  ###else ifneq ($(shell $(CXX) --version | egrep '^(Apple clang)'),) # AV for Mac (Apple clang compiler)
-  else ifeq ($(UNAME_S),Darwin) # OM for Mac (any compiler)
-    override OMPFLAGS = # AV disable OpenMP MT on Apple clang (builds fail in the CI #578)
-    ###override OMPFLAGS = -fopenmp # OM re-enable OpenMP MT on Apple clang? (AV Oct 2023: this still fails in the CI)
-  else
-    override OMPFLAGS = -fopenmp # enable OpenMP MT by default on all other platforms
-    ###override OMPFLAGS = # disable OpenMP MT on all other platforms (default before #575)
-  endif
-else
-  ###$(info USEOPENMP!=1: will build without OpenMP)
-  override OMPFLAGS =
-endif
-
-#-------------------------------------------------------------------------------
-
-#=== Configure defaults and check if user-defined choices exist for RNDGEN (legacy!), HASCURAND, HASHIPRAND
-
-# If the legacy RNDGEN exists, this take precedence over any HASCURAND choice (but a warning is printed out)
-###$(info RNDGEN=$(RNDGEN))
-ifneq ($(RNDGEN),)
-  $(warning Environment variable RNDGEN is no longer supported, please use HASCURAND instead!)
-  ifeq ($(RNDGEN),hasCurand)
-    override HASCURAND = $(RNDGEN)
-  else ifeq ($(RNDGEN),hasNoCurand)
-    override HASCURAND = $(RNDGEN)
-  else ifneq ($(RNDGEN),hasNoCurand)
-    $(error Unknown RNDGEN='$(RNDGEN)': only 'hasCurand' and 'hasNoCurand' are supported - but use HASCURAND instead!)
-  endif
-endif
-
-# Set the default HASCURAND (curand random number generator) choice, if no prior choice exists for HASCURAND
-# (NB: allow HASCURAND=hasCurand even if $(GPUCC) does not point to nvcc: assume CUDA_HOME was defined correctly...)
-ifeq ($(HASCURAND),)
-  ifeq ($(GPUCC),) # CPU-only build
-    ifneq ($(CUDA_HOME),)
-      # By default, assume that curand is installed if a CUDA installation exists
-      override HASCURAND = hasCurand
-    else
-      override HASCURAND = hasNoCurand
-    endif
-  else ifeq ($(findstring nvcc,$(GPUCC)),nvcc) # Nvidia GPU build
-    override HASCURAND = hasCurand
-  else # non-Nvidia GPU build
-    override HASCURAND = hasNoCurand
-  endif
-endif
-
-# Set the default HASHIPRAND (hiprand random number generator) choice, if no prior choice exists for HASHIPRAND
-# (NB: allow HASHIPRAND=hasHiprand even if $(GPUCC) does not point to hipcc: assume HIP_HOME was defined correctly...)
-ifeq ($(HASHIPRAND),)
-  ifeq ($(GPUCC),) # CPU-only build
-    override HASHIPRAND = hasNoHiprand
-  else ifeq ($(findstring hipcc,$(GPUCC)),hipcc) # AMD GPU build
-    override HASHIPRAND = hasHiprand
-  else # non-AMD GPU build
-    override HASHIPRAND = hasNoHiprand
-  endif
-endif
-
-#-------------------------------------------------------------------------------
-
-#=== Set the CUDA/HIP/C++ compiler flags appropriate to user-defined choices of AVX, FPTYPE, HELINL, HRDCOD
-
-# Set the build flags appropriate to OMPFLAGS
-$(info OMPFLAGS=$(OMPFLAGS))
-CXXFLAGS += $(OMPFLAGS)
-
-# Set the build flags appropriate to each BACKEND choice (example: "make BACKEND=scalar")
-# [NB MGONGPU_PVW512 is needed because "-mprefer-vector-width=256" is not exposed in a macro]
-# [Use 'g++ <buildflags> -E -dM - < /dev/null' to check which #define's are enabled]
-# [See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96476]
-ifeq ($(UNAME_P),ppc64le)
-  ifeq ($(BACKEND),simd_128)
-    override AVXFLAGS = -D__SSE4_2__ # Power9 VSX with 128 width (VSR registers)
-  else ifeq ($(BACKEND),simd_256)
-    $(error Invalid SIMD BACKEND='$(BACKEND)': only 'scalar' and 'simd_128' are supported on PowerPC for the moment)
-  else ifeq ($(BACKEND),avx512y)
-    $(error Invalid SIMD BACKEND='$(BACKEND)': only 'scalar' and 'simd_128' are supported on PowerPC for the moment)
-  else ifeq ($(BACKEND),simd_512)
-    $(error Invalid SIMD BACKEND='$(BACKEND)': only 'scalar' and 'simd_128' are supported on PowerPC for the moment)
-  endif
-else ifeq ($(UNAME_M),arm64) # ARM on Apple silicon
-  ifeq ($(BACKEND),scalar) # this internally undefines __ARM_NEON
-    override AVXFLAGS = -DMGONGPU_NOARMNEON
-  else ifeq ($(BACKEND),simd_128) # __ARM_NEON is always defined on Apple silicon
-    override AVXFLAGS =
-  else ifeq ($(BACKEND),simd_256)
-    $(error Invalid SIMD BACKEND='$(BACKEND)': only 'scalar' and 'simd_128' are supported on ARM for the moment)
-  else ifeq ($(BACKEND),avx512y)
-    $(error Invalid SIMD BACKEND='$(BACKEND)': only 'scalar' and 'simd_128' are supported on ARM for the moment)
-  else ifeq ($(BACKEND),simd_512)
-    $(error Invalid SIMD BACKEND='$(BACKEND)': only 'scalar' and 'simd_128' are supported on ARM for the moment)
-  endif
-else ifeq ($(UNAME_M),aarch64) # ARM on Linux
-  ifeq ($(BACKEND),scalar) # +nosimd ensures __ARM_NEON is absent
-    override AVXFLAGS = -march=armv8-a+nosimd
-  else ifeq ($(BACKEND),simd_128) # +simd ensures __ARM_NEON is present (128 width Q/quadword registers)
-    override AVXFLAGS = -march=armv8-a+simd
-  else ifeq ($(BACKEND),simd_256)
-    $(error Invalid SIMD BACKEND='$(BACKEND)': only 'scalar' and 'simd_128' are supported on aarch64 for the moment)
-  else ifeq ($(BACKEND),avx512y)
-    $(error Invalid SIMD BACKEND='$(BACKEND)': only 'scalar' and 'simd_128' are supported on aarch64 for the moment)
-  else ifeq ($(BACKEND),simd_512)
-    $(error Invalid SIMD BACKEND='$(BACKEND)': only 'scalar' and 'simd_128' are supported on aarch64 for the moment)
-  endif
-else ifneq ($(shell $(CXX) --version | grep ^nvc++),) # support nvc++ #531
-  ifeq ($(BACKEND),scalar)
-    override AVXFLAGS = -mno-sse3 # no SIMD
-  else ifeq ($(BACKEND),simd_128)
-    override AVXFLAGS = -mno-avx # SSE4.2 with 128 width (xmm registers)
-  else ifeq ($(BACKEND),simd_256)
-    override AVXFLAGS = -march=haswell # AVX2 with 256 width (ymm registers) [DEFAULT for clang]
-  else ifeq ($(BACKEND),avx512y)
-    override AVXFLAGS = -march=skylake -mprefer-vector-width=256 # AVX512 with 256 width (ymm registers) [DEFAULT for gcc]
-  else ifeq ($(BACKEND),simd_512)
-    override AVXFLAGS = -march=skylake -DMGONGPU_PVW512 # AVX512 with 512 width (zmm registers)
-  endif
-else
-  ifeq ($(BACKEND),scalar)
-    override AVXFLAGS = -march=x86-64 # no SIMD (see #588)
-  else ifeq ($(BACKEND),simd_128)
-    override AVXFLAGS = -march=nehalem # SSE4.2 with 128 width (xmm registers)
-  else ifeq ($(BACKEND),simd_256)
-    override AVXFLAGS = -march=haswell # AVX2 with 256 width (ymm registers) [DEFAULT for clang]
-  else ifeq ($(BACKEND),avx512y)
-    override AVXFLAGS = -march=skylake-avx512 -mprefer-vector-width=256 # AVX512 with 256 width (ymm registers) [DEFAULT for gcc]
-  else ifeq ($(BACKEND),simd_512)
-    override AVXFLAGS = -march=skylake-avx512 -DMGONGPU_PVW512 # AVX512 with 512 width (zmm registers)
-  endif
-endif
-# For the moment, use AVXFLAGS everywhere (in C++ builds): eventually, use them only in encapsulated implementations?
-ifeq ($(GPUCC),)
-  CXXFLAGS+= $(AVXFLAGS)
-endif
-
-# Set the build flags appropriate to each FPTYPE choice (example: "make FPTYPE=f")
-$(info FPTYPE='$(FPTYPE)')
-ifeq ($(FPTYPE),d)
-  CXXFLAGS += -DMGONGPU_FPTYPE_DOUBLE -DMGONGPU_FPTYPE2_DOUBLE
-  GPUFLAGS += -DMGONGPU_FPTYPE_DOUBLE -DMGONGPU_FPTYPE2_DOUBLE
-else ifeq ($(FPTYPE),f)
-  CXXFLAGS += -DMGONGPU_FPTYPE_FLOAT -DMGONGPU_FPTYPE2_FLOAT
-  GPUFLAGS += -DMGONGPU_FPTYPE_FLOAT -DMGONGPU_FPTYPE2_FLOAT
-else ifeq ($(FPTYPE),m)
-  CXXFLAGS += -DMGONGPU_FPTYPE_DOUBLE -DMGONGPU_FPTYPE2_FLOAT
-  GPUFLAGS += -DMGONGPU_FPTYPE_DOUBLE -DMGONGPU_FPTYPE2_FLOAT
-else
-  $(error Unknown FPTYPE='$(FPTYPE)': only 'd', 'f' and 'm' are supported)
-endif
-
-# Set the build flags appropriate to each HELINL choice (example: "make HELINL=1")
-$(info HELINL='$(HELINL)')
-ifeq ($(HELINL),1)
-  CXXFLAGS += -DMGONGPU_INLINE_HELAMPS
-  GPUFLAGS += -DMGONGPU_INLINE_HELAMPS
-else ifneq ($(HELINL),0)
-  $(error Unknown HELINL='$(HELINL)': only '0' and '1' are supported)
-endif
-
-# Set the build flags appropriate to each HRDCOD choice (example: "make HRDCOD=1")
-$(info HRDCOD='$(HRDCOD)')
-ifeq ($(HRDCOD),1)
-  CXXFLAGS += -DMGONGPU_HARDCODE_PARAM
-  GPUFLAGS += -DMGONGPU_HARDCODE_PARAM
-else ifneq ($(HRDCOD),0)
-  $(error Unknown HRDCOD='$(HRDCOD)': only '0' and '1' are supported)
-endif
-
-#=== Set the CUDA/HIP/C++ compiler and linker flags appropriate to user-defined choices of HASCURAND, HASHIPRAND
-
-$(info HASCURAND=$(HASCURAND))
-$(info HASHIPRAND=$(HASHIPRAND))
-override RNDCXXFLAGS=
-override RNDLIBFLAGS=
-
-# Set the RNDCXXFLAGS and RNDLIBFLAGS build flags appropriate to each HASCURAND choice (example: "make HASCURAND=hasNoCurand")
-ifeq ($(HASCURAND),hasNoCurand)
-  override RNDCXXFLAGS += -DMGONGPU_HAS_NO_CURAND
-else ifeq ($(HASCURAND),hasCurand)
-  override RNDLIBFLAGS += -L$(CUDA_HOME)/lib64/ -lcurand # NB: -lcuda is not needed here!
-else
-  $(error Unknown HASCURAND='$(HASCURAND)': only 'hasCurand' and 'hasNoCurand' are supported)
-endif
-
-# Set the RNDCXXFLAGS and RNDLIBFLAGS build flags appropriate to each HASHIPRAND choice (example: "make HASHIPRAND=hasNoHiprand")
-ifeq ($(HASHIPRAND),hasNoHiprand)
-  override RNDCXXFLAGS += -DMGONGPU_HAS_NO_HIPRAND
-else ifeq ($(HASHIPRAND),hasHiprand)
-  override RNDLIBFLAGS += -L$(HIP_HOME)/lib/ -lhiprand
-else ifneq ($(HASHIPRAND),hasHiprand)
-  $(error Unknown HASHIPRAND='$(HASHIPRAND)': only 'hasHiprand' and 'hasNoHiprand' are supported)
-endif
-
-#$(info RNDCXXFLAGS=$(RNDCXXFLAGS))
-#$(info RNDLIBFLAGS=$(RNDLIBFLAGS))
-
-#-------------------------------------------------------------------------------
-
-#=== Configure Position-Independent Code
-CXXFLAGS += -fPIC
-GPUFLAGS += $(XCOMPILERFLAG) -fPIC
-
-#-------------------------------------------------------------------------------
-
-#=== Configure build directories and build lockfiles ===
-
-# Build lockfile "full" tag (defines full specification of build options that cannot be intermixed)
-# (Rationale: avoid mixing of builds with different random number generators)
-override TAG = $(BACKEND)_$(FPTYPE)_inl$(HELINL)_hrd$(HRDCOD)_$(HASCURAND)_$(HASHIPRAND)
-
-# Export TAG (so that there is no need to check/define it again in cudacpp_src.mk)
-export TAG
-
-# Build directory: current directory by default, or build.$(DIRTAG) if USEBUILDDIR==1
-override BUILDDIR = $(CUDACPP_BUILDDIR)
-ifeq ($(USEBUILDDIR),1)
-  override LIBDIR = ../lib/$(BUILDDIR)
-  override LIBDIRRPATH = '$$ORIGIN/$(LIBDIR)'
-  $(info Building in BUILDDIR=$(BUILDDIR) for tag=$(TAG) (USEBUILDDIR == 1))
-else
-  override LIBDIR = ../lib
-  override LIBDIRRPATH = '$$ORIGIN/$(LIBDIR)'
-  $(info Building in BUILDDIR=$(BUILDDIR) for tag=$(TAG) (USEBUILDDIR != 1))
-endif
-###override INCDIR = ../../include
-###$(info Building in BUILDDIR=$(BUILDDIR) for tag=$(TAG))
-
-# On Linux, set rpath to LIBDIR to make it unnecessary to use LD_LIBRARY_PATH
-# Use relative paths with respect to the executables or shared libraries ($ORIGIN on Linux)
-# On Darwin, building libraries with absolute paths in LIBDIR makes this unnecessary
-ifeq ($(UNAME_S),Darwin)
-  override CXXLIBFLAGSRPATH =
-  override GPULIBFLAGSRPATH =
-  override CXXLIBFLAGSRPATH2 =
-  override GPULIBFLAGSRPATH2 =
-else
-  # RPATH to gpu/cpp libs when linking executables
-  override CXXLIBFLAGSRPATH = -Wl,-rpath=$(LIBDIRRPATH)
-  override GPULIBFLAGSRPATH = -Xlinker -rpath=$(LIBDIRRPATH)
-  # RPATH to common lib when linking gpu/cpp libs
-  override CXXLIBFLAGSRPATH2 = -Wl,-rpath='$$ORIGIN'
-  override GPULIBFLAGSRPATH2 = -Xlinker -rpath='$$ORIGIN'
-endif
-
-# Setting LD_LIBRARY_PATH or DYLD_LIBRARY_PATH in the RUNTIME is no longer necessary (neither on Linux nor on Mac)
-override RUNTIME =
-
-#===============================================================================
-#=== Makefile TARGETS and build rules below
-#===============================================================================
-
-.PHONY: all $(DIRS)
-
-DIRS := $(wildcard P*)
-
-# Construct the library paths
-cxx_proclibs := $(shell for dir in $(DIRS); do basename $$dir | awk -F_ '{print "-l mg5amc_"$$(NF-1)"_"$$NF"_cpp"}'; done)
-gpu_proclibs := $(shell for dir in $(DIRS); do basename $$dir | awk -F_ '{print "-l mg5amc_"$$(NF-1)"_"$$NF"_$(GPUSUFFIX)"}'; done)
-
-ifeq ($(GPUCC),)
-  cxx_rwgt=$(BUILDDIR)/rwgt_driver_cpp.exe
-  rwgtlib := $(addprefix ,$(addsuffix /librwgt_cpp.so,$(DIRS)))
-else
-  gpu_rwgt=$(BUILDDIR)/rwgt_driver_gpu.exe
-  rwgtlib := $(addprefix ,$(addsuffix /librwgt_$(GPUSUFFIX).so,$(DIRS)))
-endif
-
-# Explicitly define the default goal (this is not necessary as it is the first target, which is implicitly the default goal)
-.DEFAULT_GOAL := all.$(TAG)
-
-# First target (default goal)
-ifeq ($(GPUCC),)
-all.$(TAG): $(BUILDDIR)/.build.$(TAG) $(rwgtlib) $(cxx_rwgt)
-else
-all.$(TAG): $(BUILDDIR)/.build.$(TAG) $(rwgtlib) $(gpu_rwgt)
-endif
-
-# Target (and build options): debug
-MAKEDEBUG=
-debug: OPTFLAGS   = -g -O0
-debug: CUDA_OPTFLAGS = -G
-debug: MAKEDEBUG := debug
-debug: all.$(TAG)
-
-# Target: tag-specific build lockfiles
-override oldtagsb=`if [ -d $(BUILDDIR) ]; then find $(BUILDDIR) -maxdepth 1 -name '.build.*' ! -name '.build.$(TAG)' -exec echo $(shell pwd)/{} \; ; fi`
-$(BUILDDIR)/.build.$(TAG):
-	@if [ ! -d $(BUILDDIR) ]; then echo "mkdir -p $(BUILDDIR)"; mkdir -p $(BUILDDIR); fi
-	@if [ "$(oldtagsb)" != "" ]; then echo "Cannot build for tag=$(TAG) as old builds exist for other tags:"; echo "  $(oldtagsb)"; echo "Please run 'make clean' first\nIf 'make clean' is not enough: run 'make clean USEBUILDDIR=1 AVX=$(AVX) FPTYPE=$(FPTYPE)' or 'make cleanall'"; exit 1; fi
-	@touch $(BUILDDIR)/.build.$(TAG)
-
-# # Apply special build flags only to check_sa_<cpp|$(GPUSUFFIX)>.o (NVTX in timermap.h, #679)
-$(BUILDDIR)/rwgt_driver_cpp.o: CXXFLAGS += $(USE_NVTX) $(CUDA_INC)
-$(BUILDDIR)/rwgt_driver_gpu.o: CXXFLAGS += $(USE_NVTX) $(CUDA_INC)
-
-# Avoid "warning: builtin __has_trivial_... is deprecated; use __is_trivially_... instead" in GPUCC with icx2023 (#592)
-ifneq ($(shell $(CXX) --version | egrep '^(Intel)'),)
-ifneq ($(GPUCC),)
-GPUFLAGS += -Wno-deprecated-builtins
-endif
-endif
-
-# Generic target and build rules: objects from C++ compilation
-# (NB do not include CUDA_INC here! add it only for NVTX or curand #679)
-$(BUILDDIR)/%%_cpp.o : %%.cc *.h ../src/*.h $(BUILDDIR)/.build.$(TAG)
-	@if [ ! -d $(BUILDDIR) ]; then echo "mkdir -p $(BUILDDIR)"; mkdir -p $(BUILDDIR); fi
-	$(CXX) $(CPPFLAGS) $(INCFLAGS) $(CXXFLAGS) -c $< -o $@
-
-# Generic target and build rules: objects from CUDA or HIP compilation
-ifneq ($(GPUCC),)
-$(BUILDDIR)/%%_$(GPUSUFFIX).o : %%.cc *.h ../src/*.h $(BUILDDIR)/.build.$(TAG)
-	@if [ ! -d $(BUILDDIR) ]; then echo "mkdir -p $(BUILDDIR)"; mkdir -p $(BUILDDIR); fi
-	$(GPUCC) $(CPPFLAGS) $(INCFLAGS) $(GPUFLAGS) -c -x $(GPULANGUAGE) $< -o $@
-endif
-
-#-------------------------------------------------------------------------------
-
-# Target (and build rules): common (src) library
-commonlib : $(LIBDIR)/lib$(MG5AMC_COMMONLIB).so
-$(LIBDIR)/lib$(MG5AMC_COMMONLIB).so: ../src/*.h ../src/*.cc $(BUILDDIR)/.build.$(TAG)
-	$(MAKE) -C ../src $(MAKEDEBUG) -f $(CUDACPP_SRC_MAKEFILE)
-
-#-------------------------------------------------------------------------------
-
-#HERE LOOP MAKE OVER P DIRECTORIES AND ADD RWGT_RUNNER_LIBS
-# Ensure each librwgt.a depends on its directory being built
-$(rwgtlib): $(commonlib)
-	@$(MAKE) -C $(@D) VARIABLE=true
-
-# Target (and build rules): C++ and CUDA/HIP standalone executables
-$(cxx_rwgt): LIBFLAGS += $(CXXLIBFLAGSRPATH) # avoid the need for LD_LIBRARY_PATH
-$(cxx_rwgt): $(BUILDDIR)/rwgt_driver.o $(rwgtlib)
-	$(CXX) -o $@ $(BUILDDIR)/rwgt_driver.o $(OMPFLAGS) -ldl -pthread $(LIBFLAGS) -L$(LIBDIR) $(rwgtlib) 
-
-ifneq ($(GPUCC),)
-ifneq ($(shell $(CXX) --version | grep ^Intel),)
-$(gpu_rwgt): LIBFLAGS += -lintlc # compile with icpx and link with GPUCC (undefined reference to `_intel_fast_memcpy')
-$(gpu_rwgt): LIBFLAGS += -lsvml # compile with icpx and link with GPUCC (undefined reference to `__svml_cos4_l9')
-else ifneq ($(shell $(CXX) --version | grep ^nvc++),) # support nvc++ #531
-$(gpu_rwgt): LIBFLAGS += -L$(patsubst %%bin/nvc++,%%lib,$(subst ccache ,,$(CXX))) -lnvhpcatm -lnvcpumath -lnvc
-endif
-$(gpu_rwgt): LIBFLAGS += $(GPULIBFLAGSRPATH) # avoid the need for LD_LIBRARY_PATH
-$(gpu_rwgt): $(BUILDDIR)/rwgt_driver.o $(rwgtlib)
-	$(GPUCC) -o $@ $(BUILDDIR)/rwgt_driver.o $(CUARCHFLAGS) $(LIBFLAGS) -L$(LIBDIR) $(rwgtlib)
-endif
-
-#-------------------------------------------------------------------------------
-
-# Target: build all targets in all BACKEND modes (each BACKEND mode in a separate build directory)
-# Split the bldall target into separate targets to allow parallel 'make -j bldall' builds
-# (Obsolete hack, no longer needed as there is no INCDIR: add a fbridge.inc dependency to bldall, to ensure it is only copied once for all BACKEND modes)
-bldcuda:
-	@echo
-	$(MAKE) USEBUILDDIR=1 BACKEND=cuda -f $(CUDACPP_MAKEFILE)
-
-bldhip:
-	@echo
-	$(MAKE) USEBUILDDIR=1 BACKEND=hip -f $(CUDACPP_MAKEFILE)
-
-bldnone:
-	@echo
-	$(MAKE) USEBUILDDIR=1 BACKEND=scalar -f $(CUDACPP_MAKEFILE)
-
-bldsse4:
-	@echo
-	$(MAKE) USEBUILDDIR=1 BACKEND=simd_128 -f $(CUDACPP_MAKEFILE)
-
-bldavx2:
-	@echo
-	$(MAKE) USEBUILDDIR=1 BACKEND=simd_256 -f $(CUDACPP_MAKEFILE)
-
-bld512y:
-	@echo
-	$(MAKE) USEBUILDDIR=1 BACKEND=avx512y -f $(CUDACPP_MAKEFILE)
-
-bld512z:
-	@echo
-	$(MAKE) USEBUILDDIR=1 BACKEND=simd_512 -f $(CUDACPP_MAKEFILE)
-
-ifeq ($(UNAME_P),ppc64le)
-###bldavxs: $(INCDIR)/fbridge.inc bldnone bldsse4
-bldavxs: bldnone bldsse4
-else ifneq (,$(filter $(UNAME_M),arm64 aarch64))
-###bldavxs: $(INCDIR)/fbridge.inc bldnone bldsse4
-bldavxs: bldnone bldsse4
-else
-###bldavxs: $(INCDIR)/fbridge.inc bldnone bldsse4 bldavx2 bld512y bld512z
-bldavxs: bldnone bldsse4 bldavx2 bld512y bld512z
-endif
-
-ifneq ($(HIP_HOME),)
-ifneq ($(CUDA_HOME),)
-bldall: bldhip bldcuda bldavxs
-else
-bldall: bldhip bldavxs
-endif
-else
-ifneq ($(CUDA_HOME),)
-bldall: bldcuda bldavxs
-else
-bldall: bldavxs
-endif
-endif
-
-#-------------------------------------------------------------------------------
-
-# Target: clean the builds
-.PHONY: clean clean-rwgtlib
-
-clean: clean-rwgtlib
-ifeq ($(USEBUILDDIR),1)
-	rm -rf $(BUILDDIR)
-else
-	rm -f $(BUILDDIR)/.build.* $(BUILDDIR)/*.o $(BUILDDIR)/*.exe
-	rm -f $(LIBDIR)/lib*.so
-endif
-	$(MAKE) -C ../src clean -f $(CUDACPP_SRC_MAKEFILE)
-###	rm -rf $(INCDIR)
-
-clean-rwgtlib:
-	@for dir in $(DIRS); do $(MAKE) -C $$dir clean; done
-
-cleanall:
-	@echo
-	$(MAKE) USEBUILDDIR=0 clean -f $(CUDACPP_MAKEFILE)
-	@echo
-	$(MAKE) USEBUILDDIR=0 -C ../src cleanall -f $(CUDACPP_SRC_MAKEFILE)
-	rm -rf build.*
-
-# Target: clean the builds as well as the gtest installation(s)
-distclean: cleanall
-ifneq ($(wildcard $(TESTDIRCOMMON)),)
-	$(MAKE) -C $(TESTDIRCOMMON) clean
-endif
-	$(MAKE) -C $(TESTDIRLOCAL) clean
-
-#-------------------------------------------------------------------------------
-
-# Target: show system and compiler information
-info:
-	@echo ""
-	@uname -spn # e.g. Linux nodename.cern.ch x86_64
-ifeq ($(UNAME_S),Darwin)
-	@sysctl -a | grep -i brand
-	@sysctl -a | grep machdep.cpu | grep features || true
-	@sysctl -a | grep hw.physicalcpu:
-	@sysctl -a | grep hw.logicalcpu:
-else
-	@cat /proc/cpuinfo | grep "model name" | sort -u
-	@cat /proc/cpuinfo | grep "flags" | sort -u
-	@cat /proc/cpuinfo | grep "cpu cores" | sort -u
-	@cat /proc/cpuinfo | grep "physical id" | sort -u
-endif
-	@echo ""
-ifneq ($(shell which nvidia-smi 2>/dev/null),)
-	nvidia-smi -L
-	@echo ""
-endif
-	@echo USECCACHE=$(USECCACHE)
-ifeq ($(USECCACHE),1)
-	ccache --version | head -1
-endif
-	@echo ""
-	@echo GPUCC=$(GPUCC)
-ifneq ($(GPUCC),)
-	$(GPUCC) --version
-endif
-	@echo ""
-	@echo CXX=$(CXX)
-ifneq ($(shell $(CXX) --version | grep ^clang),)
-	@echo $(CXX) -v
-	@$(CXX) -v |& egrep -v '(Found|multilib)'
-	@readelf -p .comment `$(CXX) -print-libgcc-file-name` |& grep 'GCC: (GNU)' | grep -v Warning | sort -u | awk '{print "GCC toolchain:",$$5}'
-else
-	$(CXX) --version
-endif
-	@echo ""
-	@echo FC=$(FC)
-	$(FC) --version
-
-#-------------------------------------------------------------------------------
-
-# Target: 'make test' (execute runTest.exe, and compare check.exe with fcheck.exe)
-# [NB: THIS IS WHAT IS TESTED IN THE GITHUB CI!]
-# [NB: This used to be called 'make check' but the name has been changed as this has nothing to do with 'check.exe']
-test: runTest cmpFcheck
-
-# Target: runTest (run the C++ or CUDA/HIP test executable runTest.exe)
-runTest: all.$(TAG)
-ifeq ($(GPUCC),)
-	$(RUNTIME) $(BUILDDIR)/runTest_cpp.exe
-else
-	$(RUNTIME) $(BUILDDIR)/runTest_$(GPUSUFFIX).exe
-endif
-
-# Target: runCheck (run the C++ or CUDA/HIP standalone executable check.exe, with a small number of events)
-runCheck: all.$(TAG)
-ifeq ($(GPUCC),)
-	$(RUNTIME) $(BUILDDIR)/check_cpp.exe -p 2 32 2
-else
-	$(RUNTIME) $(BUILDDIR)/check_$(GPUSUFFIX).exe -p 2 32 2
-endif
-
-# Target: runFcheck (run the Fortran standalone executable - with C++ or CUDA/HIP MEs - fcheck.exe, with a small number of events)
-runFcheck: all.$(TAG)
-ifeq ($(GPUCC),)
-	$(RUNTIME) $(BUILDDIR)/fcheck_cpp.exe 2 32 2
-else
-	$(RUNTIME) $(BUILDDIR)/fcheck_$(GPUSUFFIX).exe 2 32 2
-endif
-
-# Target: cmpFcheck (compare ME results from the C++/CUDA/HIP and Fortran with C++/CUDA/HIP MEs standalone executables, with a small number of events)
-cmpFcheck: all.$(TAG)
-	@echo
-ifeq ($(GPUCC),)
-	@echo "$(BUILDDIR)/check_cpp.exe --common -p 2 32 2"
-	@echo "$(BUILDDIR)/fcheck_cpp.exe 2 32 2"
-	@me1=$(shell $(RUNTIME) $(BUILDDIR)/check_cpp.exe --common -p 2 32 2 | grep MeanMatrix | awk '{print $$4}'); me2=$(shell $(RUNTIME) $(BUILDDIR)/fcheck_cpp.exe 2 32 2 | grep Average | awk '{print $$4}'); echo "Avg ME (C++/C++)    = $${me1}"; echo "Avg ME (F77/C++)    = $${me2}"; if [ "$${me2}" == "NaN" ]; then echo "ERROR! Fortran calculation (F77/C++) returned NaN"; elif [ "$${me2}" == "" ]; then echo "ERROR! Fortran calculation (F77/C++) crashed"; else python3 -c "me1=$${me1}; me2=$${me2}; reldif=abs((me2-me1)/me1); print('Relative difference =', reldif); ok = reldif <= 2E-4; print ( '%%s (relative difference %%s 2E-4)' %% ( ('OK','<=') if ok else ('ERROR','>') ) ); import sys; sys.exit(0 if ok else 1)"; fi
-else
-	@echo "$(BUILDDIR)/check_$(GPUSUFFIX).exe --common -p 2 32 2"
-	@echo "$(BUILDDIR)/fcheck_$(GPUSUFFIX).exe 2 32 2"
-	@me1=$(shell $(RUNTIME) $(BUILDDIR)/check_$(GPUSUFFIX).exe --common -p 2 32 2 | grep MeanMatrix | awk '{print $$4}'); me2=$(shell $(RUNTIME) $(BUILDDIR)/fcheck_$(GPUSUFFIX).exe 2 32 2 | grep Average | awk '{print $$4}'); echo "Avg ME (C++/GPU)   = $${me1}"; echo "Avg ME (F77/GPU)   = $${me2}"; if [ "$${me2}" == "NaN" ]; then echo "ERROR! Fortran calculation (F77/GPU) crashed"; elif [ "$${me2}" == "" ]; then echo "ERROR! Fortran calculation (F77/GPU) crashed"; else python3 -c "me1=$${me1}; me2=$${me2}; reldif=abs((me2-me1)/me1); print('Relative difference =', reldif); ok = reldif <= 2E-4; print ( '%%s (relative difference %%s 2E-4)' %% ( ('OK','<=') if ok else ('ERROR','>') ) ); import sys; sys.exit(0 if ok else 1)"; fi
-endif
-
-# Target: cuda-memcheck (run the CUDA standalone executable gcheck.exe with a small number of events through cuda-memcheck)
-cuda-memcheck: all.$(TAG)
-	$(RUNTIME) $(CUDA_HOME)/bin/cuda-memcheck --check-api-memory-access yes --check-deprecated-instr yes --check-device-heap yes --demangle full --language c --leak-check full --racecheck-report all --report-api-errors all --show-backtrace yes --tool memcheck --track-unused-memory yes $(BUILDDIR)/check_$(GPUSUFFIX).exe -p 2 32 2
-
-#-------------------------------------------------------------------------------
diff --git a/madmatrix/MadtRex/makefiles/cudacpp_rex_src.mk b/madmatrix/MadtRex/makefiles/cudacpp_rex_src.mk
deleted file mode 100644
index b79e9a1ad4..0000000000
--- a/madmatrix/MadtRex/makefiles/cudacpp_rex_src.mk
+++ /dev/null
@@ -1,207 +0,0 @@
-# Copyright (C) 2020-2026 CERN and UCLouvain.
-# Licensed under the GNU Lesser General Public License (version 3 or later).
-# Created originally by: S. Roiser (Feb 2020) for the MadGraph7 CUDACPP plugin.
-# Further modified by: S. Hageboeck, O. Mattelaer, S. Roiser, J. Teig, A. Valassi, Z. Wettersten (2020-2024).
-# Integrated with the MadGraph7 project in Feb 2026.
-
-
-#=== Determine the name of this makefile (https://ftp.gnu.org/old-gnu/Manuals/make-3.80/html_node/make_17.html)
-#=== NB: assume that the same name (e.g. cudacpp.mk, Makefile...) is used in the Subprocess and src directories
-
-THISMK = $(word $(words $(MAKEFILE_LIST)),$(MAKEFILE_LIST))
-
-#-------------------------------------------------------------------------------
-
-#=== Use bash in the Makefile (https://www.gnu.org/software/make/manual/html_node/Choosing-the-Shell.html)
-
-SHELL := /bin/bash
-
-#-------------------------------------------------------------------------------
-
-#=== Configure common compiler flags for CUDA and C++
-
-INCFLAGS = -I.
-
-#-------------------------------------------------------------------------------
-
-#=== Configure the C++ compiler (note: CXXFLAGS has been exported from cudacpp.mk)
-
-###$(info CXXFLAGS=$(CXXFLAGS))
-
-# Note: AR, CXX and FC are implicitly defined if not set externally
-# See https://www.gnu.org/software/make/manual/html_node/Implicit-Variables.html
-###RANLIB = ranlib
-
-# Add -mmacosx-version-min=11.3 to avoid "ld: warning: object file was built for newer macOS version than being linked"
-LDFLAGS = -L. -lrex -ltearex -Wl,-rpath=.
-ifneq ($(shell $(CXX) --version | egrep '^Apple clang'),)
-  LDFLAGS += -mmacosx-version-min=11.3
-endif
-
-#-------------------------------------------------------------------------------
-
-#=== Configure the GPU (CUDA or HIP) compiler (note: GPUCC including ccache, GPUFLAGS, GPULANGUAGE, GPUSUFFIX have been exported from cudacpp.mk)
-
-###$(info GPUCC=$(GPUCC))
-###$(info GPUFLAGS=$(GPUFLAGS))
-###$(info GPULANGUAGE=$(GPULANGUAGE))
-###$(info GPUSUFFIX=$(GPUSUFFIX))
-
-#-------------------------------------------------------------------------------
-
-#=== Configure ccache for C++ builds (note: GPUCC has been exported from cudacpp.mk including ccache)
-
-# Enable ccache if USECCACHE=1
-ifeq ($(USECCACHE)$(shell echo $(CXX) | grep ccache),1)
-  override CXX:=ccache $(CXX)
-endif
-#ifeq ($(USECCACHE)$(shell echo $(AR) | grep ccache),1)
-#  override AR:=ccache $(AR)
-#endif
-
-#-------------------------------------------------------------------------------
-
-#=== Configure build directories and build lockfiles ===
-
-# Use the build directory exported from cudacpp.mk
-###$(info CUDACPP_BUILDDIR=$(CUDACPP_BUILDDIR))
-
-# Use the build lockfile "full" tag exported from cudacpp.mk
-###$(info TAG=$(TAG))
-
-# Build directory: current directory by default, or build.$(DIRTAG) if USEBUILDDIR==1
-###$(info Current directory is $(shell pwd))
-override BUILDDIR = $(CUDACPP_BUILDDIR)
-ifeq ($(USEBUILDDIR),1)
-  override LIBDIRREL = ../lib/$(BUILDDIR)
-  ###$(info Building in BUILDDIR=$(BUILDDIR) for tag=$(TAG) (USEBUILDDIR=1 is set))
-else
-  override LIBDIRREL = ../lib
-  ###$(info Building in BUILDDIR=$(BUILDDIR) for tag=$(TAG) (USEBUILDDIR is not set))
-endif
-######$(info Building in BUILDDIR=$(BUILDDIR) for tag=$(TAG))
-
-# Workaround for Mac #375 (I did not manage to fix rpath with @executable_path): use absolute paths for LIBDIR
-# (NB: this is quite ugly because it creates the directory if it does not exist - to avoid removing src by mistake)
-UNAME_S := $(shell uname -s)
-ifeq ($(UNAME_S),Darwin)
-  override LIBDIR = $(shell mkdir -p $(LIBDIRREL); cd $(LIBDIRREL); pwd)
-  ifeq ($(wildcard $(LIBDIR)),)
-    $(error Directory LIBDIR="$(LIBDIR)" should have been created by now)
-  endif
-else
-  override LIBDIR = $(LIBDIRREL)
-endif
-
-#===============================================================================
-#=== Makefile TARGETS and build rules below
-#===============================================================================
-
-# NB1: there are no CUDA targets in src as we avoid RDC!
-# NB2: CUDA includes for curand.h are no longer needed in the C++ code anywhere in src!
-
-ifeq ($(GPUCC),)
-MG5AMC_COMMONLIB = mg5amc_common_cpp
-else
-MG5AMC_COMMONLIB = mg5amc_common_$(GPUSUFFIX)
-endif
-
-# Explicitly define the default goal (this is not necessary as it is the first target, which is implicitly the default goal)
-.DEFAULT_GOAL := all.$(TAG)
-
-# First target (default goal)
-all.$(TAG): $(BUILDDIR)/.build.$(TAG) $(LIBDIR)/.build.$(TAG) $(LIBDIR)/lib$(MG5AMC_COMMONLIB).so $(LIBDIR)/librex.so $(LIBDIR)/libtearex.so
-
-# Target (and build options): debug
-debug: all.$(TAG)
-
-# Target: tag-specific build lockfiles
-override oldtagsb=`if [ -d $(BUILDDIR) ]; then find $(BUILDDIR) -maxdepth 1 -name '.build.*' ! -name '.build.$(TAG)' -exec echo $(shell pwd)/{} \; ; fi`
-override oldtagsl=`if [ -d $(LIBDIR) ]; then find $(LIBDIR) -maxdepth 1 -name '.build.*' ! -name '.build.$(TAG)' -exec echo $(shell pwd)/{} \; ; fi`
-
-$(BUILDDIR)/.build.$(TAG): $(LIBDIR)/.build.$(TAG)
-
-$(LIBDIR)/.build.$(TAG):
-	@if [ "$(oldtagsl)" != "" ]; then echo -e "Cannot build for tag=$(TAG) as old builds exist in $(LIBDIR) for other tags:\n$(oldtagsl)\nPlease run 'make clean' first\nIf 'make clean' is not enough: run 'make cleanall'"; exit 1; fi
-	@if [ "$(oldtagsb)" != "" ]; then echo -e "Cannot build for tag=$(TAG) as old builds exist in $(BUILDDIR) for other tags:\n$(oldtagsb)\nPlease run 'make clean' first\nIf 'make clean' is not enough: run 'make cleanall'"; exit 1; fi
-	@if [ ! -d $(LIBDIR) ]; then echo "mkdir -p $(LIBDIR)"; mkdir -p $(LIBDIR); fi
-	@touch $(LIBDIR)/.build.$(TAG)
-	@if [ ! -d $(BUILDDIR) ]; then echo "mkdir -p $(BUILDDIR)"; mkdir -p $(BUILDDIR); fi
-	@touch $(BUILDDIR)/.build.$(TAG)
-
-#-------------------------------------------------------------------------------
-
-# Generic target and build rules: objects from C++ compilation
-$(BUILDDIR)/%%_cpp.o : %%.cc *.h $(BUILDDIR)/.build.$(TAG)
-	@if [ ! -d $(BUILDDIR) ]; then echo "mkdir -p $(BUILDDIR)"; mkdir -p $(BUILDDIR); fi
-	$(CXX) $(CPPFLAGS) $(INCFLAGS) $(CXXFLAGS) -c $< -o $@
-
-# Generic target and build rules: objects from CUDA compilation
-ifneq ($(GPUCC),)
-$(BUILDDIR)/%%_$(GPUSUFFIX).o : %%.cc *.h $(BUILDDIR)/.build.$(TAG)
-	@if [ ! -d $(BUILDDIR) ]; then echo "mkdir -p $(BUILDDIR)"; mkdir -p $(BUILDDIR); fi
-	$(GPUCC) $(CPPFLAGS) $(INCFLAGS) $(GPUFLAGS) -c -x $(GPULANGUAGE) $< -o $@
-endif
-
-#-------------------------------------------------------------------------------
-
-cxx_objects=$(addprefix $(BUILDDIR)/, read_slha_cpp.o)
-cxx_objects+=$(addprefix $(BUILDDIR)/, rwgt_instance_cpp.o)
-ifeq ($(GPUCC),)
-  cxx_objects+=$(addprefix $(BUILDDIR)/, Parameters_cpp.o)
-else
-  gpu_objects=$(addprefix $(BUILDDIR)/, Parameters_$(GPUSUFFIX).o)
-endif
-
-# Target (and build rules): common (src) library
-ifeq ($(GPUCC),)
-$(LIBDIR)/lib$(MG5AMC_COMMONLIB).so : $(cxx_objects)
-	@if [ ! -d $(LIBDIR) ]; then echo "mkdir -p $(LIBDIR)"; mkdir -p $(LIBDIR); fi
-	$(CXX) -shared -o $@ $(cxx_objects) $(LDFLAGS)
-else
-$(LIBDIR)/lib$(MG5AMC_COMMONLIB).so : $(cxx_objects) $(gpu_objects)
-	@if [ ! -d $(LIBDIR) ]; then echo "mkdir -p $(LIBDIR)"; mkdir -p $(LIBDIR); fi
-	$(GPUCC) -shared -o $@ $(cxx_objects) $(gpu_objects) $(LDFLAGS)
-endif
-
-#-------------------------------------------------------------------------------
-
-# Atomic copy helper macro:
-#   1) copy to unique temp file next to the destination
-#   2) atomically rename into place
-# This is safe under parallel make and concurrent invocations.
-define ATOMIC_COPY
-	@tmp="$$(mktemp "$@.tmp.XXXXXX")"; \
-	cp -f "$(firstword $^)" "$$tmp"; \
-	mv -f "$$tmp" "$@"
-endef
-
-# Rex and teaRex: copy .so from src to LIBDIR atomically
-$(LIBDIR)/librex.so : ../src/librex.so 
-	$(ATOMIC_COPY)
-
-$(LIBDIR)/libtearex.so : ../src/libtearex.so
-	$(ATOMIC_COPY)
-
-#-------------------------------------------------------------------------------
-
-# Target: clean the builds
-.PHONY: clean
-
-clean:
-ifeq ($(USEBUILDDIR),1)
-	rm -rf $(LIBDIR)
-	rm -rf $(BUILDDIR)
-else
-	rm -f $(LIBDIR)/.build.* $(LIBDIR)/lib$(MG5AMC_COMMONLIB).so
-	rm -f $(BUILDDIR)/.build.* $(BUILDDIR)/*.o $(BUILDDIR)/*.exe
-endif
-
-cleanall:
-	@echo
-	$(MAKE) clean -f $(THISMK)
-	@echo
-	rm -rf $(LIBDIR)/build.*
-	rm -rf build.*
-
-#-------------------------------------------------------------------------------
diff --git a/madmatrix/MadtRex/makefiles/cudacpp_runner.mk b/madmatrix/MadtRex/makefiles/cudacpp_runner.mk
deleted file mode 100644
index 54eb900bd3..0000000000
--- a/madmatrix/MadtRex/makefiles/cudacpp_runner.mk
+++ /dev/null
@@ -1,908 +0,0 @@
-# Copyright (C) 2020-2026 CERN and UCLouvain.
-# Licensed under the GNU Lesser General Public License (version 3 or later).
-# Created originally by: S. Roiser (Feb 2020) for the MadGraph7 CUDACPP plugin.
-# Further modified by: S. Hageboeck, O. Mattelaer, S. Roiser, J. Teig, A. Valassi, Z. Wettersten (2020-2024).
-# Integrated with the MadGraph7 project in Feb 2026.
-
-#=== Determine the name of this makefile (https://ftp.gnu.org/old-gnu/Manuals/make-3.80/html_node/make_17.html)
-#=== NB: use ':=' to ensure that the value of CUDACPP_MAKEFILE is not modified further down after including make_opts
-#=== NB: use 'override' to ensure that the value can not be modified from the outside
-override CUDACPP_MAKEFILE := $(word $(words $(MAKEFILE_LIST)),$(MAKEFILE_LIST))
-###$(info CUDACPP_MAKEFILE='$(CUDACPP_MAKEFILE)')
-
-#=== NB: different names (e.g. cudacpp.mk and cudacpp_src.mk) are used in the Subprocess and src directories
-override CUDACPP_SRC_MAKEFILE = cudacpp_src.mk
-
-#-------------------------------------------------------------------------------
-
-#=== Include cudacpp_config.mk
-
-# Check that the user-defined choices of BACKEND, FPTYPE, HELINL, HRDCOD are supported (and configure defaults if no user-defined choices exist)
-# Stop with an error if BACKEND=cuda and nvcc is missing or if BACKEND=hip and hipcc is missing.
-# Determine CUDACPP_BUILDDIR from a DIRTAG based on BACKEND, FPTYPE, HELINL, HRDCOD and from the user-defined choice of USEBUILDDIR
-include ../../src/cudacpp_config.mk
-
-# Export CUDACPP_BUILDDIR (so that there is no need to check/define it again in cudacpp_src.mk)
-#export CUDACPP_BUILDDIR
-
-#-------------------------------------------------------------------------------
-
-#=== Use bash in the Makefile (https://www.gnu.org/software/make/manual/html_node/Choosing-the-Shell.html)
-
-SHELL := /bin/bash
-
-#-------------------------------------------------------------------------------
-
-#=== Detect O/S and architecture (assuming uname is available, https://en.wikipedia.org/wiki/Uname)
-
-# Detect O/S kernel (Linux, Darwin...)
-UNAME_S := $(shell uname -s)
-###$(info UNAME_S='$(UNAME_S)')
-
-# Detect architecture (x86_64, ppc64le...)
-UNAME_P := $(shell uname -p)
-###$(info UNAME_P='$(UNAME_P)')
-UNAME_M := $(shell uname -m)
-
-#-------------------------------------------------------------------------------
-
-#=== Include the common MadGraph7 Makefile options
-
-# OM: including make_opts is crucial for MadGraph7 flag consistency/documentation
-# AV: disable the inclusion of make_opts if the file has not been generated (standalone cudacpp)
-ifneq ($(wildcard ../../Source/make_opts),)
-  include ../../Source/make_opts
-endif
-
-#-------------------------------------------------------------------------------
-
-#=== Redefine BACKEND if the current value is 'auto'
-
-# Set the BACKEND choice corresponding to 'auto' (the 'best' C++ vectorization available)
-ifeq ($(BACKEND),auto)
-  ifeq ($(UNAME_P),ppc64le)
-    override BACKEND = simd_128
-  else ifneq (,$(filter $(UNAME_M),arm64 aarch64))
-    override BACKEND = simd_128
-  else ifeq ($(wildcard /proc/cpuinfo),)
-    override BACKEND = scalar
-    ###$(warning Using BACKEND='$(BACKEND)' because host SIMD features cannot be read from /proc/cpuinfo)
-  else ifeq ($(shell grep -m1 -c avx512vl /proc/cpuinfo)$(shell $(CXX) --version | grep ^clang),1)
-    override BACKEND = avx512y
-  else ifeq ($(shell grep -m1 -c avx2 /proc/cpuinfo),1)
-    override BACKEND = simd_256
-    ###ifneq ($(shell grep -m1 -c avx512vl /proc/cpuinfo),1)
-    ###  $(warning Using BACKEND='$(BACKEND)' because host does not support avx512vl)
-    ###else
-    ###  $(warning Using BACKEND='$(BACKEND)' because this is faster than avx512vl for clang)
-    ###endif
-  else ifeq ($(shell grep -m1 -c sse4_2 /proc/cpuinfo),1)
-    override BACKEND = simd_128
-  else
-    override BACKEND = scalar
-  endif
-  $(info BACKEND=$(BACKEND) (was auto))
-else
-  $(info BACKEND='$(BACKEND)')
-endif
-
-#-------------------------------------------------------------------------------
-
-#=== Configure the GPU compiler (CUDA or HIP)
-#=== (note, this is done also for C++, as NVTX and CURAND/ROCRAND are also needed by the C++ backends)
-
-# Set CUDA_HOME from the path to nvcc, if it exists
-#override CUDA_HOME = $(patsubst %%/bin/nvcc,%%,$(shell which nvcc 2>/dev/null))
-
-# Set HIP_HOME from the path to hipcc, if it exists
-#override HIP_HOME = $(patsubst %%/bin/hipcc,%%,$(shell which hipcc 2>/dev/null))
-
-# Configure CUDA_INC (for CURAND and NVTX) and NVTX if a CUDA installation exists
-# (FIXME? Is there any equivalent of NVTX FOR HIP? What should be configured if both CUDA and HIP are installed?)
-ifneq ($(CUDA_HOME),)
-  USE_NVTX ?=-DUSE_NVTX
-  CUDA_INC = -I$(CUDA_HOME)/include/
-else
-  override USE_NVTX=
-  override CUDA_INC=
-endif
-
-
-#=== Configure common compiler flags for C++ and CUDA/HIP
-
-INCFLAGS = -I.
-OPTFLAGS = -O3 # this ends up in GPUFLAGS too (should it?), cannot add -Ofast or -ffast-math here
-
-# Dependency on src directory
-ifeq ($(GPUCC),)
-MG5AMC_COMMONLIB = mg5amc_common_cpp
-else
-MG5AMC_COMMONLIB = mg5amc_common_$(GPUSUFFIX)
-endif
-LIBFLAGS = -L$(LIBDIR) -l$(MG5AMC_COMMONLIB) -lrex -ltearex
-INCFLAGS += -I../../src
-
-# Compiler-specific googletest build directory (#125 and #738)
-ifneq ($(shell $(CXX) --version | grep '^Intel(R) oneAPI DPC++/C++ Compiler'),)
-  override CXXNAME = icpx$(shell $(CXX) --version | head -1 | cut -d' ' -f5)
-else ifneq ($(shell $(CXX) --version | egrep '^clang'),)
-  override CXXNAME = clang$(shell $(CXX) --version | head -1 | cut -d' ' -f3)
-else ifneq ($(shell $(CXX) --version | grep '^g++ (GCC)'),)
-  override CXXNAME = gcc$(shell $(CXX) --version | head -1 | cut -d' ' -f3)
-else
-  override CXXNAME = unknown
-endif
-###$(info CXXNAME=$(CXXNAME))
-# override CXXNAMESUFFIX = _$(CXXNAME)
-
-# # Export CXXNAMESUFFIX (so that there is no need to check/define it again in cudacpp_test.mk)
-# export CXXNAMESUFFIX
-
-# Dependency on test directory
-# Within the madgraph4gpu git repo: by default use a common gtest installation in <topdir>/test (optionally use an external or local gtest)
-# Outside the madgraph4gpu git repo: by default do not build the tests (optionally use an external or local gtest)
-###GTEST_ROOT = /cvmfs/sft.cern.ch/lcg/releases/gtest/1.11.0-21e8c/x86_64-centos8-gcc11-opt/# example of an external gtest installation
-###LOCALGTEST = yes# comment this out (or use make LOCALGTEST=yes) to build tests using a local gtest installation
-TESTDIRCOMMON = ../../../../../test
-TESTDIRLOCAL = ../../test
-ifneq ($(wildcard $(GTEST_ROOT)),)
-  TESTDIR =
-else ifneq ($(LOCALGTEST),)
-  TESTDIR=$(TESTDIRLOCAL)
-  GTEST_ROOT = $(TESTDIR)/googletest/install$(CXXNAMESUFFIX)
-else ifneq ($(wildcard ../../../../../epochX/cudacpp/CODEGEN),)
-  TESTDIR = $(TESTDIRCOMMON)
-  GTEST_ROOT = $(TESTDIR)/googletest/install$(CXXNAMESUFFIX)
-else
-  TESTDIR =
-endif
-ifneq ($(GTEST_ROOT),)
-  GTESTLIBDIR = $(GTEST_ROOT)/lib64/
-  GTESTLIBS = $(GTESTLIBDIR)/libgtest.a
-  GTESTINC = -I$(GTEST_ROOT)/include
-else
-  GTESTLIBDIR =
-  GTESTLIBS =
-  GTESTINC =
-endif
-###$(info GTEST_ROOT = $(GTEST_ROOT))
-###$(info LOCALGTEST = $(LOCALGTEST))
-###$(info TESTDIR = $(TESTDIR))
-
-#-------------------------------------------------------------------------------
-
-#=== Configure PowerPC-specific compiler flags for C++ and CUDA/HIP
-
-# PowerPC-specific CXX compiler flags (being reviewed)
-ifeq ($(UNAME_P),ppc64le)
-  CXXFLAGS+= -mcpu=power9 -mtune=power9 # gains ~2-3%% both for scalar and simd_128
-  # Throughput references without the extra flags below: scalar=1.41-1.42E6, simd_128=2.15-2.19E6
-else
-  ###CXXFLAGS+= -flto # also on Intel this would increase throughputs by a factor 2 to 4...
-  ######CXXFLAGS+= -fno-semantic-interposition # no benefit (neither alone, nor combined with -flto)
-endif
-
-# PowerPC-specific CUDA/HIP compiler flags (to be reviewed!)
-ifeq ($(UNAME_P),ppc64le)
-  GPUFLAGS+= $(XCOMPILERFLAG) -mno-float128
-endif
-
-#-------------------------------------------------------------------------------
-
-#=== Configure defaults for OMPFLAGS
-
-# Disable OpenMP by default: enable OpenMP only if USEOPENMP=1 (#758)
-ifeq ($(USEOPENMP),1)
-  ###$(info USEOPENMP==1: will build with OpenMP if possible)
-  ifneq ($(findstring hipcc,$(GPUCC)),)
-    override OMPFLAGS = # disable OpenMP MT when using hipcc #802
-  else ifneq ($(shell $(CXX) --version | egrep '^Intel'),)
-    override OMPFLAGS = -fopenmp
-    ###override OMPFLAGS = # disable OpenMP MT on Intel (was ok without GPUCC but not ok with GPUCC before #578)
-  else ifneq ($(shell $(CXX) --version | egrep '^clang version 16'),)
-    ###override OMPFLAGS = # disable OpenMP on clang16 #904
-    $(error OpenMP is not supported by cudacpp on clang16 - issue #904)
-  else ifneq ($(shell $(CXX) --version | egrep '^clang version 17'),)
-    ###override OMPFLAGS = # disable OpenMP on clang17 #904
-    $(error OpenMP is not supported by cudacpp on clang17 - issue #904)
-  else ifneq ($(shell $(CXX) --version | egrep '^(clang)'),)
-    override OMPFLAGS = -fopenmp
-    ###override OMPFLAGS = # disable OpenMP MT on clang (was not ok without or with nvcc before #578)
-  ###else ifneq ($(shell $(CXX) --version | egrep '^(Apple clang)'),) # AV for Mac (Apple clang compiler)
-  else ifeq ($(UNAME_S),Darwin) # OM for Mac (any compiler)
-    override OMPFLAGS = # AV disable OpenMP MT on Apple clang (builds fail in the CI #578)
-    ###override OMPFLAGS = -fopenmp # OM re-enable OpenMP MT on Apple clang? (AV Oct 2023: this still fails in the CI)
-  else
-    override OMPFLAGS = -fopenmp # enable OpenMP MT by default on all other platforms
-    ###override OMPFLAGS = # disable OpenMP MT on all other platforms (default before #575)
-  endif
-else
-  ###$(info USEOPENMP!=1: will build without OpenMP)
-  override OMPFLAGS =
-endif
-
-#-------------------------------------------------------------------------------
-
-#=== Configure defaults and check if user-defined choices exist for RNDGEN (legacy!), HASCURAND, HASHIPRAND
-
-# If the legacy RNDGEN exists, this take precedence over any HASCURAND choice (but a warning is printed out)
-###$(info RNDGEN=$(RNDGEN))
-ifneq ($(RNDGEN),)
-  $(warning Environment variable RNDGEN is no longer supported, please use HASCURAND instead!)
-  ifeq ($(RNDGEN),hasCurand)
-    override HASCURAND = $(RNDGEN)
-  else ifeq ($(RNDGEN),hasNoCurand)
-    override HASCURAND = $(RNDGEN)
-  else ifneq ($(RNDGEN),hasNoCurand)
-    $(error Unknown RNDGEN='$(RNDGEN)': only 'hasCurand' and 'hasNoCurand' are supported - but use HASCURAND instead!)
-  endif
-endif
-
-# Set the default HASCURAND (curand random number generator) choice, if no prior choice exists for HASCURAND
-# (NB: allow HASCURAND=hasCurand even if $(GPUCC) does not point to nvcc: assume CUDA_HOME was defined correctly...)
-ifeq ($(HASCURAND),)
-  ifeq ($(GPUCC),) # CPU-only build
-    ifneq ($(CUDA_HOME),)
-      # By default, assume that curand is installed if a CUDA installation exists
-      override HASCURAND = hasCurand
-    else
-      override HASCURAND = hasNoCurand
-    endif
-  else ifeq ($(findstring nvcc,$(GPUCC)),nvcc) # Nvidia GPU build
-    override HASCURAND = hasCurand
-  else # non-Nvidia GPU build
-    override HASCURAND = hasNoCurand
-  endif
-endif
-
-# Set the default HASHIPRAND (hiprand random number generator) choice, if no prior choice exists for HASHIPRAND
-# (NB: allow HASHIPRAND=hasHiprand even if $(GPUCC) does not point to hipcc: assume HIP_HOME was defined correctly...)
-ifeq ($(HASHIPRAND),)
-  ifeq ($(GPUCC),) # CPU-only build
-    override HASHIPRAND = hasNoHiprand
-  else ifeq ($(findstring hipcc,$(GPUCC)),hipcc) # AMD GPU build
-    override HASHIPRAND = hasHiprand
-  else # non-AMD GPU build
-    override HASHIPRAND = hasNoHiprand
-  endif
-endif
-
-#-------------------------------------------------------------------------------
-
-#=== Set the CUDA/HIP/C++ compiler flags appropriate to user-defined choices of AVX, FPTYPE, HELINL, HRDCOD
-
-# Set the build flags appropriate to OMPFLAGS
-$(info OMPFLAGS=$(OMPFLAGS))
-CXXFLAGS += $(OMPFLAGS)
-
-# Set the build flags appropriate to each BACKEND choice (example: "make BACKEND=scalar")
-# [NB MGONGPU_PVW512 is needed because "-mprefer-vector-width=256" is not exposed in a macro]
-# [Use 'g++ <buildflags> -E -dM - < /dev/null' to check which #define's are enabled]
-# [See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96476]
-ifeq ($(UNAME_P),ppc64le)
-  ifeq ($(BACKEND),simd_128)
-    override AVXFLAGS = -D__SSE4_2__ # Power9 VSX with 128 width (VSR registers)
-  else ifeq ($(BACKEND),simd_256)
-    $(error Invalid SIMD BACKEND='$(BACKEND)': only 'scalar' and 'simd_128' are supported on PowerPC for the moment)
-  else ifeq ($(BACKEND),avx512y)
-    $(error Invalid SIMD BACKEND='$(BACKEND)': only 'scalar' and 'simd_128' are supported on PowerPC for the moment)
-  else ifeq ($(BACKEND),simd_512)
-    $(error Invalid SIMD BACKEND='$(BACKEND)': only 'scalar' and 'simd_128' are supported on PowerPC for the moment)
-  endif
-else ifeq ($(UNAME_M),arm64) # ARM on Apple silicon
-  ifeq ($(BACKEND),scalar) # this internally undefines __ARM_NEON
-    override AVXFLAGS = -DMGONGPU_NOARMNEON
-  else ifeq ($(BACKEND),simd_128) # __ARM_NEON is always defined on Apple silicon
-    override AVXFLAGS =
-  else ifeq ($(BACKEND),simd_256)
-    $(error Invalid SIMD BACKEND='$(BACKEND)': only 'scalar' and 'simd_128' are supported on ARM for the moment)
-  else ifeq ($(BACKEND),avx512y)
-    $(error Invalid SIMD BACKEND='$(BACKEND)': only 'scalar' and 'simd_128' are supported on ARM for the moment)
-  else ifeq ($(BACKEND),simd_512)
-    $(error Invalid SIMD BACKEND='$(BACKEND)': only 'scalar' and 'simd_128' are supported on ARM for the moment)
-  endif
-else ifeq ($(UNAME_M),aarch64) # ARM on Linux
-  ifeq ($(BACKEND),scalar) # +nosimd ensures __ARM_NEON is absent
-    override AVXFLAGS = -march=armv8-a+nosimd
-  else ifeq ($(BACKEND),simd_128) # +simd ensures __ARM_NEON is present (128 width Q/quadword registers)
-    override AVXFLAGS = -march=armv8-a+simd
-  else ifeq ($(BACKEND),simd_256)
-    $(error Invalid SIMD BACKEND='$(BACKEND)': only 'scalar' and 'simd_128' are supported on aarch64 for the moment)
-  else ifeq ($(BACKEND),avx512y)
-    $(error Invalid SIMD BACKEND='$(BACKEND)': only 'scalar' and 'simd_128' are supported on aarch64 for the moment)
-  else ifeq ($(BACKEND),simd_512)
-    $(error Invalid SIMD BACKEND='$(BACKEND)': only 'scalar' and 'simd_128' are supported on aarch64 for the moment)
-  endif
-else ifneq ($(shell $(CXX) --version | grep ^nvc++),) # support nvc++ #531
-  ifeq ($(BACKEND),scalar)
-    override AVXFLAGS = -mno-sse3 # no SIMD
-  else ifeq ($(BACKEND),simd_128)
-    override AVXFLAGS = -mno-avx # SSE4.2 with 128 width (xmm registers)
-  else ifeq ($(BACKEND),simd_256)
-    override AVXFLAGS = -march=haswell # AVX2 with 256 width (ymm registers) [DEFAULT for clang]
-  else ifeq ($(BACKEND),avx512y)
-    override AVXFLAGS = -march=skylake -mprefer-vector-width=256 # AVX512 with 256 width (ymm registers) [DEFAULT for gcc]
-  else ifeq ($(BACKEND),simd_512)
-    override AVXFLAGS = -march=skylake -DMGONGPU_PVW512 # AVX512 with 512 width (zmm registers)
-  endif
-else
-  ifeq ($(BACKEND),scalar)
-    override AVXFLAGS = -march=x86-64 # no SIMD (see #588)
-  else ifeq ($(BACKEND),simd_128)
-    override AVXFLAGS = -march=nehalem # SSE4.2 with 128 width (xmm registers)
-  else ifeq ($(BACKEND),simd_256)
-    override AVXFLAGS = -march=haswell # AVX2 with 256 width (ymm registers) [DEFAULT for clang]
-  else ifeq ($(BACKEND),avx512y)
-    override AVXFLAGS = -march=skylake-avx512 -mprefer-vector-width=256 # AVX512 with 256 width (ymm registers) [DEFAULT for gcc]
-  else ifeq ($(BACKEND),simd_512)
-    override AVXFLAGS = -march=skylake-avx512 -DMGONGPU_PVW512 # AVX512 with 512 width (zmm registers)
-  endif
-endif
-# For the moment, use AVXFLAGS everywhere (in C++ builds): eventually, use them only in encapsulated implementations?
-ifeq ($(GPUCC),)
-  CXXFLAGS+= $(AVXFLAGS)
-endif
-
-# Set the build flags appropriate to each FPTYPE choice (example: "make FPTYPE=f")
-$(info FPTYPE='$(FPTYPE)')
-ifeq ($(FPTYPE),d)
-  CXXFLAGS += -DMGONGPU_FPTYPE_DOUBLE -DMGONGPU_FPTYPE2_DOUBLE
-  GPUFLAGS += -DMGONGPU_FPTYPE_DOUBLE -DMGONGPU_FPTYPE2_DOUBLE
-else ifeq ($(FPTYPE),f)
-  CXXFLAGS += -DMGONGPU_FPTYPE_FLOAT -DMGONGPU_FPTYPE2_FLOAT
-  GPUFLAGS += -DMGONGPU_FPTYPE_FLOAT -DMGONGPU_FPTYPE2_FLOAT
-else ifeq ($(FPTYPE),m)
-  CXXFLAGS += -DMGONGPU_FPTYPE_DOUBLE -DMGONGPU_FPTYPE2_FLOAT
-  GPUFLAGS += -DMGONGPU_FPTYPE_DOUBLE -DMGONGPU_FPTYPE2_FLOAT
-else
-  $(error Unknown FPTYPE='$(FPTYPE)': only 'd', 'f' and 'm' are supported)
-endif
-
-# Set the build flags appropriate to each HELINL choice (example: "make HELINL=1")
-$(info HELINL='$(HELINL)')
-ifeq ($(HELINL),1)
-  CXXFLAGS += -DMGONGPU_INLINE_HELAMPS
-  GPUFLAGS += -DMGONGPU_INLINE_HELAMPS
-else ifneq ($(HELINL),0)
-  $(error Unknown HELINL='$(HELINL)': only '0' and '1' are supported)
-endif
-
-# Set the build flags appropriate to each HRDCOD choice (example: "make HRDCOD=1")
-$(info HRDCOD='$(HRDCOD)')
-ifeq ($(HRDCOD),1)
-  CXXFLAGS += -DMGONGPU_HARDCODE_PARAM
-  GPUFLAGS += -DMGONGPU_HARDCODE_PARAM
-else ifneq ($(HRDCOD),0)
-  $(error Unknown HRDCOD='$(HRDCOD)': only '0' and '1' are supported)
-endif
-
-#=== Set the CUDA/HIP/C++ compiler and linker flags appropriate to user-defined choices of HASCURAND, HASHIPRAND
-
-$(info HASCURAND=$(HASCURAND))
-$(info HASHIPRAND=$(HASHIPRAND))
-override RNDCXXFLAGS=
-override RNDLIBFLAGS=
-
-# Set the RNDCXXFLAGS and RNDLIBFLAGS build flags appropriate to each HASCURAND choice (example: "make HASCURAND=hasNoCurand")
-ifeq ($(HASCURAND),hasNoCurand)
-  override RNDCXXFLAGS += -DMGONGPU_HAS_NO_CURAND
-else ifeq ($(HASCURAND),hasCurand)
-  override RNDLIBFLAGS += -L$(CUDA_HOME)/lib64/ -lcurand # NB: -lcuda is not needed here!
-else
-  $(error Unknown HASCURAND='$(HASCURAND)': only 'hasCurand' and 'hasNoCurand' are supported)
-endif
-
-# Set the RNDCXXFLAGS and RNDLIBFLAGS build flags appropriate to each HASHIPRAND choice (example: "make HASHIPRAND=hasNoHiprand")
-ifeq ($(HASHIPRAND),hasNoHiprand)
-  override RNDCXXFLAGS += -DMGONGPU_HAS_NO_HIPRAND
-else ifeq ($(HASHIPRAND),hasHiprand)
-  override RNDLIBFLAGS += -L$(HIP_HOME)/lib/ -lhiprand
-else ifneq ($(HASHIPRAND),hasHiprand)
-  $(error Unknown HASHIPRAND='$(HASHIPRAND)': only 'hasHiprand' and 'hasNoHiprand' are supported)
-endif
-
-#$(info RNDCXXFLAGS=$(RNDCXXFLAGS))
-#$(info RNDLIBFLAGS=$(RNDLIBFLAGS))
-
-#-------------------------------------------------------------------------------
-
-#=== Configure Position-Independent Code
-CXXFLAGS += -fPIC
-GPUFLAGS += $(XCOMPILERFLAG) -fPIC
-
-#-------------------------------------------------------------------------------
-
-#=== Configure build directories and build lockfiles ===
-
-# Build lockfile "full" tag (defines full specification of build options that cannot be intermixed)
-# (Rationale: avoid mixing of builds with different random number generators)
-override TAG = $(BACKEND)_$(FPTYPE)_inl$(HELINL)_hrd$(HRDCOD)_$(HASCURAND)_$(HASHIPRAND)
-
-# Export TAG (so that there is no need to check/define it again in cudacpp_src.mk)
-export TAG
-
-# Build directory: current directory by default, or build.$(DIRTAG) if USEBUILDDIR==1
-override BUILDDIR = $(CUDACPP_BUILDDIR)
-ifeq ($(USEBUILDDIR),1)
-  override LIBDIR = ../../lib/$(BUILDDIR)
-  override LIBDIRRPATH = '$$ORIGIN/../$(LIBDIR)'
-  $(info Building in BUILDDIR=$(BUILDDIR) for tag=$(TAG) (USEBUILDDIR == 1))
-else
-  override LIBDIR = ../../lib
-  override LIBDIRRPATH = '$$ORIGIN/$(LIBDIR)'
-  $(info Building in BUILDDIR=$(BUILDDIR) for tag=$(TAG) (USEBUILDDIR != 1))
-endif
-###override INCDIR = ../../include
-###$(info Building in BUILDDIR=$(BUILDDIR) for tag=$(TAG))
-
-# On Linux, set rpath to LIBDIR to make it unnecessary to use LD_LIBRARY_PATH
-# Use relative paths with respect to the executables or shared libraries ($ORIGIN on Linux)
-# On Darwin, building libraries with absolute paths in LIBDIR makes this unnecessary
-ifeq ($(UNAME_S),Darwin)
-  override CXXLIBFLAGSRPATH =
-  override GPULIBFLAGSRPATH =
-  override CXXLIBFLAGSRPATH2 =
-  override GPULIBFLAGSRPATH2 =
-else
-  # RPATH to gpu/cpp libs when linking executables
-  override CXXLIBFLAGSRPATH = -Wl,-rpath=$(LIBDIRRPATH)
-  override GPULIBFLAGSRPATH = -Xlinker -rpath=$(LIBDIRRPATH)
-  # RPATH to common lib when linking gpu/cpp libs
-  override CXXLIBFLAGSRPATH2 = -Wl,-rpath='$$ORIGIN'
-  override GPULIBFLAGSRPATH2 = -Xlinker -rpath='$$ORIGIN'
-endif
-
-# Setting LD_LIBRARY_PATH or DYLD_LIBRARY_PATH in the RUNTIME is no longer necessary (neither on Linux nor on Mac)
-override RUNTIME =
-
-#===============================================================================
-#=== Makefile TARGETS and build rules below
-#===============================================================================
-
-
-ifeq ($(GPUCC),)
-  cxx_rwgtlib=$(BUILDDIR)/librwgt_cpp.so
-else
-  gpu_rwgtlib=$(BUILDDIR)/librwgt_$(GPUSUFFIX).so
-endif
-
-# Explicitly define the default goal (this is not necessary as it is the first target, which is implicitly the default goal)
-.DEFAULT_GOAL := all.$(TAG)
-
-# First target (default goal)
-ifeq ($(GPUCC),)
-all.$(TAG): $(BUILDDIR)/.build.$(TAG) $(cxx_rwgtlib) 
-else
-all.$(TAG): $(BUILDDIR)/.build.$(TAG) $(gpu_rwgtlib) 
-endif
-
-# Target (and build options): debug
-MAKEDEBUG=
-debug: OPTFLAGS   = -g -O0
-debug: CUDA_OPTFLAGS = -G
-debug: MAKEDEBUG := debug
-debug: all.$(TAG)
-
-# Target: tag-specific build lockfiles
-override oldtagsb=`if [ -d $(BUILDDIR) ]; then find $(BUILDDIR) -maxdepth 1 -name '.build.*' ! -name '.build.$(TAG)' -exec echo $(shell pwd)/{} \; ; fi`
-$(BUILDDIR)/.build.$(TAG):
-	@if [ ! -d $(BUILDDIR) ]; then echo "mkdir -p $(BUILDDIR)"; mkdir -p $(BUILDDIR); fi
-	@if [ "$(oldtagsb)" != "" ]; then echo "Cannot build for tag=$(TAG) as old builds exist for other tags:"; echo "  $(oldtagsb)"; echo "Please run 'make clean' first\nIf 'make clean' is not enough: run 'make clean USEBUILDDIR=1 AVX=$(AVX) FPTYPE=$(FPTYPE)' or 'make cleanall'"; exit 1; fi
-	@touch $(BUILDDIR)/.build.$(TAG)
-
-# Apply special build flags only to CrossSectionKernel_<cpp|$(GPUSUFFIX)>.o (no fast math, see #117 and #516)
-# Added edgecase for HIP compilation
-ifeq ($(shell $(CXX) --version | grep ^nvc++),)
-$(BUILDDIR)/CrossSectionKernels_cpp.o: CXXFLAGS := $(filter-out -ffast-math,$(CXXFLAGS))
-$(BUILDDIR)/CrossSectionKernels_cpp.o: CXXFLAGS += -fno-fast-math
-$(BUILDDIR)/CrossSectionKernels_$(GPUSUFFIX).o: GPUFLAGS += $(XCOMPILERFLAG) -fno-fast-math
-endif
-
-# Apply special build flags only to check_sa_<cpp|$(GPUSUFFIX)>.o (NVTX in timermap.h, #679)
-$(BUILDDIR)/check_sa_cpp.o: CXXFLAGS += $(USE_NVTX) $(CUDA_INC)
-$(BUILDDIR)/rwgt_runner_cpp.o: CXXFLAGS += $(USE_NVTX) $(CUDA_INC)
-$(BUILDDIR)/check_sa_$(GPUSUFFIX).o: CXXFLAGS += $(USE_NVTX) $(CUDA_INC)
-$(BUILDDIR)/rwgt_runner_$(GPUSUFFIX).o: CXXFLAGS += $(USE_NVTX) $(CUDA_INC)
-
-# Apply special build flags only to check_sa_<cpp|$(GPUSUFFIX)>.o and (Cu|Hip)randRandomNumberKernel_<cpp|$(GPUSUFFIX)>.o
-$(BUILDDIR)/check_sa_cpp.o: CXXFLAGS += $(RNDCXXFLAGS)
-$(BUILDDIR)/rwgt_runner_cpp.o: CXXFLAGS += $(RNDCXXFLAGS)
-$(BUILDDIR)/check_sa_$(GPUSUFFIX).o: GPUFLAGS += $(RNDCXXFLAGS)
-$(BUILDDIR)/rwgt_runner_$(GPUSUFFIX).o: GPUFLAGS += $(RNDCXXFLAGS)
-$(BUILDDIR)/CurandRandomNumberKernel_cpp.o: CXXFLAGS += $(RNDCXXFLAGS)
-$(BUILDDIR)/CurandRandomNumberKernel_$(GPUSUFFIX).o: GPUFLAGS += $(RNDCXXFLAGS)
-$(BUILDDIR)/HiprandRandomNumberKernel_cpp.o: CXXFLAGS += $(RNDCXXFLAGS)
-$(BUILDDIR)/HiprandRandomNumberKernel_$(GPUSUFFIX).o: GPUFLAGS += $(RNDCXXFLAGS)
-ifeq ($(HASCURAND),hasCurand) # curand headers, #679
-$(BUILDDIR)/CurandRandomNumberKernel_cpp.o: CXXFLAGS += $(CUDA_INC)
-endif
-ifeq ($(HASHIPRAND),hasHiprand) # hiprand headers
-$(BUILDDIR)/HiprandRandomNumberKernel_cpp.o: CXXFLAGS += $(HIP_INC)
-endif
-
-# Avoid "warning: builtin __has_trivial_... is deprecated; use __is_trivially_... instead" in GPUCC with icx2023 (#592)
-ifneq ($(shell $(CXX) --version | egrep '^(Intel)'),)
-ifneq ($(GPUCC),)
-GPUFLAGS += -Wno-deprecated-builtins
-endif
-endif
-
-# Generic target and build rules: objects from C++ compilation
-# (NB do not include CUDA_INC here! add it only for NVTX or curand #679)
-$(BUILDDIR)/%%_cpp.o : %%.cc *.h ../../src/*.h $(BUILDDIR)/.build.$(TAG)
-	@if [ ! -d $(BUILDDIR) ]; then echo "mkdir -p $(BUILDDIR)"; mkdir -p $(BUILDDIR); fi
-	$(CXX) $(CPPFLAGS) $(INCFLAGS) $(CXXFLAGS) -c $< -o $@
-
-# Generic target and build rules: objects from CUDA or HIP compilation
-ifneq ($(GPUCC),)
-$(BUILDDIR)/%%_$(GPUSUFFIX).o : %%.cc *.h ../../src/*.h $(BUILDDIR)/.build.$(TAG)
-	@if [ ! -d $(BUILDDIR) ]; then echo "mkdir -p $(BUILDDIR)"; mkdir -p $(BUILDDIR); fi
-	$(GPUCC) $(CPPFLAGS) $(INCFLAGS) $(GPUFLAGS) -c -x $(GPULANGUAGE) $< -o $@
-endif
-
-#-------------------------------------------------------------------------------
-
-# Target (and build rules): common (src) library
-commonlib : $(LIBDIR)/lib$(MG5AMC_COMMONLIB).so
-
-$(LIBDIR)/lib$(MG5AMC_COMMONLIB).so: ../../src/*.h ../../src/*.cc $(BUILDDIR)/.build.$(TAG)
-	$(MAKE) -C ../../src $(MAKEDEBUG) -f $(CUDACPP_SRC_MAKEFILE)
-
-#-------------------------------------------------------------------------------
-
-processid_short=$(shell basename $(CURDIR) | awk -F_ '{print $$(NF-1)"_"$$NF}')
-###$(info processid_short=$(processid_short))
-
-MG5AMC_CXXLIB = mg5amc_$(processid_short)_cpp
-cxx_objects_lib=$(BUILDDIR)/CPPProcess_cpp.o $(BUILDDIR)/color_sum_cpp.o $(BUILDDIR)/MatrixElementKernels_cpp.o $(BUILDDIR)/BridgeKernels_cpp.o $(BUILDDIR)/CrossSectionKernels_cpp.o
-cxx_objects_exe=$(BUILDDIR)/CommonRandomNumberKernel_cpp.o $(BUILDDIR)/RamboSamplingKernels_cpp.o
-
-ifneq ($(GPUCC),)
-MG5AMC_GPULIB = mg5amc_$(processid_short)_$(GPUSUFFIX)
-gpu_objects_lib=$(BUILDDIR)/CPPProcess_$(GPUSUFFIX).o $(BUILDDIR)/color_sum_$(GPUSUFFIX).o $(BUILDDIR)/MatrixElementKernels_$(GPUSUFFIX).o $(BUILDDIR)/BridgeKernels_$(GPUSUFFIX).o $(BUILDDIR)/CrossSectionKernels_$(GPUSUFFIX).o
-gpu_objects_exe=$(BUILDDIR)/CommonRandomNumberKernel_$(GPUSUFFIX).o $(BUILDDIR)/RamboSamplingKernels_$(GPUSUFFIX).o
-endif
-
-# Target (and build rules): C++ and CUDA/HIP shared libraries
-$(LIBDIR)/lib$(MG5AMC_CXXLIB).so: $(BUILDDIR)/fbridge_cpp.o
-$(LIBDIR)/lib$(MG5AMC_CXXLIB).so: cxx_objects_lib += $(BUILDDIR)/fbridge_cpp.o
-$(LIBDIR)/lib$(MG5AMC_CXXLIB).so: $(LIBDIR)/lib$(MG5AMC_COMMONLIB).so $(cxx_objects_lib)
-	$(CXX) -shared -o $@ $(cxx_objects_lib) $(CXXLIBFLAGSRPATH2) -L$(LIBDIR) -l$(MG5AMC_COMMONLIB)
-
-ifneq ($(GPUCC),)
-$(LIBDIR)/lib$(MG5AMC_GPULIB).so: $(BUILDDIR)/fbridge_$(GPUSUFFIX).o
-$(LIBDIR)/lib$(MG5AMC_GPULIB).so: gpu_objects_lib += $(BUILDDIR)/fbridge_$(GPUSUFFIX).o
-$(LIBDIR)/lib$(MG5AMC_GPULIB).so: $(LIBDIR)/lib$(MG5AMC_COMMONLIB).so $(gpu_objects_lib)
-	$(GPUCC) --shared -o $@ $(gpu_objects_lib) $(GPULIBFLAGSRPATH2) -L$(LIBDIR) -l$(MG5AMC_COMMONLIB)
-# Bypass std::filesystem completely to ease portability on LUMI #803
-#ifneq ($(findstring hipcc,$(GPUCC)),)
-#	$(GPUCC) --shared -o $@ $(gpu_objects_lib) $(GPULIBFLAGSRPATH2) -L$(LIBDIR) -l$(MG5AMC_COMMONLIB) -lstdc++fs
-#else
-#	$(GPUCC) --shared -o $@ $(gpu_objects_lib) $(GPULIBFLAGSRPATH2) -L$(LIBDIR) -l$(MG5AMC_COMMONLIB)
-#endif
-endif
-
-#-------------------------------------------------------------------------------
-
-# Target (and build rules): C++ rwgt libraries
-# ZW: the -Bsymbolic flag ensures that function calls will be handled internally by the library, rather than going to global context
-cxx_rwgtfiles := $(BUILDDIR)/rwgt_runner_cpp.o $(LIBDIR)/lib$(MG5AMC_COMMONLIB).so $(BUILDDIR)/fbridge_cpp.o $(cxx_objects_lib) $(cxx_objects_exe) $(BUILDDIR)/CurandRandomNumberKernel_cpp.o $(BUILDDIR)/HiprandRandomNumberKernel_cpp.o
-$(cxx_rwgtlib): LIBFLAGS += $(CXXLIBFLAGSRPATH)
-$(cxx_rwgtlib): $(LIBDIR)/lib$(MG5AMC_COMMONLIB).so $(cxx_rwgtfiles) $(cxx_objects_lib)
-	$(CXX) -shared -Wl,-Bsymbolic -o $@ $(BUILDDIR)/rwgt_runner_cpp.o  $(OMPFLAGS) -ldl -pthread $(LIBFLAGS) -L$(LIBDIR) $(BUILDDIR)/fbridge_cpp.o $(cxx_objects_lib) $(cxx_objects_exe) $(BUILDDIR)/CurandRandomNumberKernel_cpp.o $(BUILDDIR)/HiprandRandomNumberKernel_cpp.o $(RNDLIBFLAGS)
-
-ifneq ($(GPUCC),)
-ifneq ($(shell $(CXX) --version | grep ^Intel),)
-$(gpu_checkmain): LIBFLAGS += -lintlc # compile with icpx and link with GPUCC (undefined reference to `_intel_fast_memcpy')
-$(gpu_checkmain): LIBFLAGS += -lsvml # compile with icpx and link with GPUCC (undefined reference to `__svml_cos4_l9')
-else ifneq ($(shell $(CXX) --version | grep ^nvc++),) # support nvc++ #531
-$(gpu_checkmain): LIBFLAGS += -L$(patsubst %%bin/nvc++,%%lib,$(subst ccache ,,$(CXX))) -lnvhpcatm -lnvcpumath -lnvc
-endif
-$(gpu_checkmain): LIBFLAGS += $(GPULIBFLAGSRPATH) # avoid the need for LD_LIBRARY_PATH
-$(gpu_checkmain): $(BUILDDIR)/check_sa_$(GPUSUFFIX).o $(gp_objects_lib) $(gpu_objects_exe) $(BUILDDIR)/CurandRandomNumberKernel_$(GPUSUFFIX).o $(BUILDDIR)/HiprandRandomNumberKernel_$(GPUSUFFIX).o
-	$(GPUCC) -o $@ $(BUILDDIR)/check_sa_$(GPUSUFFIX).o $(LIBFLAGS) -L$(LIBDIR) -l$(MG5AMC_GPULIB) $(gpu_objects_exe) $(BUILDDIR)/CurandRandomNumberKernel_$(GPUSUFFIX).o $(BUILDDIR)/HiprandRandomNumberKernel_$(GPUSUFFIX).o $(RNDLIBFLAGS)
-ifneq ($(shell $(CXX) --version | grep ^Intel),)
-$(gpu_rwgtlib): LIBFLAGS += -lintlc # compile with icpx and link with GPUCC (undefined reference to `_intel_fast_memcpy')
-$(gpu_rwgtlib): LIBFLAGS += -lsvml # compile with icpx and link with GPUCC (undefined reference to `__svml_cos4_l9')
-else ifneq ($(shell $(CXX) --version | grep ^nvc++),) # support nvc++ #531
-$(gpu_rwgtlib): LIBFLAGS += -L$(patsubst %%bin/nvc++,%%lib,$(subst ccache ,,$(CXX))) -lnvhpcatm -lnvcpumath -lnvc
-endif
-$(gpu_rwgtlib): LIBFLAGS += $(GPULIBFLAGSRPATH) # avoid the need for LD_LIBRARY_PATH
-gpu_rwgtfiles := $(BUILDDIR)/rwgt_runner_$(GPUSUFFIX).o $(LIBDIR)/lib$(MG5AMC_COMMONLIB).so $(gpu_objects_lib) $(gpu_objects_exe) $(BUILDDIR)/fbridge_$(GPUSUFFIX).o $(BUILDDIR)/CurandRandomNumberKernel_$(GPUSUFFIX).o $(BUILDDIR)/HiprandRandomNumberKernel_$(GPUSUFFIX).o
-$(gpu_rwgtlib): $(LIBDIR)/lib$(MG5AMC_COMMONLIB).so $(gpu_rwgtfiles) $(gpu_objects_lib)
-	$(GPUCC) -shared -Xcompiler \"-Wl,-Bsymbolic\" -o $@ $(BUILDDIR)/rwgt_runner_$(GPUSUFFIX).o $(LIBFLAGS) -L$(LIBDIR) $(BUILDDIR)/fbridge_$(GPUSUFFIX).o $(gpu_objects_exe) $(gpu_objects_lib) $(BUILDDIR)/CurandRandomNumberKernel_$(GPUSUFFIX).o $(BUILDDIR)/HiprandRandomNumberKernel_$(GPUSUFFIX).o $(RNDLIBFLAGS)
-endif
-
-#-------------------------------------------------------------------------------
-
-# Target (and build rules): test objects and test executable
-ifeq ($(GPUCC),)
-$(BUILDDIR)/testxxx_cpp.o: $(GTESTLIBS)
-$(BUILDDIR)/testxxx_cpp.o: INCFLAGS += $(GTESTINC)
-$(BUILDDIR)/testxxx_cpp.o: testxxx_cc_ref.txt
-$(cxx_testmain): $(BUILDDIR)/testxxx_cpp.o
-$(cxx_testmain): cxx_objects_exe += $(BUILDDIR)/testxxx_cpp.o # Comment out this line to skip the C++ test of xxx functions
-else
-$(BUILDDIR)/testxxx_$(GPUSUFFIX).o: $(GTESTLIBS)
-$(BUILDDIR)/testxxx_$(GPUSUFFIX).o: INCFLAGS += $(GTESTINC)
-$(BUILDDIR)/testxxx_$(GPUSUFFIX).o: testxxx_cc_ref.txt
-$(gpu_testmain): $(BUILDDIR)/testxxx_$(GPUSUFFIX).o
-$(gpu_testmain): gpu_objects_exe += $(BUILDDIR)/testxxx_$(GPUSUFFIX).o # Comment out this line to skip the CUDA/HIP test of xxx functions
-endif
-
-ifneq ($(UNAME_S),Darwin) # Disable testmisc on Darwin (workaround for issue #838)
-ifeq ($(GPUCC),)
-$(BUILDDIR)/testmisc_cpp.o: $(GTESTLIBS)
-$(BUILDDIR)/testmisc_cpp.o: INCFLAGS += $(GTESTINC)
-$(cxx_testmain): $(BUILDDIR)/testmisc_cpp.o
-$(cxx_testmain): cxx_objects_exe += $(BUILDDIR)/testmisc_cpp.o # Comment out this line to skip the C++ miscellaneous tests
-else
-$(BUILDDIR)/testmisc_$(GPUSUFFIX).o: $(GTESTLIBS)
-$(BUILDDIR)/testmisc_$(GPUSUFFIX).o: INCFLAGS += $(GTESTINC)
-$(gpu_testmain): $(BUILDDIR)/testmisc_$(GPUSUFFIX).o
-$(gpu_testmain): gpu_objects_exe += $(BUILDDIR)/testmisc_$(GPUSUFFIX).o # Comment out this line to skip the CUDA/HIP miscellaneous tests
-endif
-endif
-
-ifeq ($(GPUCC),)
-$(BUILDDIR)/runTest_cpp.o: $(GTESTLIBS)
-$(BUILDDIR)/runTest_cpp.o: INCFLAGS += $(GTESTINC)
-$(cxx_testmain): $(BUILDDIR)/runTest_cpp.o
-$(cxx_testmain): cxx_objects_exe += $(BUILDDIR)/runTest_cpp.o
-else
-$(BUILDDIR)/runTest_$(GPUSUFFIX).o: $(GTESTLIBS)
-$(BUILDDIR)/runTest_$(GPUSUFFIX).o: INCFLAGS += $(GTESTINC)
-ifneq ($(shell $(CXX) --version | grep ^Intel),)
-$(gpu_testmain): LIBFLAGS += -lintlc # compile with icpx and link with GPUCC (undefined reference to `_intel_fast_memcpy')
-$(gpu_testmain): LIBFLAGS += -lsvml # compile with icpx and link with GPUCC (undefined reference to `__svml_cos4_l9')
-else ifneq ($(shell $(CXX) --version | grep ^nvc++),) # support nvc++ #531
-$(gpu_testmain): LIBFLAGS += -L$(patsubst %%bin/nvc++,%%lib,$(subst ccache ,,$(CXX))) -lnvhpcatm -lnvcpumath -lnvc
-endif
-$(gpu_testmain): $(BUILDDIR)/runTest_$(GPUSUFFIX).o
-$(gpu_testmain): gpu_objects_exe  += $(BUILDDIR)/runTest_$(GPUSUFFIX).o
-endif
-
-ifeq ($(GPUCC),)
-$(cxx_testmain): $(GTESTLIBS)
-$(cxx_testmain): INCFLAGS +=  $(GTESTINC)
-$(cxx_testmain): LIBFLAGS += -L$(GTESTLIBDIR) -lgtest # adding also -lgtest_main is no longer necessary since we added main() to testxxx.cc
-else
-$(gpu_testmain): $(GTESTLIBS)
-$(gpu_testmain): INCFLAGS +=  $(GTESTINC)
-$(gpu_testmain): LIBFLAGS += -L$(GTESTLIBDIR) -lgtest # adding also -lgtest_main is no longer necessary since we added main() to testxxx.cc
-endif
-
-ifeq ($(GPUCC),) # if at all, OMP is used only in CXX builds (not in GPU builds)
-ifneq ($(OMPFLAGS),)
-ifneq ($(shell $(CXX) --version | egrep '^Intel'),)
-$(cxx_testmain): LIBFLAGS += -liomp5 # see #578 (not '-qopenmp -static-intel' as in https://stackoverflow.com/questions/45909648)
-else ifneq ($(shell $(CXX) --version | egrep '^clang'),)
-$(cxx_testmain): LIBFLAGS += -L $(shell dirname $(shell $(CXX) -print-file-name=libc++.so)) -lomp # see #604
-###else ifneq ($(shell $(CXX) --version | egrep '^Apple clang'),)
-###$(cxx_testmain): LIBFLAGS += ???? # OMP is not supported yet by cudacpp for Apple clang (see #578 and #604)
-else
-$(cxx_testmain): LIBFLAGS += -lgomp
-endif
-endif
-endif
-
-# Test quadmath in testmisc.cc tests for constexpr_math #627
-###ifeq ($(GPUCC),)
-###$(cxx_testmain): LIBFLAGS += -lquadmath
-###else
-###$(gpu_testmain): LIBFLAGS += -lquadmath
-###endif
-
-# Bypass std::filesystem completely to ease portability on LUMI #803
-###ifneq ($(findstring hipcc,$(GPUCC)),)
-###$(gpu_testmain): LIBFLAGS += -lstdc++fs
-###endif
-
-ifeq ($(GPUCC),) # link only runTest_cpp.o
-$(cxx_testmain): LIBFLAGS += $(CXXLIBFLAGSRPATH) # avoid the need for LD_LIBRARY_PATH
-$(cxx_testmain): $(LIBDIR)/lib$(MG5AMC_COMMONLIB).so $(cxx_objects_lib) $(cxx_objects_exe) $(GTESTLIBS)
-	$(CXX) -o $@ $(cxx_objects_lib) $(cxx_objects_exe) -ldl -pthread $(LIBFLAGS)
-else # link only runTest_$(GPUSUFFIX).o (new: in the past, this was linking both runTest_cpp.o and runTest_$(GPUSUFFIX).o)
-$(gpu_testmain): LIBFLAGS += $(GPULIBFLAGSRPATH) # avoid the need for LD_LIBRARY_PATH
-$(gpu_testmain): $(LIBDIR)/lib$(MG5AMC_COMMONLIB).so $(gpu_objects_lib) $(gpu_objects_exe) $(GTESTLIBS)
-ifneq ($(findstring hipcc,$(GPUCC)),) # link fortran/c++/hip using $FC when hipcc is used #802
-	$(FC) -o $@ $(gpu_objects_lib) $(gpu_objects_exe) -ldl $(LIBFLAGS) -lstdc++ -lpthread  -L$(shell dirname $(shell $(GPUCC) -print-prog-name=clang))/../../lib -lamdhip64
-else
-	$(GPUCC) -o $@ $(gpu_objects_lib) $(gpu_objects_exe) -ldl $(LIBFLAGS) -lcuda
-endif
-endif
-
-# Use target gtestlibs to build only googletest
-ifneq ($(GTESTLIBS),)
-gtestlibs: $(GTESTLIBS)
-endif
-
-# Use flock (Linux only, no Mac) to allow 'make -j' if googletest has not yet been downloaded https://stackoverflow.com/a/32666215
-$(GTESTLIBS):
-ifneq ($(shell which flock 2>/dev/null),)
-	@if [ ! -d $(BUILDDIR) ]; then echo "mkdir -p $(BUILDDIR)"; mkdir -p $(BUILDDIR); fi
-	flock $(BUILDDIR)/.make_test.lock $(MAKE) -C $(TESTDIR)
-else
-	if [ -d $(TESTDIR) ]; then $(MAKE) -C $(TESTDIR); fi
-endif
-
-#-------------------------------------------------------------------------------
-
-# Target: build all targets in all BACKEND modes (each BACKEND mode in a separate build directory)
-# Split the bldall target into separate targets to allow parallel 'make -j bldall' builds
-# (Obsolete hack, no longer needed as there is no INCDIR: add a fbridge.inc dependency to bldall, to ensure it is only copied once for all BACKEND modes)
-bldcuda:
-	@echo
-	$(MAKE) USEBUILDDIR=1 BACKEND=cuda -f $(CUDACPP_MAKEFILE)
-
-bldhip:
-	@echo
-	$(MAKE) USEBUILDDIR=1 BACKEND=hip -f $(CUDACPP_MAKEFILE)
-
-bldnone:
-	@echo
-	$(MAKE) USEBUILDDIR=1 BACKEND=scalar -f $(CUDACPP_MAKEFILE)
-
-bldsse4:
-	@echo
-	$(MAKE) USEBUILDDIR=1 BACKEND=simd_128 -f $(CUDACPP_MAKEFILE)
-
-bldavx2:
-	@echo
-	$(MAKE) USEBUILDDIR=1 BACKEND=simd_256 -f $(CUDACPP_MAKEFILE)
-
-bld512y:
-	@echo
-	$(MAKE) USEBUILDDIR=1 BACKEND=avx512y -f $(CUDACPP_MAKEFILE)
-
-bld512z:
-	@echo
-	$(MAKE) USEBUILDDIR=1 BACKEND=simd_512 -f $(CUDACPP_MAKEFILE)
-
-ifeq ($(UNAME_P),ppc64le)
-###bldavxs: $(INCDIR)/fbridge.inc bldnone bldsse4
-bldavxs: bldnone bldsse4
-else ifneq (,$(filter $(UNAME_M),arm64 aarch64))
-###bldavxs: $(INCDIR)/fbridge.inc bldnone bldsse4
-bldavxs: bldnone bldsse4
-else
-###bldavxs: $(INCDIR)/fbridge.inc bldnone bldsse4 bldavx2 bld512y bld512z
-bldavxs: bldnone bldsse4 bldavx2 bld512y bld512z
-endif
-
-ifneq ($(HIP_HOME),)
-ifneq ($(CUDA_HOME),)
-bldall: bldhip bldcuda bldavxs
-else
-bldall: bldhip bldavxs
-endif
-else
-ifneq ($(CUDA_HOME),)
-bldall: bldcuda bldavxs
-else
-bldall: bldavxs
-endif
-endif
-
-#-------------------------------------------------------------------------------
-
-# Target: clean the builds
-.PHONY: clean
-
-clean:
-ifeq ($(USEBUILDDIR),1)
-	rm -rf $(BUILDDIR)
-else
-	rm -f $(BUILDDIR)/.build.* $(BUILDDIR)/*.o $(BUILDDIR)/*.so $(BUILDDIR)/*.exe
-	rm -f $(LIBDIR)/lib*.so
-endif
-	$(MAKE) -C ../../src clean -f $(CUDACPP_SRC_MAKEFILE)
-###	rm -rf $(INCDIR)
-
-cleanall:
-	@echo
-	$(MAKE) USEBUILDDIR=0 clean -f $(CUDACPP_MAKEFILE)
-	@echo
-	$(MAKE) USEBUILDDIR=0 -C ../../src cleanall -f $(CUDACPP_SRC_MAKEFILE)
-	rm -rf build.*
-
-# Target: clean the builds as well as the gtest installation(s)
-distclean: cleanall
-ifneq ($(wildcard $(TESTDIRCOMMON)),)
-	$(MAKE) -C $(TESTDIRCOMMON) clean
-endif
-	$(MAKE) -C $(TESTDIRLOCAL) clean
-
-#-------------------------------------------------------------------------------
-
-# Target: show system and compiler information
-info:
-	@echo ""
-	@uname -spn # e.g. Linux nodename.cern.ch x86_64
-ifeq ($(UNAME_S),Darwin)
-	@sysctl -a | grep -i brand
-	@sysctl -a | grep machdep.cpu | grep features || true
-	@sysctl -a | grep hw.physicalcpu:
-	@sysctl -a | grep hw.logicalcpu:
-else
-	@cat /proc/cpuinfo | grep "model name" | sort -u
-	@cat /proc/cpuinfo | grep "flags" | sort -u
-	@cat /proc/cpuinfo | grep "cpu cores" | sort -u
-	@cat /proc/cpuinfo | grep "physical id" | sort -u
-endif
-	@echo ""
-ifneq ($(shell which nvidia-smi 2>/dev/null),)
-	nvidia-smi -L
-	@echo ""
-endif
-	@echo USECCACHE=$(USECCACHE)
-ifeq ($(USECCACHE),1)
-	ccache --version | head -1
-endif
-	@echo ""
-	@echo GPUCC=$(GPUCC)
-ifneq ($(GPUCC),)
-	$(GPUCC) --version
-endif
-	@echo ""
-	@echo CXX=$(CXX)
-ifneq ($(shell $(CXX) --version | grep ^clang),)
-	@echo $(CXX) -v
-	@$(CXX) -v |& egrep -v '(Found|multilib)'
-	@readelf -p .comment `$(CXX) -print-libgcc-file-name` |& grep 'GCC: (GNU)' | grep -v Warning | sort -u | awk '{print "GCC toolchain:",$$5}'
-else
-	$(CXX) --version
-endif
-	@echo ""
-	@echo FC=$(FC)
-	$(FC) --version
-
-#-------------------------------------------------------------------------------
-
-# Target: 'make test' (execute runTest.exe, and compare check.exe with fcheck.exe)
-# [NB: THIS IS WHAT IS TESTED IN THE GITHUB CI!]
-# [NB: This used to be called 'make check' but the name has been changed as this has nothing to do with 'check.exe']
-test: runTest cmpFcheck
-
-# Target: runTest (run the C++ or CUDA/HIP test executable runTest.exe)
-runTest: all.$(TAG)
-ifeq ($(GPUCC),)
-	$(RUNTIME) $(BUILDDIR)/runTest_cpp.exe
-else
-	$(RUNTIME) $(BUILDDIR)/runTest_$(GPUSUFFIX).exe
-endif
-
-# Target: runCheck (run the C++ or CUDA/HIP standalone executable check.exe, with a small number of events)
-runCheck: all.$(TAG)
-ifeq ($(GPUCC),)
-	$(RUNTIME) $(BUILDDIR)/check_cpp.exe -p 2 32 2
-else
-	$(RUNTIME) $(BUILDDIR)/check_$(GPUSUFFIX).exe -p 2 32 2
-endif
-
-# Target: runFcheck (run the Fortran standalone executable - with C++ or CUDA/HIP MEs - fcheck.exe, with a small number of events)
-runFcheck: all.$(TAG)
-ifeq ($(GPUCC),)
-	$(RUNTIME) $(BUILDDIR)/fcheck_cpp.exe 2 32 2
-else
-	$(RUNTIME) $(BUILDDIR)/fcheck_$(GPUSUFFIX).exe 2 32 2
-endif
-
-# Target: cmpFcheck (compare ME results from the C++/CUDA/HIP and Fortran with C++/CUDA/HIP MEs standalone executables, with a small number of events)
-cmpFcheck: all.$(TAG)
-	@echo
-ifeq ($(GPUCC),)
-	@echo "$(BUILDDIR)/check_cpp.exe --common -p 2 32 2"
-	@echo "$(BUILDDIR)/fcheck_cpp.exe 2 32 2"
-	@me1=$(shell $(RUNTIME) $(BUILDDIR)/check_cpp.exe --common -p 2 32 2 | grep MeanMatrix | awk '{print $$4}'); me2=$(shell $(RUNTIME) $(BUILDDIR)/fcheck_cpp.exe 2 32 2 | grep Average | awk '{print $$4}'); echo "Avg ME (C++/C++)    = $${me1}"; echo "Avg ME (F77/C++)    = $${me2}"; if [ "$${me2}" == "NaN" ]; then echo "ERROR! Fortran calculation (F77/C++) returned NaN"; elif [ "$${me2}" == "" ]; then echo "ERROR! Fortran calculation (F77/C++) crashed"; else python3 -c "me1=$${me1}; me2=$${me2}; reldif=abs((me2-me1)/me1); print('Relative difference =', reldif); ok = reldif <= 2E-4; print ( '%%s (relative difference %%s 2E-4)' %% ( ('OK','<=') if ok else ('ERROR','>') ) ); import sys; sys.exit(0 if ok else 1)"; fi
-else
-	@echo "$(BUILDDIR)/check_$(GPUSUFFIX).exe --common -p 2 32 2"
-	@echo "$(BUILDDIR)/fcheck_$(GPUSUFFIX).exe 2 32 2"
-	@me1=$(shell $(RUNTIME) $(BUILDDIR)/check_$(GPUSUFFIX).exe --common -p 2 32 2 | grep MeanMatrix | awk '{print $$4}'); me2=$(shell $(RUNTIME) $(BUILDDIR)/fcheck_$(GPUSUFFIX).exe 2 32 2 | grep Average | awk '{print $$4}'); echo "Avg ME (C++/GPU)   = $${me1}"; echo "Avg ME (F77/GPU)   = $${me2}"; if [ "$${me2}" == "NaN" ]; then echo "ERROR! Fortran calculation (F77/GPU) crashed"; elif [ "$${me2}" == "" ]; then echo "ERROR! Fortran calculation (F77/GPU) crashed"; else python3 -c "me1=$${me1}; me2=$${me2}; reldif=abs((me2-me1)/me1); print('Relative difference =', reldif); ok = reldif <= 2E-4; print ( '%%s (relative difference %%s 2E-4)' %% ( ('OK','<=') if ok else ('ERROR','>') ) ); import sys; sys.exit(0 if ok else 1)"; fi
-endif
-
-# Target: cuda-memcheck (run the CUDA standalone executable gcheck.exe with a small number of events through cuda-memcheck)
-cuda-memcheck: all.$(TAG)
-	$(RUNTIME) $(CUDA_HOME)/bin/cuda-memcheck --check-api-memory-access yes --check-deprecated-instr yes --check-device-heap yes --demangle full --language c --leak-check full --racecheck-report all --report-api-errors all --show-backtrace yes --tool memcheck --track-unused-memory yes $(BUILDDIR)/check_$(GPUSUFFIX).exe -p 2 32 2
-
-#-------------------------------------------------------------------------------
diff --git a/madmatrix/MadtRex/rex.mk b/madmatrix/MadtRex/rex.mk
deleted file mode 100644
index 03f858c8fa..0000000000
--- a/madmatrix/MadtRex/rex.mk
+++ /dev/null
@@ -1,26 +0,0 @@
-CXX = g++
-CXXFLAGS = -std=c++17 -Wall -Wextra -fPIC -O3
-LDFLAGS = -L. -lrex -Wl,-rpath=.
-
-REX_SRC = Rex.cc
-REX_HDR = Rex.h
-REX_OBJ = Rex.o
-REX_TARGET = librex.so
-
-TEA_SRC = teaRex.cc
-TEA_HDR = teaRex.h
-TEA_OBJ = teaRex.o
-TEA_TARGET = libtearex.so
-
-all: $(REX_TARGET) $(TEA_TARGET) 
-
-
-# Build shared library
-$(REX_TARGET): $(REX_SRC)
-	$(CXX) $(CXXFLAGS) -shared -o $@ $^
-
-$(TEA_TARGET): $(TEA_SRC) $(REX_TARGET)
-	$(CXX) $(CXXFLAGS) -shared -o $@ $(TEA_SRC) $(LDFLAGS)
-
-clean:
-	rm -f $(REX_TARGET) $(TEA_TARGET) $(REX_OBJ) $(TEA_OBJ)
\ No newline at end of file
diff --git a/madmatrix/MadtRex/teaRex.cc b/madmatrix/MadtRex/teaRex.cc
deleted file mode 100644
index d47b45fba3..0000000000
--- a/madmatrix/MadtRex/teaRex.cc
+++ /dev/null
@@ -1,939 +0,0 @@
-/***
- *     _            ______
- *    | |           | ___ \
- *    | |_ ___  __ _| |_/ /_____  __
- *    | __/ _ \/ _` |    // _ \ \/ /
- *    | ||  __/ (_| | |\ \  __/>  <
- *     \__\___|\__,_\_| \_\___/_/\_\
- *
- ***/
-//
-// *t*ensorial *e*vent *a*daption with *R*e*x* Version 1.0.0
-// teaRex is an extension to the Rex library for the generic reweighting of parton-level events.
-// It provides a flexible framework for applying weight modifications to events based on user-defined criteria,
-// using the underlying Rex formats to sort, extract, and rewrite event-level information,
-// and extending it to allow for generic reweighting using any information stored in an LHE file as input for a
-// user-provided reweighting function acting on REX::process objects, which are SoA (Structure of Arrays)
-// objects for storing event information. Users can either provide the REX::process objects themselves,
-// or use the flexible Rex sorting architecture to extract the necessary information from an LHE file.
-//
-// Copyright © 2023-2026 CERN, CERN Author Zenny Wettersten.
-// Licensed under the GNU Lesser General Public License (version 3 or later).
-// All rights not expressly granted are reserved.
-//
-
-#ifndef _TEAREX_CPP_
-#define _TEAREX_CPP_
-
-#include "teaRex.h"
-
-namespace REX::tea
-{
-
-    bool true_function()
-    {
-        return true;
-    }
-
-    void rwgt_slha::rwgt_card::add_param(const std::string &block_name, std::pair<int, double> param)
-    {
-        if (this->blocks.find(block_name) == this->blocks.end())
-            this->blocks[block_name] = rwgt_block{block_name, {}};
-        this->blocks[block_name].params.push_back(param);
-    }
-
-    void rwgt_slha::rwgt_card::add_param(const std::string &block_name, int param_id, double param_value)
-    {
-        this->add_param(block_name, std::make_pair(param_id, param_value));
-    }
-
-    void rwgt_slha::parse_rwgt_card(std::istream &is)
-    {
-        this->cards.clear();
-        size_t curr_launch = 0;
-        this->cards.emplace_back();
-        std::string line;
-        bool first_launch = false;
-        while (std::getline(is, line))
-        {
-            if (line.empty() || line[0] == '#')
-                continue; // Skip empty lines and comments
-            std::string iss = line.c_str();
-            if (iss.find("launch") != std::string::npos)
-            {
-                if (!first_launch)
-                {
-                    first_launch = true;
-                }
-                else
-                {
-                    this->cards.emplace_back();
-                    curr_launch++;
-                }
-                size_t name_pos = iss.find("rwgt_name");
-                if (name_pos != std::string::npos)
-                {
-                    std::string curr_name = iss.substr(name_pos + 10);
-                    auto name_lines = REX::blank_splitter(curr_name);
-                    this->cards.back().launch_name = std::string(name_lines[0]);
-                }
-                continue;
-            }
-            if (iss.find("set") != std::string::npos)
-            {
-                auto rwgt_line = REX::blank_splitter(iss);
-                if (rwgt_line[0] != "set")
-                    continue;
-                if (rwgt_line.size() < 3)
-                    continue;
-                if (rwgt_line.size() == 3)
-                    throw std::runtime_error("rwgt_slha::parse_rwgt_card: \"set\" line appears to use parameter names. This is not supported by teaRex.");
-                if (!first_launch)
-                {
-                    REX::warning("rwgt_slha::parse_rwgt_card: \"set\" line appears before first launch command. Assuming launch command missed. May end up appending meaningless reweighting iterations.");
-                    first_launch = true;
-                }
-                size_t curr_ind_of_line = 1;
-                if (rwgt_line[curr_ind_of_line] == "param_card")
-                    ++curr_ind_of_line;
-                std::string curr_block = std::string(rwgt_line[curr_ind_of_line]);
-                ++curr_ind_of_line;
-                int curr_param = REX::ctoi(rwgt_line[curr_ind_of_line]);
-                ++curr_ind_of_line;
-                double curr_val = REX::ctod(rwgt_line[curr_ind_of_line]);
-                this->cards[curr_launch].add_param(curr_block, curr_param, curr_val);
-                this->cards[curr_launch].rwgt_com += line + "\n";
-                continue;
-            }
-        }
-    }
-
-    rwgt_slha &rwgt_slha::set_card_path(const std::string &path)
-    {
-        this->card_path = path;
-        return *this;
-    }
-
-    rwgt_slha &rwgt_slha::set_mod_card_path(const std::string &path)
-    {
-        this->mod_card_path = path;
-        return *this;
-    }
-
-    bool rwgt_slha::move_param_card(const std::string &new_path)
-    {
-        if (this->card_path.empty())
-            throw std::runtime_error("rwgt_slha::move_param_card: card_path not set");
-        if (new_path.empty() && this->mod_card_path.empty())
-        {
-            this->mod_card_path = this->card_path + ".mod";
-        }
-        else if (!new_path.empty())
-        {
-            this->mod_card_path = new_path;
-        }
-        try
-        {
-            std::filesystem::rename(this->card_path, this->mod_card_path);
-        }
-        catch (const std::filesystem::filesystem_error &e)
-        {
-            throw std::runtime_error("rwgt_slha::move_param_card: failed to rename original card to .mod backup: " + std::string(e.what()));
-        }
-        return true;
-    }
-
-    bool rwgt_slha::remove_mod_card()
-    {
-        if (this->card_path.empty())
-            throw std::runtime_error("rwgt_slha::remove_mod_card: card_path not set");
-        if (this->mod_card_path.empty())
-            throw std::runtime_error("rwgt_slha::remove_mod_card: mod_card_path not set");
-        try
-        {
-            std::filesystem::remove(this->card_path);
-        }
-        catch (const std::filesystem::filesystem_error &e)
-        {
-            throw std::runtime_error("rwgt_slha::remove_mod_card: failed to remove mod card: " + std::string(e.what()));
-        }
-        try
-        {
-            std::filesystem::rename(this->mod_card_path, this->card_path);
-        }
-        catch (const std::filesystem::filesystem_error &e)
-        {
-            throw std::runtime_error("rwgt_slha::remove_mod_card: failed to rename .mod backup to original card: " + std::string(e.what()));
-        }
-        return true;
-    }
-
-    bool rwgt_slha::write_rwgt_card(size_t idx)
-    {
-        if (idx >= this->cards.size())
-            throw std::out_of_range("rwgt_slha::write_rwgt_card: index out of range");
-        if (this->card_path.empty())
-            throw std::runtime_error("rwgt_slha::write_rwgt_card: card_path not set");
-        if (this->mod_card_path.empty())
-            this->mod_card_path = this->card_path + ".mod";
-        if (!std::filesystem::exists(this->mod_card_path))
-        {
-            this->move_param_card();
-        }
-        if (!this->orig_params.blocks.empty())
-        {
-            for (const auto &[key, val] : this->orig_params.blocks)
-            {
-                if (REX::to_upper(key) == "DECAY")
-                {
-                    for (const auto &[pid, width] : val.params)
-                    {
-                        this->REX::slha::set_decay(pid, width);
-                    }
-                }
-                else
-                {
-                    for (auto [param_id, param_value] : val.params)
-                    {
-                        this->REX::slha::set(key, param_id, param_value);
-                    }
-                }
-            }
-            this->orig_params.blocks.clear();
-        }
-        for (const auto &[key, val] : this->cards[idx].blocks)
-        {
-            if (REX::to_upper(key) == "DECAY")
-            {
-                for (const auto &[pid, width] : val.params)
-                {
-                    auto curr_decay = this->REX::slha::get_decay(pid);
-                    this->orig_params.add_param("DECAY", pid, curr_decay);
-                    this->REX::slha::set_decay(pid, width);
-                }
-            }
-            else
-            {
-                for (auto [param_id, param_value] : val.params)
-                {
-                    auto curr_param = this->REX::slha::get(key, param_id);
-                    this->orig_params.add_param(key, param_id, curr_param);
-                    this->REX::slha::set(key, param_id, param_value);
-                }
-            }
-        }
-        std::ofstream ofs(this->card_path);
-        if (!ofs)
-            throw std::runtime_error("rwgt_slha::write_rwgt_card: failed to open file");
-        this->REX::slha::write(ofs);
-        return true;
-    }
-
-    std::vector<std::function<bool()>> rwgt_slha::get_card_writers()
-    {
-        std::vector<std::function<bool()>> writers;
-        for (size_t i = 0; i < this->cards.size(); ++i)
-        {
-            writers.push_back([this, i]()
-                              { return this->write_rwgt_card(i); });
-        }
-        return writers;
-    }
-
-    std::vector<std::string> rwgt_slha::get_launch_names()
-    {
-        std::vector<std::string> names;
-        for (const auto &card : this->cards)
-        {
-            names.push_back(card.launch_name);
-        }
-        return names;
-    }
-
-    std::vector<std::string> rwgt_slha::get_rwgt_commands()
-    {
-        std::vector<std::string> commands;
-        for (const auto &card : this->cards)
-        {
-            commands.push_back(card.rwgt_com);
-        }
-        return commands;
-    }
-
-    threadPool::threadPool(unsigned nthreads)
-        : stop_(false), active_(0)
-    {
-        workers_.reserve(nthreads);
-        for (unsigned i = 0; i < nthreads; ++i)
-        {
-            workers_.emplace_back([this]
-                                  {
-                for (;;) {
-                    Task task;
-                    {
-                        std::unique_lock<std::mutex> lk(m_);
-                        cv_.wait(lk, [this]{ return stop_ || !q_.empty(); });
-                        if (stop_ && q_.empty()) return;
-                        task = std::move(q_.front());
-                        q_.pop();
-                        ++active_;
-                    }
-                    try {
-                        task();
-                    } catch (...) {
-                        // Record first exception and signal cancellation
-                        {
-                            std::lock_guard<std::mutex> g(err_m_);
-                            if (!first_error_) first_error_ = std::current_exception();
-                        }
-                        cancel_.store(true, std::memory_order_relaxed);
-                    }
-                    {
-                        std::lock_guard<std::mutex> lk(m_);
-                        --active_;
-                        if (q_.empty() && active_ == 0) drained_.notify_all();
-                    }
-                } });
-        }
-    }
-
-    threadPool::~threadPool()
-    {
-        {
-            std::lock_guard<std::mutex> lk(m_);
-            stop_ = true;
-        }
-        cv_.notify_all();
-        for (auto &t : workers_)
-            t.join();
-    }
-
-    void threadPool::enqueue(Task t)
-    {
-        {
-            std::lock_guard<std::mutex> lk(m_);
-            q_.push(std::move(t));
-        }
-        cv_.notify_one();
-    }
-
-    void threadPool::begin_batch()
-    {
-        cancel_.store(false, std::memory_order_relaxed);
-        std::lock_guard<std::mutex> g(err_m_);
-        first_error_ = nullptr;
-    }
-
-    void threadPool::wait_batch()
-    {
-        std::unique_lock<std::mutex> lk(m_);
-        drained_.wait(lk, [this]
-                      { return (q_.empty() && active_ == 0) || first_error_; });
-        lk.unlock();
-        if (first_error_)
-            std::rethrow_exception(first_error_);
-    }
-
-    bool threadPool::cancel_requested() const noexcept
-    {
-        return cancel_.load(std::memory_order_relaxed);
-    }
-
-    procReweightor::procReweightor(weightor reweight_function)
-    {
-        this->reweight_functions.push_back(reweight_function);
-    }
-
-    procReweightor::procReweightor(weightor reweight_function, eventBelongs selector)
-    {
-        this->reweight_functions.push_back(reweight_function);
-        this->event_checker = std::make_shared<eventBelongs>(std::move(selector));
-        this->event_checker_fn = this->event_checker->get_event_bool();
-    }
-
-    procReweightor::procReweightor(weightor reweight_function, std::shared_ptr<eventBelongs> selector)
-    {
-        this->reweight_functions.push_back(reweight_function);
-        this->event_checker = selector;
-        this->event_checker_fn = selector->get_event_bool();
-    }
-
-    procReweightor::procReweightor(std::vector<weightor> rwgts)
-    {
-        this->reweight_functions = rwgts;
-    }
-
-    procReweightor::procReweightor(std::vector<weightor> rwgts, std::shared_ptr<eventBelongs> selector)
-    {
-        this->reweight_functions = rwgts;
-        this->event_checker = selector;
-        this->event_checker_fn = selector->get_event_bool();
-    }
-
-    procReweightor::procReweightor(std::vector<weightor> rwgts, eventBelongs selector)
-    {
-        this->reweight_functions = rwgts;
-        this->event_checker = std::make_shared<eventBelongs>(std::move(selector));
-        this->event_checker_fn = this->event_checker->get_event_bool();
-    }
-
-    procReweightor::procReweightor(std::vector<weightor> rwgts, eventBelongs selector, weightor normaliser)
-    {
-        this->reweight_functions = rwgts;
-        this->event_checker = std::make_shared<eventBelongs>(std::move(selector));
-        this->event_checker_fn = this->event_checker->get_event_bool();
-        this->normaliser = normaliser;
-    }
-
-    procReweightor &procReweightor::set_event_checker(eventBelongs checker)
-    {
-        this->event_checker = std::make_shared<eventBelongs>(std::move(checker));
-        this->event_checker_fn = this->event_checker->get_event_bool();
-        return *this;
-    }
-
-    procReweightor &procReweightor::set_event_checker(REX::event_bool_fn checker)
-    {
-        this->event_checker = nullptr;
-        this->event_checker_fn = checker;
-        return *this;
-    }
-
-    procReweightor &procReweightor::set_normaliser(weightor normaliser)
-    {
-        this->normaliser = normaliser;
-        return *this;
-    }
-
-    procReweightor &procReweightor::set_reweight_functions(weightor rwgt)
-    {
-        this->reweight_functions = {rwgt};
-        if (!this->normaliser)
-            this->normaliser = rwgt;
-        return *this;
-    }
-
-    procReweightor &procReweightor::set_reweight_functions(std::vector<weightor> rwgts)
-    {
-        this->reweight_functions = rwgts;
-        return *this;
-    }
-
-    procReweightor &procReweightor::add_reweight_function(weightor rwgt)
-    {
-        this->reweight_functions.push_back(rwgt);
-        return *this;
-    }
-
-    procReweightor &procReweightor::set_process(std::shared_ptr<process> p)
-    {
-        this->proc = p;
-        return *this;
-    }
-
-    // Member functions for handling reweighting
-    void procReweightor::initialise()
-    {
-        if (!this->proc)
-            throw std::runtime_error("procReweightor::initialise: process not set before initialisation");
-        if (!this->normaliser)
-        {
-            if (this->reweight_functions.empty())
-            {
-                warning("procReweightor::initialise: no reweight functions set, process will only yield zero weights.");
-                this->normalisation = std::vector<double>(this->proc->weight_.size(), 0.0);
-                return;
-            }
-            if (this->reweight_functions.size() != 1)
-                warning("procReweightor::initialise: multiple reweight functions set, assuming first is default evaluator and using it for normalisation.");
-            this->normaliser = this->reweight_functions[0];
-        }
-        auto normalised = this->normaliser(*this->proc);
-        if (!normalised)
-            throw std::runtime_error("procReweightor::initialise: normaliser function returned null pointer");
-        if (normalised->size() != this->proc->weight_.size())
-            throw std::runtime_error("procReweightor::initialise: normalisation vector size does not match number of original weights in process");
-        this->normalisation = *normalised;
-        std::transform(this->normalisation.begin(), this->normalisation.end(), this->normalisation.begin(),
-                       [](double val)
-                       { return (val == 0.0) ? 0.0 : 1.0 / val; });
-        this->normalisation = *REX::vec_elem_mult<double>(this->normalisation, this->proc->weight_);
-    }
-
-    void procReweightor::initialise(std::shared_ptr<process> p)
-    {
-        this->proc = p;
-        initialise();
-    }
-
-    void procReweightor::evaluate()
-    {
-        return this->evaluate(0);
-    }
-
-    void procReweightor::evaluate(size_t amp)
-    {
-        if (!this->proc)
-            throw std::runtime_error("procReweightor::evaluate: process not set before evaluation");
-        if (this->reweight_functions.size() <= amp)
-            return this->append_zero_weights();
-        if (this->normalisation.empty())
-            this->initialise();
-        auto newweights = this->reweight_functions[amp](*this->proc);
-        if (!newweights)
-            throw std::runtime_error("procReweightor::evaluate: reweight function returned null pointer");
-        this->backlog.push_back(std::move(*newweights));
-    }
-
-    void procReweightor::append_zero_weights()
-    {
-        if (!this->proc)
-            throw std::runtime_error("procReweightor::append_zero_weights: process not set before appending zero weights");
-        this->backlog.push_back(std::vector<double>(this->proc->weight_.size(), 0.0));
-    }
-
-    void procReweightor::append_backlog()
-    {
-        if (this->normalisation.empty())
-            throw std::runtime_error("procReweightor::append_backlog: normalisation is empty; call initialise() first");
-
-        for (auto &weights : this->backlog)
-        {
-            if (weights.size() != this->normalisation.size())
-                throw std::runtime_error("procReweightor::append_backlog: size mismatch between weights and normalisation");
-
-            this->proc->append_wgts(*REX::vec_elem_mult<double>(weights, this->normalisation));
-        }
-        this->backlog.clear();
-    }
-
-    reweightor::reweightor(lhe &&mother) : lhe(std::move(mother)) {}
-
-    reweightor::reweightor(const lhe &mother) : lhe(mother) {}
-
-    reweightor::reweightor(lhe &&mother, std::vector<std::shared_ptr<procReweightor>> rws) : lhe(std::move(mother)), reweightors(rws) {}
-
-    reweightor::reweightor(const lhe &mother, std::vector<std::shared_ptr<procReweightor>> rws) : lhe(mother), reweightors(rws) {}
-
-    reweightor::reweightor(lhe &&mother, std::vector<std::shared_ptr<procReweightor>> rws, std::vector<iterator> iters) : lhe(std::move(mother)), reweightors(rws), iterators(std::move(iters)) {}
-
-    reweightor::reweightor(const lhe &mother, std::vector<std::shared_ptr<procReweightor>> rws, std::vector<iterator> iters) : lhe(mother), reweightors(rws), iterators(std::move(iters)) {}
-
-    reweightor::reweightor(lhe &&mother, std::vector<procReweightor> rws) : lhe(std::move(mother))
-    {
-        this->set_reweightors(rws);
-    }
-
-    reweightor::reweightor(const lhe &mother, std::vector<procReweightor> rws) : lhe(mother)
-    {
-        this->set_reweightors(rws);
-    }
-
-    reweightor::reweightor(lhe &&mother, std::vector<procReweightor> rws, std::vector<iterator> iters) : lhe(std::move(mother)), iterators(std::move(iters))
-    {
-        this->set_reweightors(rws);
-    }
-
-    reweightor::reweightor(const lhe &mother, std::vector<procReweightor> rws, std::vector<iterator> iters) : lhe(mother), iterators(std::move(iters))
-    {
-        this->set_reweightors(rws);
-    }
-
-    reweightor &reweightor::set_reweightors(std::vector<std::shared_ptr<procReweightor>> rws)
-    {
-        this->reweightors = rws;
-        return *this;
-    }
-
-    reweightor &reweightor::set_reweightors(std::vector<procReweightor> rws)
-    {
-        this->reweightors.clear();
-        for (auto &rw : rws)
-        {
-            this->reweightors.push_back(std::make_shared<procReweightor>(std::move(rw)));
-        }
-        return *this;
-    }
-
-    reweightor &reweightor::add_reweightor(procReweightor &rw)
-    {
-        this->reweightors.push_back(std::make_shared<procReweightor>(rw));
-        return *this;
-    }
-
-    reweightor &reweightor::add_reweightor(procReweightor &&rw)
-    {
-        this->reweightors.push_back(std::make_shared<procReweightor>(std::move(rw)));
-        return *this;
-    }
-
-    reweightor &reweightor::add_reweightor(std::shared_ptr<procReweightor> rw)
-    {
-        this->reweightors.push_back(rw);
-        return *this;
-    }
-
-    reweightor &reweightor::set_initialise(iterator init)
-    {
-        this->initialise = init;
-        return *this;
-    }
-
-    reweightor &reweightor::set_finalise(iterator fin)
-    {
-        this->finalise = fin;
-        return *this;
-    }
-
-    reweightor &reweightor::set_iterators(const std::vector<iterator> &iters)
-    {
-        this->iterators = iters;
-        return *this;
-    }
-
-    reweightor &reweightor::add_iterator(const iterator &iter)
-    {
-        this->iterators.push_back(iter);
-        return *this;
-    }
-
-    reweightor &reweightor::add_iterator(iterator &&iter)
-    {
-        this->iterators.push_back(std::move(iter));
-        return *this;
-    }
-
-    reweightor &reweightor::set_launch_names(const std::vector<std::string> &names)
-    {
-        this->launch_names = names;
-        return *this;
-    }
-
-    reweightor &reweightor::add_launch_name(const std::string &name)
-    {
-        this->launch_names.push_back(name);
-        return *this;
-    }
-
-    void reweightor::calc_norm()
-    {
-        if (this->events.empty())
-            throw std::runtime_error("reweightor::calc_norm: no events loaded, cannot calculate norm");
-        this->norm_factor = 1.0;
-        if (std::abs(this->idWgt_) == 3)
-        {
-            this->norm_factor = std::accumulate(this->xSec_.begin(), this->xSec_.end(), 0.0);
-            this->norm_factor /= this->events.size();
-        }
-        else if (std::abs(this->idWgt_) == 4)
-        {
-            this->norm_factor = 1. / this->events.size();
-        }
-        else
-        {
-            if (std::abs(this->idWgt_) > 2 || this->idWgt_ == 0)
-                warning("reweightor::calc_norm: idWgt is not set to a value defined in the LHE standard. Assuming weighted events.");
-            this->norm_factor = std::accumulate(this->xSec_.begin(), this->xSec_.end(), 0.0);
-            double accumulated_wgts = 0.0;
-            for (const auto &proc : this->processes)
-            {
-                accumulated_wgts += std::accumulate(proc->weight_.begin(), proc->weight_.end(), 0.0);
-            }
-            if (accumulated_wgts == 0.0)
-            {
-                for (auto ev : this->events)
-                {
-                    accumulated_wgts += ev->weight_;
-                }
-            }
-            if (accumulated_wgts == 0.0)
-                throw std::runtime_error("reweightor::calc_norm: total weight is zero, cannot calculate norm");
-            this->norm_factor /= accumulated_wgts;
-        }
-    }
-
-    void reweightor::set_norm(double norm)
-    {
-        this->norm_factor = norm;
-    }
-
-    void reweightor::setup_pool()
-    {
-        if (!pool)
-        {
-            unsigned hc = std::thread::hardware_concurrency();
-            if (hc == 0)
-                hc = 1;
-            unsigned want = this->pool_threads ? this->pool_threads : hc;
-            unsigned n = static_cast<unsigned>(
-                std::max<size_t>(1, std::min<size_t>(reweightors.size(), want)));
-            this->pool = std::make_unique<threadPool>(n);
-        }
-    }
-
-    void reweightor::extract_sorter()
-    {
-        if (this->reweightors.empty())
-            throw std::runtime_error("reweightor::extract_sorter: no procReweightors set in reweightor");
-
-        std::vector<event_bool_fn> preds;
-        preds.reserve(this->reweightors.size());
-        for (const auto &rw : this->reweightors)
-        {
-            if (rw->event_checker_fn)
-            {
-                preds.push_back(rw->event_checker_fn);
-            }
-            else
-            {
-                preds.push_back(rw->event_checker->get_event_bool());
-            }
-        }
-
-        this->set_sorter(eventSorter(std::move(preds)));
-        this->sorted_events.clear();
-        this->processes.clear();
-        this->sort_events();
-        this->events_to_processes();
-
-        const size_t R = this->reweightors.size();
-        const size_t B = this->processes.size();
-        const bool has_unsorted = (B == R + 1);
-        auto processes_full = this->processes;
-
-        std::vector<size_t> keep;
-        keep.reserve(R);
-        for (size_t i = 0; i < R; ++i)
-            if (!processes_full[i]->events.empty())
-                keep.push_back(i);
-
-        std::vector<std::shared_ptr<process>> procs;
-        std::vector<std::shared_ptr<procReweightor>> rwgs;
-        procs.reserve(keep.size() + (has_unsorted ? 1u : 0u));
-        rwgs.reserve(keep.size() + (has_unsorted ? 1u : 0u));
-
-        for (size_t i : keep)
-        {
-            procs.push_back(processes_full[i]);
-            rwgs.push_back(this->reweightors[i]);
-        }
-
-        if (has_unsorted && !processes_full.back()->events.empty())
-        {
-            procs.push_back(processes_full.back());
-
-            auto dummy = std::make_shared<procReweightor>();
-            rwgs.push_back(std::move(dummy));
-        }
-
-        this->processes = std::move(procs);
-        this->reweightors = std::move(rwgs);
-
-        if (this->processes.size() != this->reweightors.size())
-            throw std::runtime_error("reweightor::extract_sorter: number of processes does not match number of reweightors.");
-
-        for (size_t i = 0; i < this->reweightors.size(); ++i)
-        {
-            auto &p = this->processes[i];
-            p->validate();
-            this->reweightors[i]->set_process(p);
-        }
-    }
-
-    void reweightor::initialise_reweightors()
-    {
-        if (this->reweightors.size() != this->processes.size())
-            throw std::runtime_error("initialise_reweightors: reweightors/processes size mismatch");
-
-        for (size_t i = 0; i < this->reweightors.size(); ++i)
-            this->reweightors[i]->initialise(this->processes[i]);
-    }
-
-    void reweightor::finalise_reweighting()
-    {
-        for (auto proc : this->processes)
-        {
-            proc->transpose_wgts();
-            proc->validate();
-        }
-        if (!this->finalise())
-            warning("reweightor::finalise_reweighting: finalise iterator returned false, something might have gone wrong. Validate output manually.");
-        if (this->launch_names.size() > 0)
-        {
-            this->extract_weight_ids();
-            size_t nWgts = this->weight_ids->size();
-            for (size_t i = 0; i < this->launch_names.size(); ++i)
-            {
-                std::string curr_name = (this->launch_names[i].empty()) ? "rwgt_" + std::to_string(i + nWgts + 1) : this->launch_names[i];
-                this->weight_ids->push_back(curr_name);
-            }
-        }
-        this->calc_xSecs();
-        this->calc_xErrs();
-    }
-
-    void reweightor::setup()
-    {
-        if (!this->initialise())
-            throw std::runtime_error("reweightor::setup: initialise iterator returned false, something went wrong.");
-        this->extract_sorter();
-        this->initialise_reweightors();
-        this->n_amps = 0;
-        for (auto &rwgt : this->reweightors)
-        {
-            size_t amps = rwgt->reweight_functions.size();
-            this->n_amps = std::max(this->n_amps, amps);
-        }
-
-        if (this->n_amps == 0)
-        {
-            throw std::runtime_error("reweightor::setup: no reweight functions found, something went wrong.");
-        }
-        this->setup_pool();
-    }
-
-    void reweightor::run_iteration()
-    {
-        if (!this->iterators[this->curr_iter]())
-            throw std::runtime_error("reweightor::run_iteration: iterator returned false, something went wrong in iteration " + std::to_string(this->curr_iter) + ".");
-        this->curr_iter++;
-
-        // Nothing to do?
-        const size_t N = this->reweightors.size();
-        if (N == 0 || this->n_amps == 0)
-            return;
-
-        // Ensure pool exists (persisted across iterations)
-        setup_pool();
-
-        // Parallel "reweightor" phase
-        pool->begin_batch();
-        for (size_t i = 0; i < N; ++i)
-        {
-            pool->enqueue([this, i]
-                          {
-                    // Early cancel check (best-effort)
-                    if (pool->cancel_requested()) return;
-        
-                    auto &rwgt = this->reweightors[i];
-                    for (size_t amp = 0; amp < this->n_amps; ++amp) {
-                        // Safe because each task owns a distinct rwgt
-                        rwgt->evaluate(amp);
-                        if (pool->cancel_requested()) return; // responsive cancellation
-                    } });
-        }
-        // Wait for completion (or rethrow first error)
-        pool->wait_batch();
-    }
-
-    void reweightor::run_all_iterations()
-    {
-        while (this->curr_iter < this->iterators.size())
-        {
-            this->run_iteration();
-            for (auto &rwgt : this->reweightors)
-            {
-                rwgt->append_backlog();
-            }
-#pragma optimize("", off)
-            std::cout << ".";
-#pragma optimize("", on)
-            std::cout.flush();
-        }
-    }
-
-    void reweightor::run()
-    {
-        this->setup();
-        this->run_all_iterations();
-        this->finalise_reweighting();
-    }
-
-    void reweightor::calc_xSecs()
-    {
-        if (this->norm_factor == 0.0)
-            this->calc_norm();
-        this->rwgt_xSec = std::vector<double>(this->events[0]->wgts_.size(), 0.0);
-        for (auto ev : this->events)
-        {
-            for (size_t i = 0; i < ev->wgts_.size(); ++i)
-            {
-                this->rwgt_xSec[i] += ev->wgts_[i];
-            }
-        }
-        for (auto &x : this->rwgt_xSec)
-            x *= this->norm_factor;
-    }
-
-    void reweightor::calc_xErrs()
-    {
-        if (this->rwgt_xSec.size() == 0)
-            this->calc_xSecs();
-        double loc_xSec = std::accumulate(this->xSec_.begin(), this->xSec_.end(), 0.0);
-        double loc_xErr = std::sqrt(std::accumulate(this->xSecErr_.begin(), this->xSecErr_.end(), 0.0, [](double a, double b)
-                                                    { return a + b * b; }));
-        size_t nEvs = this->events.size();
-        if (nEvs == 0)
-        {
-            this->transpose();
-            nEvs = this->events.size();
-            if (nEvs == 0)
-                throw std::runtime_error("reweightor::calc_xErrs: no events found, cannot calculate errors");
-        }
-        this->rwgt_xErr = std::vector<double>(this->rwgt_xSec.size(), 0.0);
-        auto omg = std::vector<double>(this->rwgt_xSec.size(), 0.0);
-        auto omgSq = std::vector<double>(this->rwgt_xSec.size(), 0.0);
-        for (auto ev : this->events)
-        {
-            for (size_t k = 0; k < ev->wgts_.size(); ++k)
-            {
-                double ratio = ev->wgts_[k] / ev->weight_;
-                omg[k] += ratio;
-                omgSq[k] += std::pow(ratio, 2);
-            }
-        }
-        double invNoEvs = 1.0 / double(nEvs);
-        double sqrtInvNoEvs = std::sqrt(invNoEvs);
-        for (size_t k = 0; k < this->rwgt_xSec.size(); ++k)
-        {
-            double variance = (omgSq[k] - std::pow(omg[k], 2) * invNoEvs) * invNoEvs;
-            variance = std::max(variance, 0.0);
-            this->rwgt_xErr[k] = loc_xSec * std::sqrt(variance) * sqrtInvNoEvs + loc_xErr * omg[k] * invNoEvs;
-            if (std::isnan(this->rwgt_xErr[k]) || std::isinf(this->rwgt_xErr[k]) || this->rwgt_xErr[k] <= 0.0)
-            {
-                warning("reweightor::calc_xErrs: Error propagation failed for weight " + std::to_string(k) + ". Approximating the error at the level of the cross section.");
-                this->rwgt_xErr[k] = loc_xErr * std::max(loc_xSec / this->rwgt_xSec[k], this->rwgt_xSec[k] / loc_xSec);
-            }
-        }
-    }
-
-    param_rwgt::param_rwgt(const lhe &mother, std::vector<std::shared_ptr<procReweightor>> rws, const std::string &slha_path, const std::string &rwgt_path)
-        : reweightor(mother, rws)
-    {
-        this->read_slha_rwgt(slha_path, rwgt_path);
-    }
-
-    void param_rwgt::read_slha_rwgt(std::istream &slha_in, std::istream &rwgt_in)
-    {
-        this->card_iter = rwgt_slha::create(slha_in, rwgt_in);
-        this->initialise = [&]()
-        { return this->card_iter.move_param_card(); };
-        this->finalise = [&]()
-        { return this->card_iter.remove_mod_card(); };
-        this->iterators = this->card_iter.get_card_writers();
-        this->launch_names = this->card_iter.get_launch_names();
-        this->weight_context = this->card_iter.get_rwgt_commands();
-    }
-
-    void param_rwgt::read_slha_rwgt(const std::string &slha_file, const std::string &rwgt_file)
-    {
-        std::ifstream slha_in(slha_file);
-        std::ifstream rwgt_in(rwgt_file);
-        if (!slha_in || !rwgt_in)
-            throw std::runtime_error("param_rwgt::read_slha_rwgt: failed to open input files");
-        this->read_slha_rwgt(slha_in, rwgt_in);
-        this->card_iter.set_card_path(slha_file);
-    }
-
-} // namespace REX::tea
-#endif // _TEAREX_CPP_
diff --git a/madmatrix/MadtRex/template_files/rwgt_driver.inc b/madmatrix/MadtRex/template_files/rwgt_driver.inc
deleted file mode 100644
index ff099589b5..0000000000
--- a/madmatrix/MadtRex/template_files/rwgt_driver.inc
+++ /dev/null
@@ -1,201 +0,0 @@
-//==========================================================================
-// Copyright (C) 2023-2026 CERN
-// Licensed under the GNU Lesser General Public License (version 3 or later).
-// Written originally by: Z. Wettersten (Jan 2024) for the MadGraph7 CUDACPP plugin.
-// Integrated with the MadGraph7 project in Feb 2026.
-//==========================================================================
-//==========================================================================
-// This file has been automatically generated for C++ Standalone by
-%(info_lines)s
-//==========================================================================
-//==========================================================================
-// Driver for reweighting events for processes
-%(multiprocess_lines)s
-//--------------------------------------------------------------------------
-
-#include "rwgt_instance.h"
-#include <cstdlib>
-#include <typeinfo>
-#include <memory>
-%(include_lines)s
-
-    int usage(char *argv0, int ret = 1)
-{
-    std::cout << "Usage: " << argv0
-              << " [--lhefile=\"/YOUR/PATH/HERE\"|-lhe=\"/YOUR/PATH/HERE\"] [--rwgtcard=/YOUR/PATH/HERE|-rwgt=\"/YOUR/PATH/HERE\"]\n"
-              << "[--output=/YOUR/PATH/HERE\"|-out=\"/YOUR/PATH/HERE\"]\n"
-              << "[--param_card=/YOUR/PATH/HERE\"|-slha=\"/YOUR/PATH/HERE\"]\n";
-    std::cout << "\n";
-    std::cout << "The LHE file path should be with respect to the directory you are running\n";
-    std::cout << "this program from, and similarly the rwgt_card should be as well.\n";
-    return ret;
-}
-
-void writeRwgtCsv(std::string path, std::shared_ptr<std::vector<std::string>> names, std::shared_ptr<std::vector<double>> xSecs, std::shared_ptr<std::vector<double>> errXSecs)
-{
-    std::ofstream outFile;
-    outFile.open(path);
-    if (!outFile.is_open())
-        throw std::runtime_error("Failed to open output file for writing.");
-    if (names->size() != xSecs->size() || names->size() != errXSecs->size())
-        throw std::runtime_error("Mismatch in number of processes, cross-sections, and errors when logging results.");
-    for (size_t k = 0; k < names->size(); ++k)
-    {
-        outFile << names->at(k) << ", " << xSecs->at(k) << ", " << errXSecs->at(k) << "\n";
-    }
-    outFile.close();
-    return;
-}
-
-void writeRwgtCsv(std::string path, std::vector<std::string> names, std::vector<double> xSecs, std::vector<double> errXSecs)
-{
-    std::ofstream outFile;
-    outFile.open(path);
-    if (!outFile.is_open())
-        throw std::runtime_error("Failed to open output file for writing.");
-    if (names.size() != xSecs.size() || names.size() != errXSecs.size())
-        throw std::runtime_error("Mismatch in number of processes, cross-sections, and errors when logging results.");
-    for (size_t k = 0; k < names.size(); ++k)
-    {
-        outFile << names.at(k) << ", " << xSecs.at(k) << ", " << errXSecs.at(k) << "\n";
-    }
-    outFile.close();
-    return;
-}
-
-int main(int argc, char **argv)
-{
-
-    std::string banner =
-        "#################################################\n"
-        "#    ___  ___          _ _  ______              #\n"
-        "#    |  \\/  |         | | | | ___ \\             #\n"
-        "#    | .  . | __ _  __| | |_| |_/ /_____  __    #\n"
-        "#    | |\\/| |/ _` |/ _` | __|    // _ \\ \\/ /    #\n"
-        "#    | |  | | (_| | (_| | |_| |\\ \\  __/>  <     #\n"
-        "#    \\_|  |_/\\__,_|\\__,_|\\__\\_| \\_\\___/_/\\_\\    #\n"
-        "#                                               #\n"
-        "#        Data-parallel event reweighting        #\n"
-        "#             in MadGraph7              #\n"
-        "#                                               #\n"
-        "#################################################\n";
-
-    std::cout << banner;
-    std::cout << "Starting MadtRex driver...\n";
-    std::string lheFilePath;
-    std::string rwgtCardPath;
-    std::string outputPath;
-    std::string slhaPath;
-    size_t nb_threads = 1;
-    size_t batch_size = 32;
-
-    if (argc < 2)
-    {
-        return usage(argv[0]);
-    }
-
-    for (int i = 1; i < argc; i++)
-    {
-        auto currArg = std::string(argv[i]);
-        if (currArg.substr(0, 9) == "--lhefile" || currArg.substr(0, 4) == "-lhe")
-        {
-            lheFilePath = currArg.substr(currArg.find("=") + 1);
-        }
-        else if (currArg.substr(0, 10) == "--rwgtcard" || currArg.substr(0, 5) == "-rwgt")
-        {
-            rwgtCardPath = currArg.substr(currArg.find("=") + 1);
-        }
-        else if (currArg.substr(0, 8) == "--output" || currArg.substr(0, 4) == "-out")
-        {
-            outputPath = currArg.substr(currArg.find("=") + 1);
-        }
-        else if (currArg.substr(0, 12) == "--param_card" || currArg.substr(0, 5) == "-slha")
-        {
-            slhaPath = currArg.substr(currArg.find("=") + 1);
-        }
-        else if (currArg.substr(0, 10) == "--nb_threads" || currArg.substr(0, 3) == "-nt")
-        {
-            nb_threads = std::stoi(currArg.substr(currArg.find("=") + 1));
-        }
-        else if (currArg.substr(0, 12) == "--batch_size" || currArg.substr(0, 5) == "-warp")
-        {
-            batch_size = std::stoi(currArg.substr(currArg.find("=") + 1));
-        }
-        else
-        {
-            return usage(argv[0]);
-        }
-    }
-
-    if (lheFilePath.empty() || rwgtCardPath.empty())
-    {
-        return usage(argv[0]);
-    }
-
-    std::string currPath = argv[0];
-
-    size_t slashPos = currPath.find_last_of("/");
-    bool onWindows = false;
-    if (slashPos == std::string::npos)
-    {
-        slashPos = currPath.find_last_of("\\");
-        onWindows = true;
-    }
-    if (slashPos == std::string::npos)
-        throw std::runtime_error("Failed to determine current working directory -- need to know where program is run from to identify where to pull and push param_card.dat.");
-
-    if (slhaPath.empty())
-    {
-        if (onWindows)
-        {
-            if (currPath.substr(currPath.find_last_of("\\", slashPos - 1) + 1, 2) == "P1")
-            {
-                slhaPath = "..\\..\\Cards\\param_card.dat";
-            }
-            else if (currPath.substr(currPath.find_last_of("\\", slashPos - 1) + 1, 3) == "Sub")
-            {
-                slhaPath = "..\\Cards\\param_card.dat";
-            }
-            else
-            {
-                slhaPath = "\\Cards\\param_card.dat";
-            }
-        }
-        else
-        {
-            if (currPath.substr(currPath.find_last_of("/", slashPos - 1) + 1, 2) == "P1")
-            {
-                slhaPath = "../../Cards/param_card.dat";
-            }
-            else if (currPath.substr(currPath.find_last_of("/", slashPos - 1) + 1, 3) == "Sub")
-            {
-                slhaPath = "../Cards/param_card.dat";
-            }
-            else
-            {
-                slhaPath = "/Cards/param_card.dat";
-            }
-        }
-    }
-
-    static std::vector<std::shared_ptr<REX::tea::procReweightor>> rwgtRun = {%(make_rwgt)s};
-
-    auto rwgt_runner = REX::tea::param_rwgt(REX::load_lhef(lheFilePath), rwgtRun);
-    rwgt_runner.read_slha_rwgt(slhaPath, rwgtCardPath);
-    rwgt_runner.pool_threads = nb_threads;
-
-    rwgt_runner.run();
-
-    std::cout << "\nReweighting procedure finished.\n";
-
-    std::ofstream lhe_out(outputPath);
-    if (!lhe_out)
-        throw std::runtime_error("Failed to open output LHE file for writing.");
-    rwgt_runner.print(lhe_out, true);
-
-    std::cout << "Reweighted LHE file written to " << outputPath << ".\n";
-
-    writeRwgtCsv("rwgt_results.csv", *rwgt_runner.weight_ids, rwgt_runner.rwgt_xSec, rwgt_runner.rwgt_xErr);
-
-    return 0;
-}
diff --git a/madmatrix/MadtRex/template_files/rwgt_instance.cc b/madmatrix/MadtRex/template_files/rwgt_instance.cc
deleted file mode 100644
index 6c72c99a45..0000000000
--- a/madmatrix/MadtRex/template_files/rwgt_instance.cc
+++ /dev/null
@@ -1,177 +0,0 @@
-//==========================================================================
-// Copyright (C) 2023-2026 CERN
-// Licensed under the GNU Lesser General Public License (version 3 or later).
-// Written originally by: Z. Wettersten (Jan 2024) for the MadGraph7 CUDACPP plugin.
-// Integrated with the MadGraph7 project in Feb 2026.
-//==========================================================================
-//==========================================================================
-// Library including generic functions and classes for event reweighting.
-// Process-specific rwgt_runner files are generated by mg5amc@nlo and use
-// this library, while the rwgt_driver file is a wrapping program that
-// calls the process-specific runners for given subprocesses.
-//==========================================================================
-
-#ifndef _RWGT_INSTANCE_CC_
-#define _RWGT_INSTANCE_CC_
-
-#include "rwgt_instance.h"
-
-namespace rwgt
-{
-
-    // ZW: Function for calculating the number of remaining events in a warp
-    //  in order to pad the input arrays to a multiple of the warp size
-    unsigned int warpRemain(unsigned int nEvt, unsigned int nWarp)
-    {
-        return (nWarp - (nEvt % nWarp)) % nWarp;
-    }
-
-    // ZW: Function for padding the input arrays to a multiple of the warp size
-    template <typename T>
-    void warpPad(std::vector<T> &input, unsigned int nWarp = 32)
-    {
-        auto nEvt = input.size();
-        auto nWarpRemain = warpRemain(nEvt, nWarp);
-        input.reserve(nEvt + nWarpRemain);
-        for (size_t k = nEvt - nWarpRemain; k < nEvt; ++k)
-        {
-            input.push_back(input[k]);
-        }
-        return;
-    }
-
-    fBridge::fBridge(REX::process &proc, unsigned int warpSize)
-    {
-        this->nPar = proc.events[0]->view().size();
-        this->nEvt = proc.events.size();
-        this->nWarp = warpSize;
-        this->nWarpRemain = warpRemain(nEvt, nWarp);
-        this->fauxNEvt = nEvt + nWarpRemain;
-        this->gs = proc.gS();
-        this->rndHel = std::vector<FORTRANFPTYPE>(fauxNEvt, 0.);
-        this->rndCol = std::vector<FORTRANFPTYPE>(fauxNEvt, 0.);
-        this->selHel = std::vector<int>(fauxNEvt, 0.);
-        this->selCol = std::vector<int>(fauxNEvt, 0.);
-        this->goodHel = false;
-    }
-
-    fBridge::fBridge(std::shared_ptr<REX::process> proc, unsigned int warpSize)
-    {
-        if (proc == nullptr)
-        {
-            throw std::runtime_error("fBridge: Provided process is null or contains no subprocesses.");
-        }
-        // Assuming all subprocesses have the same number of particles
-        this->nPar = proc->events[0]->view().size();
-        this->nEvt = proc->events.size();
-        this->nWarp = warpSize;
-        this->nWarpRemain = warpRemain(nEvt, nWarp);
-        this->fauxNEvt = nEvt + nWarpRemain;
-        this->gs = proc->gS();
-        this->rndHel = std::vector<FORTRANFPTYPE>(fauxNEvt, 0.);
-        this->rndCol = std::vector<FORTRANFPTYPE>(fauxNEvt, 0.);
-        this->selHel = std::vector<int>(fauxNEvt, 0.);
-        this->selCol = std::vector<int>(fauxNEvt, 0.);
-        this->goodHel = false;
-    }
-
-    void fBridge::bridgeSetup(unsigned int &noEvts, unsigned int warpSize)
-    {
-        this->nEvt = noEvts;
-        this->nWarp = warpSize;
-        this->nWarpRemain = warpRemain(nEvt, nWarp);
-        this->fauxNEvt = nEvt + nWarpRemain;
-        this->rndHel = std::vector<FORTRANFPTYPE>(fauxNEvt, 0.);
-        this->rndCol = std::vector<FORTRANFPTYPE>(fauxNEvt, 0.);
-        this->selHel = std::vector<int>(fauxNEvt, 0.);
-        this->selCol = std::vector<int>(fauxNEvt, 0.);
-    }
-    void fBridge::bridgeSetup(std::vector<FORTRANFPTYPE> &evVec, unsigned int warpSize)
-    {
-        this->nEvt = evVec.size();
-        this->nWarp = warpSize;
-        this->nWarpRemain = warpRemain(nEvt, nWarp);
-        this->fauxNEvt = nEvt + nWarpRemain;
-        this->rndHel = std::vector<FORTRANFPTYPE>(fauxNEvt, 0.);
-        this->rndCol = std::vector<FORTRANFPTYPE>(fauxNEvt, 0.);
-        this->selHel = std::vector<int>(fauxNEvt, 0.);
-        this->selCol = std::vector<int>(fauxNEvt, 0.);
-    }
-    void fBridge::bridgeSetup(std::shared_ptr<std::vector<FORTRANFPTYPE>> &evVec, unsigned int warpSize)
-    {
-        this->bridgeSetup(*evVec, warpSize);
-    }
-
-    void fBridge::bridgeSetup(REX::process &proc, unsigned int warpSize)
-    {
-        this->nPar = proc.events[0]->view().size();
-        this->nEvt = proc.events.size();
-        this->nWarp = warpSize;
-        this->nWarpRemain = warpRemain(nEvt, nWarp);
-        this->fauxNEvt = nEvt + nWarpRemain;
-        this->gs = proc.gS();
-        this->rndHel = std::vector<FORTRANFPTYPE>(fauxNEvt, 0.);
-        this->rndCol = std::vector<FORTRANFPTYPE>(fauxNEvt, 0.);
-        this->selHel = std::vector<int>(fauxNEvt, 0.);
-        this->selCol = std::vector<int>(fauxNEvt, 0.);
-        this->goodHel = false;
-    }
-
-    void fBridge::bridgeSetup(std::shared_ptr<REX::process> proc, unsigned int warpSize)
-    {
-        this->bridgeSetup(*proc, warpSize);
-    }
-
-    void fBridge::setBridge(bridgeWrapper amp)
-    {
-        if (this->bridge == nullptr)
-        {
-            this->bridge = amp;
-        }
-        else
-            throw std::runtime_error("fBridge object doubly defined.");
-    }
-    std::shared_ptr<std::vector<FORTRANFPTYPE>> fBridge::bridgeCall(std::vector<FORTRANFPTYPE> &momenta)
-    {
-        if (this->nEvt == 0)
-            this->bridgeSetup(this->gs);
-        if (this->bridge == nullptr)
-            throw std::runtime_error("fBridge object not defined.");
-        if (this->gs.size() != fauxNEvt)
-            warpPad(gs, nWarp);
-        warpPad(momenta, nWarp * nPar * nMom);
-        auto evalScatAmps = this->bridge(
-            fauxNEvt, nPar, nMom,
-            momenta, gs, rndHel, rndCol,
-            selHel, selCol, chanId, goodHel);
-        momenta.resize(nEvt * nPar * nMom);
-        evalScatAmps->resize(nEvt);
-        return evalScatAmps;
-    }
-
-    std::shared_ptr<std::vector<FORTRANFPTYPE>> fBridge::bridgeCall(REX::process &proc)
-    {
-        if (this->nEvt == 0)
-            this->bridgeSetup(proc);
-        return this->bridgeCall(proc.momenta_.flat_vector());
-    }
-
-    std::shared_ptr<std::vector<FORTRANFPTYPE>> fBridge::bridgeCall(std::shared_ptr<REX::process> proc)
-    {
-        return this->bridgeCall(proc->momenta_.flat_vector());
-    }
-
-    REX::tea::weightor fBridge::getAmp()
-    {
-        if (this->bridge == nullptr)
-            throw std::runtime_error("fBridge object not defined.");
-        REX::tea::weightor amp = [this](REX::process &process)
-        {
-            return this->bridgeCall(process);
-        };
-        return amp;
-    }
-
-}
-
-#endif
diff --git a/madmatrix/MadtRex/template_files/rwgt_instance.h b/madmatrix/MadtRex/template_files/rwgt_instance.h
deleted file mode 100644
index 22271eec2e..0000000000
--- a/madmatrix/MadtRex/template_files/rwgt_instance.h
+++ /dev/null
@@ -1,75 +0,0 @@
-//==========================================================================
-// Copyright (C) 2023-2026 CERN
-// Licensed under the GNU Lesser General Public License (version 3 or later).
-// Written originally by: Z. Wettersten (Jan 2024) for the MadGraph7 CUDACPP plugin.
-// Integrated with the MadGraph7 project in Feb 2026.
-//==========================================================================
-//==========================================================================
-// Library including generic functions and classes for event reweighting.
-// Process-specific rwgt_runner files are generated by mg5amc@nlo and use
-// this library, while the rwgt_driver file is a wrapping program that
-// calls the process-specific runners for given subprocesses.
-//==========================================================================
-
-#ifndef _RWGT_INSTANCE_H_
-#define _RWGT_INSTANCE_H_
-
-#include "teaRex.h"
-
-/**
- * The floating point precision used in Fortran arrays.
- * This is presently hardcoded to double precision (REAL*8).
- */
-using FORTRANFPTYPE = double; // for Fortran double precision (REAL*8) arrays
-// using FORTRANFPTYPE = float; // for Fortran single precision (REAL*4) arrays
-
-namespace rwgt
-{
-
-    // ZW: Function for calculating the number of remaining events in a warp
-    //  in order to pad the input arrays to a multiple of the warp size
-    unsigned int warpRemain(unsigned int nEvt, unsigned int nWarp = 32);
-
-    // ZW: bridgeWrapper needs args: nEvs, nPar, nMom, moms, gs, rndhel, rndcol, selhel, selcol, chanId
-    using bridgeWrapper = std::function<std::shared_ptr<std::vector<FORTRANFPTYPE>>(int &, int &, int &, std::vector<FORTRANFPTYPE> &, std::vector<FORTRANFPTYPE> &, std::vector<FORTRANFPTYPE> &, std::vector<FORTRANFPTYPE> &, std::vector<int> &, std::vector<int> &, unsigned int &, bool &)>;
-
-    struct fBridge
-    {
-        // Default constructors
-        fBridge() = default;
-        fBridge(const fBridge &source) = default;
-        fBridge(fBridge &&source) = default;
-
-        std::vector<FORTRANFPTYPE> rndHel = {};
-        std::vector<FORTRANFPTYPE> rndCol = {};
-        std::vector<FORTRANFPTYPE> gs = {};
-        std::vector<int> selHel = {};
-        std::vector<int> selCol = {};
-        unsigned int chanId = 0;
-        int nMom = 4;
-        int nWarp;
-        int nWarpRemain;
-        int nEvt;
-        int fauxNEvt;
-        int nPar;
-        bool goodHel = false;
-        bridgeWrapper bridge;
-        fBridge(REX::process &proc, unsigned int warpSize = 32);
-        fBridge(std::shared_ptr<REX::process> proc, unsigned int warpSize = 32);
-        void init(std::vector<REX::event> &process, unsigned int warpSize = 32);
-        void init(std::vector<std::shared_ptr<REX::event>> process, unsigned int warpSize = 32);
-        void bridgeSetup(unsigned int &noEvts, unsigned int warpSize = 32);
-        void bridgeSetup(std::vector<FORTRANFPTYPE> &evVec, unsigned int warpSize = 32);
-        void bridgeSetup(std::shared_ptr<std::vector<FORTRANFPTYPE>> &evVec, unsigned int warpSize = 32);
-        void bridgeSetup(REX::process &proc, unsigned int warpSize = 32);
-        void bridgeSetup(std::shared_ptr<REX::process> proc, unsigned int warpSize = 32);
-        void setBridge(bridgeWrapper amp);
-        std::shared_ptr<std::vector<FORTRANFPTYPE>> bridgeCall(std::vector<FORTRANFPTYPE> &momenta);
-        std::shared_ptr<std::vector<FORTRANFPTYPE>> bridgeCall(REX::process &process);
-        std::shared_ptr<std::vector<FORTRANFPTYPE>> bridgeCall(std::shared_ptr<REX::process> process);
-        REX::tea::weightor getAmp();
-    };
-
-}
-
-#endif
diff --git a/madmatrix/MadtRex/template_files/rwgt_runner_cc.inc b/madmatrix/MadtRex/template_files/rwgt_runner_cc.inc
deleted file mode 100644
index 9cee1e4170..0000000000
--- a/madmatrix/MadtRex/template_files/rwgt_runner_cc.inc
+++ /dev/null
@@ -1,65 +0,0 @@
-//==========================================================================
-// Copyright (C) 2023-2026 CERN
-// Licensed under the GNU Lesser General Public License (version 3 or later).
-// Written originally by: Z. Wettersten (Jan 2024) for the MadGraph7 CUDACPP plugin.
-// Integrated with the MadGraph7 project in Feb 2026.
-//==========================================================================
-//==========================================================================
-// This file has been automatically generated for the CUDACPP plugin by
-%(info_lines)s
-//==========================================================================
-//==========================================================================
-// A class for reweighting matrix elements for
-%(process_lines)s
-//--------------------------------------------------------------------------
-#ifndef _TREX_
-#define _TREX_
-#endif
-#include "rwgt_instance.h"
-#include "fbridge.h"
-
-namespace %(process_namespace)s
-{
-
-    std::shared_ptr<REX::eventBelongs> get_comp()
-    {
-        static std::vector<std::vector<short int>> status = {%(parton_status)s};
-        static std::vector<std::vector<long int>> pdg = {%(parton_ids)s};
-        if (status.size() != pdg.size())
-            throw std::runtime_error("Inconsistent event data in rwgt_runner");
-        static std::vector<std::shared_ptr<REX::event>> loc_evs;
-        for (size_t i = 0; i < status.size(); ++i)
-        {
-            auto ev = std::make_shared<REX::event>(pdg[i].size());
-            ev->set_status(status[i]);
-            ev->set_pdg(pdg[i]);
-            loc_evs.push_back(ev);
-        }
-        return std::make_shared<REX::eventBelongs>(loc_evs, REX::external_legs_comparator);
-    }
-
-    std::shared_ptr<std::vector<FORTRANFPTYPE>> amp(int &nEvt, int &nPar, int &nMom, std::vector<FORTRANFPTYPE> &momenta, std::vector<FORTRANFPTYPE> &alphaS, std::vector<FORTRANFPTYPE> &rndHel, std::vector<FORTRANFPTYPE> &rndCol, std::vector<int> &selHel, std::vector<int> &selCol, unsigned int &chanId, bool &goodHel)
-    {
-        CppObjectInFortran *bridgeInst;
-        auto evalScatAmps = std::make_shared<std::vector<FORTRANFPTYPE>>(nEvt);
-        fbridgecreate_(&bridgeInst, &nEvt, &nPar, &nMom);
-        fbridgesequence_nomultichannel_(&bridgeInst, &momenta.at(0), &alphaS.at(0), &rndHel[0], &rndCol[0], &evalScatAmps->at(0), &selHel[0], &selCol[0], &goodHel);
-        fbridgedelete_(&bridgeInst);
-        return evalScatAmps;
-    }
-
-    rwgt::fBridge bridgeConstr(size_t warpSize)
-    {
-        rwgt::fBridge bridge;
-        bridge.setBridge(amp);
-        bridge.nWarp = warpSize;
-        return bridge;
-    }
-
-    std::shared_ptr<REX::tea::procReweightor> make_reweightor(size_t warpSize)
-    {
-        static rwgt::fBridge bridge = bridgeConstr(warpSize);
-        auto comp = get_comp();
-        return std::make_shared<REX::tea::procReweightor>(bridge.getAmp(), comp);
-    }
-}
diff --git a/madmatrix/MadtRex/template_files/rwgt_runner_h.inc b/madmatrix/MadtRex/template_files/rwgt_runner_h.inc
deleted file mode 100644
index 96db8c26a3..0000000000
--- a/madmatrix/MadtRex/template_files/rwgt_runner_h.inc
+++ /dev/null
@@ -1,32 +0,0 @@
-//==========================================================================
-// Copyright (C) 2023-2026 CERN
-// Licensed under the GNU Lesser General Public License (version 3 or later).
-// Written originally by: Z. Wettersten (June 2024) for the MadGraph7 CUDACPP plugin.
-// Integrated with the MadGraph7 project in Feb 2026.
-//==========================================================================
-//==========================================================================
-// This file has been automatically generated for the CUDACPP plugin by
-%(info_lines)s
-//==========================================================================
-//==========================================================================
-// A class for reweighting matrix elements for
-%(process_lines)s
-//--------------------------------------------------------------------------
-
-#ifndef _%(process_namespace)s_RUNNER_H_
-#define _%(process_namespace)s_RUNNER_H_
-
-#include "rwgt_instance.h"
-
-namespace %(process_namespace)s {
-
-    REX::eventBelongs get_comp();
-    std::shared_ptr<std::vector<FORTRANFPTYPE>> amp( int& nEvt, int& nPar, int& nMom, std::vector<FORTRANFPTYPE>& momenta, std::vector<FORTRANFPTYPE>& alphaS, std::vector<FORTRANFPTYPE>& rndHel, std::vector<FORTRANFPTYPE>& rndCol, std::vector<int>& selHel, std::vector<int>& selCol, int& chanId, bool& goodHel );
-    rwgt::fBridge bridgeConstr(size_t warpSize = 32);
-    std::shared_ptr<REX::tea::procReweightor> make_reweightor(size_t warpSize = 32);
-
-}
-
-
-
-#endif
diff --git a/madmatrix/__init__.py b/madmatrix/__init__.py
index a02a01ba4c..5ee4bfa93a 100644
--- a/madmatrix/__init__.py
+++ b/madmatrix/__init__.py
@@ -30,9 +30,8 @@
     #    Example: new_output = {'myformat': MYCLASS}
     #    allows the command "output myformat PATH" in madgraph.
     #    MYCLASS should inherit from class madgraph.iolibs.export_v4.VirtualExporter
-    from . import output, trex
-    new_output = { 'standalone_trex' : trex.TREX_ProcessExporter }
-    new_reweight = {'madtrex': trex.TREX_ReweightInterface}
+    from . import output
+    new_output = {}
 
     # 2. Define new way to handle the cluster.
     #    Example: new_cluster = {'mycluster': MYCLUSTERCLASS}
diff --git a/madmatrix/density.py b/madmatrix/density.py
index 09aa7d0117..b23b05804b 100644
--- a/madmatrix/density.py
+++ b/madmatrix/density.py
@@ -38,6 +38,10 @@
 
 PROCESSES_FILE = 'madmatrix_processes.json'
 
+# The FPTYPE of the libraries this module builds and loads: the default of
+# madmatrix.mk, made explicit since it is part of the library names
+FPTYPE = 'm'
+
 
 class MadDensityError(Exception):
     pass
@@ -65,7 +69,7 @@ def make_variables(backend='scalar', libtag=''):
     the standalone check_sa.exe) when make runs in SubProcesses. DENSITY=1
     adds Density.o, which only this module calls: other builds of the same
     output (check_sa, mg7) leave it out."""
-    variables = ['PGOAL=lib', 'BACKEND=%s' % backend, 'DENSITY=1']
+    variables = ['PGOAL=lib', 'BACKEND=%s' % backend, 'FPTYPE=%s' % FPTYPE, 'DENSITY=1']
     if libtag:
         variables.append('LIBTAG=%s' % libtag)
     return variables
@@ -76,11 +80,30 @@ def build_command(nb_core=1, backend='scalar', libtag=''):
     return ['make', '-j%d' % max(1, int(nb_core))] + make_variables(backend, libtag)
 
 
+def common_library(output_dir):
+    """The name of the common library of output_dir, with '{libtag}', '{backend}'
+    and '{fptype}' left open, as its process table gives it (the name without the
+    model and FPTYPE for an output whose table predates them)."""
+    try:
+        with open(os.path.join(output_dir, 'SubProcesses', PROCESSES_FILE)) as fsock:
+            data = json.load(fsock)
+    except (IOError, OSError, ValueError):
+        data = {}
+    return data.get('common_library', 'libmadmatrix_common{libtag}_{backend}.so')
+
+
+def common_library_path(output_dir, libtag, backend='scalar'):
+    """The path of the common library of output_dir built with libtag."""
+    return os.path.join(output_dir, 'lib', common_library(output_dir).format(
+        libtag=libtag, backend=backend, fptype=FPTYPE))
+
+
 def find_libtag(output_dir, backend='scalar'):
     """The LIBTAG the libraries of output_dir were built with, read off the
     name of its common library."""
     lib = os.path.join(output_dir, 'lib')
-    prefix, suffix = 'libmadmatrix_common', '_%s.so' % backend
+    prefix, suffix = common_library(output_dir).format(
+        libtag='{libtag}', backend=backend, fptype=FPTYPE).split('{libtag}')
     found = sorted(name[len(prefix):-len(suffix)] for name in
                    (os.listdir(lib) if os.path.isdir(lib) else [])
                    if name.startswith(prefix) and name.endswith(suffix))
@@ -224,8 +247,7 @@ def __init__(self, output_dir, backend='scalar', libtag=None):
         self.libtag = libtag
         # the common library first, by its path: the process libraries only name
         # it, and the check of _load has to see it
-        _load(os.path.join(output_dir, 'lib',
-                           'libmadmatrix_common%s_%s.so' % (libtag, backend)))
+        _load(common_library_path(output_dir, libtag, backend))
         self.subprocesses = []
         # raw pdg combination (leg order of the process) -> (subprocess, flavor index)
         self._by_pdgs = {}
@@ -236,7 +258,8 @@ def __init__(self, output_dir, backend='scalar', libtag=None):
         self._process_slots = []
         for entry in entries:
             lib_path = os.path.join(output_dir, 'lib',
-                                    entry['library'].format(libtag=libtag, backend=backend))
+                                    entry['library'].format(libtag=libtag, backend=backend,
+                                                            fptype=FPTYPE))
             if not os.path.exists(lib_path):
                 raise MadDensityError('missing %s: build the libraries first (%s in %s)'
                                       % (lib_path, ' '.join(build_command(backend=backend,
diff --git a/madmatrix/model_handling.py b/madmatrix/model_handling.py
index 3cbac6a3e2..42eae5c79d 100644
--- a/madmatrix/model_handling.py
+++ b/madmatrix/model_handling.py
@@ -1359,7 +1359,29 @@ def write_flv_couplings(self, params):
 
     # AV - overload export_cpp.UFOModelConverterCPP method (improve formatting)
     def write_set_parameters(self, params):
-        res = self.super_write_set_parameters_donotfixMajorana(params)
+        # Independent, SLHA-card-level parameters (self.params_ext_names) get
+        # an inline check for a runtime override (e.g. set by name via the
+        # UMAMI set_parameter interface) ahead of their normal SLHA-driven
+        # assignment; every other parameter/coupling is unaffected. NB: this
+        # must NOT be folded into super_write_set_parameters_donotfixMajorana,
+        # since write_hardcoded_parameters() calls that method directly and
+        # requires each parameter's expr to remain a single plain assignment.
+        res_strings = []
+        for param in params:
+            # 'aS' is excluded here even though it is SLHA-card-level: its
+            # assignment is entirely commented out below (by name-matching on
+            # a plain "  aS = ..." line), since this backend always retrieves
+            # alphaS event-by-event (as G) rather than from the param card.
+            # Wrapping it in an override check would break that name-match.
+            if param.name in self.params_ext_names and param.name != 'aS':
+                res_strings.append(
+                    'if (m_param_overrides.find("%s") != m_param_overrides.end())\n'
+                    '%s = m_param_overrides["%s"];\nelse\n%s' %
+                    (param.name, param.name, param.name, param.expr))
+            else:
+                res_strings.append(param.expr)
+        res = "\n".join(res_strings)
+        res = res.replace('ABS(','std::abs(') # for SMEFT #614 and #616
         res = res.replace('(','( ')
         res = res.replace(')',' )')
         res = res.replace('+',' + ')
@@ -1598,6 +1620,31 @@ def super_generate_parameters_class_files(self):
         replace_dict['set_independent_couplings'] = self.write_set_parameters(self.coups_indep)
         replace_dict['set_dependent_parameters'] = self.write_set_parameters(self.params_dep)
         replace_dict['set_dependent_couplings'] = self.write_set_parameters(list(self.coups_dep.values()))
+        # NB: unlike the plain (mg7_v5) backend, dependent (aS-dependent) parameters
+        # and couplings are NOT persistent members of this Parameters class at all
+        # (their declarations are commented out above): they are recomputed fresh
+        # every event batch by computeDependentCouplings(), bypassing the Parameters
+        # singleton entirely. Only independent parameters/couplings (and never 'aS',
+        # whose declaration is likewise commented out) can be registered here.
+        # Two separate calls/maps, not one combined list into one map: model
+        # parameters (mdl_*) always stay plain cxsmpl<double> (registered
+        # into m_complex_model_params below), while independent couplings
+        # (GC_*) are cxtype -- FPTYPE-templated, so possibly cxsmpl<float>
+        # -- and go into m_complex_params instead. A single combined
+        # registration into one map only compiled by accident at FPTYPE=m/d
+        # (where cxtype happens to also be cxsmpl<double>) and failed to
+        # compile at FPTYPE=f -- see write_register_parameters()'s own doc
+        # comment.
+        replace_dict['register_parameters'] = \
+                               self.write_register_parameters(
+                                   [p for p in self.params_indep if p.name != 'aS'],
+                                   [n for n in self.params_ext_names if n != 'aS'],
+                                   complex_map_name='m_complex_model_params') + \
+                               '\n' + \
+                               self.write_register_parameters(self.coups_indep, [])
+        replace_dict['register_slha_keys'] = \
+                               self.write_register_slha_keys(
+                                   [e for e in self.params_ext_slha if e[0] != 'aS'])
         print_params_indep = [ line.replace('std::cout','//std::cout') + ' // now retrieved event-by-event (as G) from Fortran (running alphas #373)'
                                if '"aS =' in line else
                                line for line in self.write_print_parameters(self.params_indep).split('\n') ]
@@ -1985,6 +2032,8 @@ def get_process_function_definitions(self, write=True):
         # process-specific count (ProcessTables::nb_tmp_jamp) rather than
         # hardcoded per-process like the rest of calculate_jamps.
         self._nb_tmp_jamp = getattr(self.helas_call_writer, 'nb_tmp_jamp', 0)
+        replace_dict['refreshExternalMasses_lines'] = \
+                               self.get_refresh_external_masses_lines(self.matrix_elements[0])
         replace_dict['nparams'] = len(self.params2order)
         replace_dict['coupling_list'] = ' '
         replace_dict['hel_amps_cc'] = '#include \"HelAmps_%s.cc\"' % self.model_name # AV
@@ -2656,6 +2705,17 @@ def get_hardcoded_initProc_lines(self, matrix_element):
             initProc_lines.append('    m_masses.push_back( Parameters::%s );' % part.get('mass'))
         return '\n'.join(initProc_lines)
 
+    # Companion to get_initProc_lines: the same external-mass list, but pushed into an
+    # arbitrary "masses" vector instead of m_masses, for CPPProcess::refreshExternalMasses()
+    # (called after umami_set_parameter() overrides a mass so UMAMI_META_MASSES stays in sync;
+    # see madgraph/iolibs/template_files/madmatrix/umami.cc).
+    def get_refresh_external_masses_lines(self, matrix_element):
+        """Get refreshExternalMasses_lines for function definition for CPPProcess::refreshExternalMasses"""
+        lines = ['    // Set external particle masses for this matrix element']
+        for part in matrix_element.get_external_wavefunctions():
+            lines.append('    masses.push_back( m_pars->%s );' % part.get('mass'))
+        return '\n'.join(lines)
+
     # AV - replace the export_cpp.OneProcessExporterCPP method (fix helicity order and improve formatting)
     def get_helicity_matrix(self, matrix_element):
         """Return the Helicity matrix definition lines for this matrix element"""
@@ -3386,10 +3446,10 @@ def get_external_line(self, wf, argument):
             # part of the call key, but the leg number and the mass are wf's own:
             # argument is whichever wavefunction first had this key.
             if argument.get('offshell') and str(wf.get('mass')).upper() != 'ZERO':
-                call = call + 'offshell_mass<M_ACCESS>( m_pars->%%s, momenta, %d ), cHel[ihel][%%d],' \
+                call = call + 'offshell_mass<M_ACCESS>( m_pars->%%s, momenta, %d ), hel_sv[%%d],' \
                     % (wf.get('number_external') - 1)
             else:
-                call = call + 'm_pars->%s, cHel[ihel][%d],'
+                call = call + 'm_pars->%s, hel_sv[%d],'
         else:
             # AV This seems to be for scalars (spin==1???), pass neither mass nor helicity (#351)
             ###call = call + 'm_pars->%s,'
diff --git a/madmatrix/output.py b/madmatrix/output.py
index 27ef9a6c35..29908633ad 100644
--- a/madmatrix/output.py
+++ b/madmatrix/output.py
@@ -119,10 +119,13 @@ class ProcessExporterMadMatrix(export_cpp.ProcessExporterMG7):
                      # Default cards for the optional post-processing tools
                      # (Pythia8/Delphes/MadSpin/reweight/analysis) so that
                      # bin/generate_events can offer to enable and edit them.
+                     # The reweight card is the MG7 one: it documents both
+                     # reweighting backends (classic module, MadtRex).
                      'Cards': relative_path_list(pjoin(MG5DIR, 'Template', 'Common', 'Cards'),
-                                  ['madspin_card_default.dat', 'reweight_card_default.dat',
+                                  ['madspin_card_default.dat',
                                    'density_card_default.dat',
                                    'delphes_card_default.dat']) +
+                              relative_path_list(mg7_templates, ['reweight_card_default.dat']) +
                               relative_path_list(pjoin(MG5DIR, 'Template', 'LO', 'Cards'),
                                   ['pythia8_card_default.dat',
                                    'madanalysis5_parton_card_default.dat',
@@ -188,7 +191,18 @@ class ProcessExporterMadMatrix(export_cpp.ProcessExporterMG7):
     # AV (default from OM's tutorial) - add a debug printout
     def __init__(self, *args, **kwargs):
         self.in_madevent_mode = False # see MR #747
-        args[1]["me_lib_format"] = pjoin("lib", "libmadmatrix_{process_id}_{{device}}.so")
+        # {model} is resolved by ProcessExporterMG7.generate_subprocess_directory()
+        # alongside {process_id}, from the matrix element's own model -- same
+        # sanitized name (get_model_name()) used for MADMATRIX_LIB/
+        # MADMATRIX_COMMONLIB's %(model)s in madmatrix.mk. {{device}}/{{fptype}}
+        # stay literal ("{device}"/"{fptype}") in the resulting me_path: which
+        # concrete BACKEND/FPTYPE a given process directory was compiled for is
+        # a per-build choice made when the libraries are compiled and loaded
+        # (see compile_matrix_elements() in mg7/launch.py), not something
+        # subprocesses.json commits to at generation time.
+        args[1]["me_lib_format"] = pjoin(
+            "lib", "libmadmatrix_{model}_{process_id}_{{device}}_{{fptype}}.so"
+        )
         super().__init__(*args, **kwargs)
         # Honor the output command's --mask=True|False (flavor-mask
         # optimization for grouped/merged flavors). Default: enabled.
@@ -415,8 +429,9 @@ def finalize(self, *args, **kwargs):
 
     def write_processes_file(self):
         """SubProcesses/madmatrix_processes.json: per subprocess, its library
-        (with the make variables LIBTAG and BACKEND left as '{libtag}' and
-        '{backend}'), its legs, the processes it evaluates, its off-shell legs
+        (with the make variables LIBTAG, BACKEND and FPTYPE left as '{libtag}',
+        '{backend}' and '{fptype}', as for the common library, common_library),
+        its legs, the processes it evaluates, its off-shell legs
         and, per flavor index, every pdg combination (in the leg order of the
         process) that this index stands for, and what of it depends on the
         renormalisation scale beyond alpha_s (scale_dependence)."""
@@ -432,7 +447,8 @@ def write_processes_file(self):
                 flavors[flavor['index']].extend(list(pdgs) for pdgs in flavor['options'])
             subprocesses.append({
                 'path': path,
-                'library': 'libmadmatrix{libtag}_%s_{backend}.so' % path,
+                'library': 'libmadmatrix{libtag}_%s_%s_{backend}_{fptype}.so'
+                           % (self.lib_model_name, path),
                 'ninitial': len(info['incoming']),
                 'legs': list(info['incoming']) + list(info['outgoing']),
                 'processes': processes,
@@ -441,4 +457,6 @@ def write_processes_file(self):
                 'scale_dependence': scale,
                 'flavors': flavors})
         with open(pjoin(self.dir_path, 'SubProcesses', self.processes_file), 'w') as fsock:
-            json.dump({'subprocesses': subprocesses}, fsock, indent=1)
+            json.dump({'common_library': 'libmadmatrix_common{libtag}_%s_{backend}_{fptype}.so'
+                                         % self.lib_model_name,
+                       'subprocesses': subprocesses}, fsock, indent=1)
diff --git a/madmatrix/trex.py b/madmatrix/trex.py
deleted file mode 100644
index 30f549c918..0000000000
--- a/madmatrix/trex.py
+++ /dev/null
@@ -1,827 +0,0 @@
-# Copyright (C) 2023-2026 CERN.
-# Licensed under the GNU Lesser General Public License (version 3 or later).
-# Created originally by: Z. Wettersten (Sep 2024) for the MadGraph7 CUDACPP plugin.
-# Further modified by: Z. Wettersten (2024-2025).
-# Integrated with the MadGraph7 project in Feb 2026.
-
-import os
-import subprocess
-import re
-import sys
-import importlib.util
-
-# AV - PLUGIN_NAME can be one of PLUGIN/CUDACPP_OUTPUT or MG5aMC_PLUGIN/CUDACPP_OUTPUT
-PLUGIN_NAME = __name__.rsplit('.',1)[0]
-
-# AV - use templates for source code, scripts and Makefiles from PLUGINDIR instead of MG5DIR
-###from madgraph import MG5DIR
-PLUGINDIR = os.path.dirname( __file__ )
-
-__import__('%s.output'%PLUGIN_NAME)
-output = sys.modules['%s.output'%PLUGIN_NAME]
-__import__('%s.model_handling'%PLUGIN_NAME)
-model_handling = sys.modules['%s.model_handling'%PLUGIN_NAME]
-
-import importlib.util
-SPEC_EXPORTCPP = importlib.util.find_spec('madgraph.iolibs.export_cpp')
-PLUGIN_export_cpp = importlib.util.module_from_spec(SPEC_EXPORTCPP)
-SPEC_EXPORTCPP.loader.exec_module(PLUGIN_export_cpp)
-###sys.modules['PLUGIN.CUDACPP_OUTPUT.PLUGIN_export_cpp'] = PLUGIN_export_cpp # allow 'import PLUGIN.CUDACPP_OUTPUT.PLUGIN_export_cpp' in model_handling.py
-sys.modules['%s.PLUGIN_export_cpp'%PLUGIN_NAME] = PLUGIN_export_cpp # allow 'import <PLUGIN_NAME>.PLUGIN_export_cpp' in model_handling.py
-del SPEC_EXPORTCPP
-
-# AV - create a plugin-specific logger
-import logging
-logger = logging.getLogger('madgraph.%s.model_handling'%PLUGIN_NAME)
-
-#------------------------------------------------------------------------------------
-
-
-#------------------------------------------------------------------------------------
-
-from os.path import join as pjoin
-import madgraph
-import madgraph.iolibs.files as files
-import madgraph.iolibs.export_v4 as export_v4
-import madgraph.various.misc as misc
-import madgraph.interface.reweight_interface as rwgt_interface
-import madgraph.various.banner as banner
-import madgraph.models.check_param_card as check_param_card
-import madgraph.interface.extended_cmd as extended_cmd
-import madgraph.interface.common_run_interface as common_run_interface
-
-from . import launch_plugin
-
-class TREX_OneProcessExporter(model_handling.OneProcessExporterMadMatrix):
-    """A custom OneProcessExporter for the TREX reweighting"""
-    
-    rex_path = os.path.join( PLUGINDIR, 'MadtRex' )
-    
-    # ZW - rwgt functions
-    def get_rwgt_legs(self, process):
-        """Return string with particle ids and status in the REX std::pair format"""
-        return ",".join(["{%i,%i}" % (leg.get('state'), leg.get('id')) \
-            for leg in process.get('legs')]).replace('0', '-1')
-    
-    def get_rwgt_legs_status(self, process):
-        """Return string with particle statuses as a C++ vector<int>"""
-        return "{" + ",".join(["%i" % leg.get('state') for leg in process.get('legs')]).replace('0', '-1') + "}"
-
-    def get_rwgt_legs_pdg(self, process):
-        """Return string with particle PDG IDs as a C++ vector<int>"""
-        return "{" + ",".join(["%i" % leg.get('id') for leg in process.get('legs')]) + "}"
-
-    def get_rwgt_status_vec(self, processes):
-        """Return string with vectors of particle statuses"""
-        prtSets = []
-        for k in range(len(processes)):
-            prtSets.append( self.get_rwgt_legs_status(processes[k]))
-        return ",".join(prtSets)
-    
-    def get_rwgt_pdg_vec(self, processes):
-        """Return string with vectors of particle PDG IDs"""
-        prtSets = []
-        for k in range(len(processes)):
-            prtSets.append(self.get_rwgt_legs_pdg(processes[k]))
-        return ",".join(prtSets)
-
-    def get_rwgt_legs_vec(self, processes):
-        """Return string with vectors of particle ids and statuses"""
-        prtSets = []
-        for k in range(len(processes)):
-            prtSets.append("{" + self.get_rwgt_legs(processes[k]) + "}")
-        return ",".join(prtSets)
-        
-    def get_init_prts_vec(self, process):
-        """Return string with initial state particle ids for use in REX event sorting"""
-        prts = ",".join(["%i" % leg.get('id') for leg in process.get('legs') if leg.get('state') == 0])
-        return "{" + prts + "}"
-    
-    def get_init_prts_vecs(self, processes):
-        """Return string with vectors of initial state particle ids"""
-        prtSets = []
-        for k in range(len(processes)):
-            prtSets.append(self.get_init_prts_vec(processes[k]))
-        return ",".join(prtSets)
-    
-    def get_fin_prts_vec(self, process):
-        """Return string with final state particle ids for use in REX event sorting"""
-        prts = ",".join(["%i" % leg.get('id') for leg in process.get('legs') if leg.get('state') == 1])
-        return "{" + prts + "}"
-    
-    def get_fin_prts_vecs(self, processes):
-        """Return string with vectors of final state particle ids"""
-        prtSets = []
-        for k in range(len(processes)):
-            prtSets.append(self.get_fin_prts_vec(processes[k]))
-        return ",".join(prtSets)
-        
-    def get_rwgt_procMap(self, process):
-        """Return string with particle states and order in the REX procMap format"""
-        currState = False
-        retString = "thisProc{{-1,{"
-        for leg in process.get('legs'):
-            if currState == leg.get('state'):
-                retString += "%i," % leg.get('id')
-            else:
-                currState = leg.get('state')
-                retString += "}},{1,{%i," % leg.get('id')
-        retString = retString[:-1] + "}}}"
-        return retString
-    
-    def get_proc_dir(self):
-        """Return process directory name for the current process"""
-        return "P%d_%s" % (self.process_number, self.process_name)
-        
-    def get_rwgt_runner(self):
-        """Return string to initialise the rwgtRunners in tRex"""
-        return "%s::runner" % (self.get_proc_dir())
-    
-    def get_rwgt_includes(self):
-        """Return string with the include directives for the tRex reweighting"""
-        return "#include \"P%d_%s/rwgt_runner.cc\"" % (self.process_number, self.process_name)
-    
-    def write_rwgt_header(self):
-        """Writes a simple rwgt_runner.h file to forward declare the runner object"""
-        # Adjust the placeholders for use with `.format()`
-        rwgt_h = """#ifndef {namespace}_RWGT_RUNNER_H
-    #define {namespace}_RWGT_RUNNER_H
-    #include \"rwgt_instance.h\"
-    namespace {namespace} {{
-        extern rwgt::instance runner;
-    }}
-    #endif""".format(namespace=self.get_proc_dir())
-        
-        # Using `with` statement for better file handling
-        with open(os.path.join(self.path, 'rwgt_runner.h'), 'w') as ff:
-            ff.write(rwgt_h)
-    
-    def edit_rwgt_header(self):
-        """Adds process-specific details to the rwgt_runner.h template"""
-        replace_dict = super().get_process_class_definitions(write=False)
-        replace_dict['process_namespace'] = self.get_proc_dir()
-        replace_dict['info_lines'] = model_handling.PLUGIN_export_cpp.get_mg5_info_lines()
-        template = open(pjoin(self.rex_path,'template_files', 'rwgt_runner_h.inc'),'r').read()
-        ff = open(pjoin(self.path, 'rwgt_runner.h'),'w')
-        ff.write(template % replace_dict)
-        ff.close()
-    
-    def edit_rwgt_runner(self):
-        """Create the rwgt_runner.cc file for the tRex reweighting"""
-        ###misc.sprint('Entering OneProcessExporterMadMatrixRwgt.edit_rwgt_runner')
-        # Create the rwgt_runner.cc file
-        replace_dict = super().get_process_class_definitions(write=False)
-        replace_dict['process_namespace'] = self.get_proc_dir()
-        replace_dict['info_lines'] = model_handling.PLUGIN_export_cpp.get_mg5_info_lines()
-        replace_dict['parton_ids'] = self.get_rwgt_pdg_vec(self.matrix_elements[0].get('processes'))
-        replace_dict['parton_status'] = self.get_rwgt_status_vec(self.matrix_elements[0].get('processes'))
-        replace_dict['no_events'] = len(self.matrix_elements[0].get('processes'))
-        template = open(pjoin(self.rex_path,'template_files', 'rwgt_runner_cc.inc'),'r').read()
-        ff = open(pjoin(self.path, 'rwgt_runner.cc'),'w')
-        ff.write(template % replace_dict)
-        ff.close()
-    
-    # ZW - override the PLUGIN method to generate the rwgt_runner.cc file as well
-    # note: also generating standard check_sa.cc and gcheck_sa.cu files, which
-    # are not used in the REX reweighting
-    def generate_process_files(self):
-        """Generate mgOnGpuConfig.h, CPPProcess.cc, CPPProcess.h, check_sa.cc, gXXX.cu links"""
-        super().generate_process_files()
-        # for TREX for now no /backend
-        self.edit_mgonGPU()
-        self.edit_rwgt_header()
-        self.edit_rwgt_runner()
-        
-class TREX_ProcessExporter(output.ProcessExporterMadMatrix):
-    
-    oneprocessclass = TREX_OneProcessExporter
-    
-    rwgt_names = []
-    proc_lines = []
-    
-    s = PLUGINDIR + '/madgraph/iolibs/template_files/'
-    t = PLUGINDIR + '/MadtRex/'
-    r = PLUGINDIR + '/MadtRex/template_files/'
-    m = PLUGINDIR + '/MadtRex/makefiles/'
-    from_template = dict(output.ProcessExporterMadMatrix.from_template)
-    # for TREX for now no /backend
-    from_template['src'] = from_template['src'] + output.relative_path_list(
-        output.ProcessExporterMadMatrix.madmatrix_templates,
-        ['mgOnGpuFptypes.h', 'mgOnGpuCxtypes.h', 'mgOnGpuVectors.h', 'constexpr_math.h']
-    ) + [t+'librex.so', t+'libtearex.so',
-        t+'Rex.h', t+'teaRex.h',
-        r+'rwgt_instance.h', r+'rwgt_instance.cc']
-    from_template['SubProcesses'] = from_template['SubProcesses'] + [m+'cudacpp_driver.mk',
-                                                                     r+'rwgt_instance.h', t+'Rex.h', t+'teaRex.h']
-
-    to_link_in_P = output.ProcessExporterMadMatrix.to_link_in_P + ['rwgt_instance.h', 'Rex.h', 'teaRex.h']
-
-    template_src_make = pjoin(m,'cudacpp_rex_src.mk')
-    # template_tst_make = pjoin(m,'cudacpp_test.mk')
-    template_Sub_make = pjoin(m,'cudacpp_runner.mk')
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.template_path = PLUGINDIR
-        self.rwgt_names = []
-        self.proc_lines = []
-        self.compile_library()
-    
-    def check_library(self, temp_dir=None):
-        """Checks whether librex.so and libtearex.so exist in the rex template directory"""
-        if temp_dir is None:
-            temp_dir = pjoin(PLUGINDIR, 'MadtRex')
-        if not os.path.exists(pjoin(temp_dir, 'librex.so')):
-            return False
-        if not os.path.exists(pjoin(temp_dir, 'libtearex.so')):
-            return False
-        return True
-
-    def compile_library(self):
-        """Compile librex.so and libtearex.so in the rex template directory"""
-        rex_template_dir = pjoin(PLUGINDIR, 'MadtRex')
-        if not self.check_library(rex_template_dir):
-            logger.info('Rex or teaRex libraries not found, compiling them now...')
-            logger.info('This may take a while, but only needs to be done once.')
-            misc.compile(arg=['-f', 'rex.mk'],cwd=rex_template_dir,mode='cpp')
-        return
-    
-    def generate_subprocess_directory(self, matrix_element, cpp_helas_call_writer,
-                                      proc_number=None):
-        """Generate the Pxxxxx directory for a subprocess in C++ standalone,
-        including the necessary .h and .cc files"""
-
-        
-        process_exporter_cpp = self.oneprocessclass(matrix_element,cpp_helas_call_writer)
-        
-        self.rwgt_names.append("P%d_%s" % (process_exporter_cpp.process_number, 
-                                             process_exporter_cpp.process_name))
-        
-        process_lines = "\n".join([process_exporter_cpp.get_process_info_lines(me) for me in \
-                                   process_exporter_cpp.matrix_elements])
-        self.proc_lines.append(process_lines)
-        
-        # Create the directory PN_xx_xxxxx in the specified path
-        dirpath = pjoin(self.dir_path, 'SubProcesses', "P%d_%s" % (process_exporter_cpp.process_number, 
-                                             process_exporter_cpp.process_name))
-        try:
-            os.mkdir(dirpath)
-        except os.error as error:
-            logger.warning(error.strerror + " " + dirpath)
-    
-        with misc.chdir(dirpath):
-            logger.info('Creating files in directory %s' % dirpath)
-            process_exporter_cpp.path = dirpath
-            # Create the process .h and .cc files
-            process_exporter_cpp.generate_process_files()
-            for file in self.to_link_in_P:
-                files.ln('../%s' % file)
-        return
-    
-    def export_driver(self):
-        replace_dict = {}
-        replace_dict['info_lines'] = model_handling.PLUGIN_export_cpp.get_mg5_info_lines()
-        replace_dict['multiprocess_lines'] = "\n".join(self.proc_lines)
-        replace_dict['include_lines'] = ''
-        replace_dict['make_rwgt'] = ''
-        for name in self.rwgt_names:
-            replace_dict['include_lines'] += '#include "%s/rwgt_runner.h"\n' % name
-            replace_dict['make_rwgt'] += '%s::make_reweightor(batch_size),' % name
-        replace_dict['make_rwgt'] = replace_dict['make_rwgt'][:-1]
-        template_path = os.path.join( PLUGINDIR, 'madgraph', 'iolibs', 'template_files' )
-        template = open(pjoin(self.r, 'rwgt_driver.inc'),'r').read()
-        ff = open(pjoin(self.dir_path, 'SubProcesses', 'rwgt_driver.cc'),'w')
-        ff.write(template % replace_dict)
-        ff.close()
-    
-    def link_makefile(self):
-        """Link the makefile for the tRex reweighting"""
-        files.ln(pjoin(self.dir_path, 'SubProcesses', 'cudacpp_driver.mk'), starting_dir=pjoin(self.dir_path, 'SubProcesses'), name='makefile')
-    
-    def finalize(self, matrix_element, cmdhistory, MG5options, outputflag):
-        self.export_driver()
-        self.link_makefile()
-        return super().finalize(matrix_element, cmdhistory, MG5options, outputflag)
-    
-class TREX_ReweightInterface(rwgt_interface.ReweightInterface):
-    """A custom ReweightInterface for the tRex reweighting"""
-    
-    prompt = 'MadtRex>'
-    sa_class = 'standalone_trex'
-    debug_output = 'tRex_debug'
-    
-    def __init__(self, *args, **kwargs):
-        """Initialise the tRex reweighting interface
-        Currently no (substantial) changes compared to upstream are necessary,
-        but adding an __init__ method allows for future modifications"""
-        super().__init__(*args, **kwargs)
-        self.param_card = None
-        self.reweight_card = []
-        self.reweight_names = []
-        self.backend = 'auto' # default backend for tRex reweighting
-        self.fptype = 'm' # default floating point type for tRex reweighting
-        self.nb_threads = 1
-        if self.mother:
-            self.nb_threads = self.mother.options['nb_core'] if self.mother.options['run_mode'] !=0 else 1
-        else:
-            self.nb_threads = 1
-        self.batch_size = 32
-        self.compile_library()
-
-    def check_library(self, temp_dir=None):
-        """Checks whether librex.so and libtearex.so exist in the rex template directory"""
-        if temp_dir is None:
-            temp_dir = pjoin(PLUGINDIR, 'MadtRex')
-        if not os.path.exists(pjoin(temp_dir, 'librex.so')):
-            return False
-        if not os.path.exists(pjoin(temp_dir, 'libtearex.so')):
-            return False
-        return True
-
-    def compile_library(self):
-        """Compile librex.so and libtearex.so in the rex template directory"""
-        if self.multicore=='wait':
-            return
-        rex_template_dir = pjoin(PLUGINDIR, 'MadtRex')
-        if not self.check_library(rex_template_dir):
-            misc.compile(arg=['-f', 'rex.mk'],cwd=rex_template_dir,mode='cpp')
-        return
-
-    def setup_f2py_interface(self):
-        """"Override native setup_f2py_interface to avoid parsing things not necessary for tRex reweighting"""
-        self.create_standalone_directory()
-        self.compile()
-        
-    def launch_actual_reweighting(self, *args, **kwargs):
-        """override standard launch command to instead call the tRex reweighting"""
-        
-        import csv
-        
-        if self.rwgt_dir:
-            path_me =self.rwgt_dir
-        else:
-            path_me = self.me_dir 
-        
-        if self.second_model or self.second_process or self.dedicated_path:
-            rw_dir = pjoin(path_me, 'rw_me_%s' % self.nb_library)
-        else:
-            rw_dir = pjoin(path_me, 'rw_me')
-
-        run_path = pjoin(rw_dir, 'SubProcesses')
-        input_file = os.path.relpath(self.lhe_input.path, run_path)
-        output_file = input_file + 'rw'
-        output_path = self.lhe_input.path + 'rw'
-        param_card = pjoin(rw_dir, 'Cards', 'param_card.dat')
-
-        #ZW: Exceptions, making sure all the necessary files for teaREX are accessible
-        if( misc.is_executable(pjoin(run_path,'rwgt_driver_gpu.exe')) ):
-            driver = pjoin(run_path, 'rwgt_driver_gpu.exe')
-        elif(misc.is_executable(pjoin(run_path,'rwgt_driver_cpp.exe')) ):
-            driver = pjoin(run_path,'rwgt_driver_cpp.exe')
-        else:
-            raise Exception('No MadtRex driver found for parallel reweighting')
-        if not os.path.exists(param_card):
-            try:
-                files.cp(os.path.join(path_me, 'Cards', 'param_card_default.dat'), param_card)
-            except:
-                raise Exception("No param_card.dat file found in %s" % pjoin(path_me, 'Cards'))
-        param_path = os.path.relpath(param_card, run_path)    
-
-        rwgt_card = os.path.join(path_me, 'Cards', 'reweight_card.dat')
-        
-        self.write_reweight_card(rwgt_card)
-
-        if not os.path.exists(rwgt_card):
-            try:
-                files.cp(os.path.join(path_me, 'Cards', 'reweight_card_default.dat'), rwgt_card)
-            except:
-                raise Exception("No reweight_card.dat file found in %s" % pjoin(path_me, 'Cards'))
-        rwgt_path = os.path.relpath(rwgt_card, run_path)
-        target = ''
-        if not self.mother:
-            name, ext = self.lhe_input.name.rsplit('.',1)
-            target = '%s_out.%s' % (name, ext)            
-        elif self.output_type != "default" :
-            target = pjoin(self.mother.me_dir, 'Events', self.mother.run_name, 'events.lhe')
-        else:
-            target = self.lhe_input.path
-        n_threads = self.nb_threads
-        if n_threads < 1:
-            n_threads = 1
-        nb_warps = self.batch_size
-        if nb_warps < 1:
-            nb_warps = 32
-
-        #ZW: rwgt_driver is written and compiled properly, now just to figure out how to run it through MG
-        subprocess.call([driver, '-lhe=%s' % input_file, '-slha=%s' % param_card, '-rwgt=%s' % rwgt_card, '-out=%s' % output_file, '-nt=%s' % n_threads, '-warp=%s' % nb_warps], cwd=run_path)
-
-        # ZW: check if output exists, if not nicely raise an exception
-        if not os.path.exists(output_path):
-            if os.path.exists(target):
-                files.mv(self.lhe_input.path, target)
-                logger.info('Error in reweighting: output file not found. Returning original LHE file.')
-                return
-            else:
-                raise Exception('Error in reweighting: output file not found. Input file not found. Exiting.')
-        else:
-            files.mv(output_path, target)
-            csv_file = pjoin(run_path, 'rwgt_results.csv')
-            with open(csv_file, newline='') as results:
-                iters = csv.reader(results)
-                for row in iters:
-                    self.all_cross_section[(row[0],'')] = (float(row[1]), float(row[2]))
-
-            return
-    
-    def compile(self):
-        """override compile to use the TREX makefiles"""
-        
-        if self.multicore=='wait':
-            return
-        
-        if not self.rwgt_dir:
-            path_me = self.me_dir
-        else:
-            path_me = self.rwgt_dir
-        
-        rwgt_dir_possibility =   ['rw_me','rw_me_%s' % self.nb_library,'rw_mevirt','rw_mevirt_%s' % self.nb_library]
-        for onedir in rwgt_dir_possibility:
-            if not os.path.isdir(pjoin(path_me,onedir)):
-                continue
-            pdir = pjoin(path_me, onedir, 'SubProcesses')
-            if self.mother:
-                nb_core = self.mother.options['nb_core'] if self.mother.options['run_mode'] !=0 else 1
-            else:
-                nb_core = 1
-            misc.compile(arg=['BACKEND=%s' % self.backend,'FPTYPE=%s' % self.fptype],cwd=pdir, nb_core=nb_core,mode='cpp')
-        return
-    
-    def load_module(self):
-        """override load_module since we do not use it"""
-        return
-    
-    # def import_command_file(self, filepath):
-    #     """override import_command_file to simply launch TREX"""
-    #     self.exec_cmd('launch', precmd=True)
-    #     return
-    
-    def do_change(self,line):
-        """extend do_change to include the backend and floating point type options"""
-        args = self.split_arg(line)
-        if len(args) < 2:
-            return super().do_change(line)
-        if args[0].lower() in ['cudacpp_backend', 'backend']:
-            cudacpp_supported_backends = [ 'cuda', 'hip', 'scalar', 'simd_128', 'simd_256', 'avx512y', 'simd_512', 'auto' ]
-            if args[1].lower() in cudacpp_supported_backends:
-                self.backend = args[1].lower()
-                logger.info("Setting tRex reweighting backend to '%s'" % self.backend)
-            else:
-                raise Exception( "Invalid cudacpp_backend='%s': supported backends are %s"%(args[1], cudacpp_supported_backends) )
-        elif args[0].lower() in ['fptype', 'floating_type']:
-            fptype_supported = ['f', 'm', 'd']
-            fptype_semisupported = ['float', 'mixed', 'double']
-            if args[1].lower() in fptype_semisupported:
-                args[1] = args[1].lower()[0]
-            if args[1].lower() in fptype_supported:
-                self.fptype = args[1].lower()
-                logger.info("Setting tRex reweighting floating point type to '%s'" % self.fptype)
-            else:
-                raise Exception( "Invalid fptype='%s': supported types are %s"%(args[1], fptype_supported) )
-        elif args[0].lower() in ['nb_core', 'nb_threads']:
-            try:
-                self.nb_threads = int(args[1])
-                if self.nb_threads < 1:
-                    raise ValueError
-                logger.info("Setting tRex reweighting number of threads to '%s'" % self.nb_threads)
-            except ValueError:
-                raise Exception( "Invalid nb_core='%s': must be a positive integer"%(args[1]) )
-        elif args[0].lower() in ['batch_size', 'nb_batch', 'vector_size', 'vec_size', 'nb_warp', 'warp_size']:
-            try:
-                self.batch_size = int(args[1])
-                if self.batch_size < 1:
-                    raise ValueError
-                logger.info("Setting tRex reweighting batch size to '%s'" % self.batch_size)
-            except ValueError:
-                raise Exception( "Invalid batch_size='%s': must be a positive integer"%(args[1]) )
-        else:
-            return super().do_change(line)
-        return
-        
-    def do_launch(self, line):
-        """override do_launch to instead overwrite the reweight_card
-        to fit the expected input for TREX without having to extend TREX itself"""
-        args = self.split_arg(line)
-        opts = self.check_launch(args)
-        mgcmd = self.mg5cmd
-        if opts['rwgt_name']:
-            self.options['rwgt_name'] = opts['rwgt_name']
-        if opts['rwgt_info']:
-            self.options['rwgt_info'] = opts['rwgt_info']
-        model_line = self.banner.get('proc_card', 'full_model_line')
-
-        # TV: Load model: needed for the combine_ij function: maybe not needed everyt time??? 
-        model = self.banner.get('proc_card', 'model')
-        self.load_model( model, True, False)
-
-        if not self.has_standalone_dir:
-            out = self.setup_f2py_interface()
-            if out:
-                return
-
-        if not self.param_card:
-            s_orig = self.banner['slha']
-            self.param_card = check_param_card.ParamCard(s_orig.splitlines())
-
-        # get the mode of reweighting #LO/NLO/NLO_tree/...
-        type_rwgt = self.get_weight_names()
-        
-        if self.rwgt_dir:
-            path_me =self.rwgt_dir
-        else:
-            path_me = self.me_dir 
-        
-            
-        # get iterator over param_card and the name associated to the current reweighting.
-        param_card_iterator, tag_name = self.handle_param_card(model_line, args, type_rwgt)
-
-        self.reweight_names.append(tag_name)
-        
-        # perform the scanning
-        if param_card_iterator:
-            if self.options['rwgt_name']:
-                reweight_name = self.options['rwgt_name'].rsplit('_',1)[0] # to avoid side effect during the scan
-            else:
-                reweight_name = None
-            for i,card in enumerate(param_card_iterator):
-                if reweight_name:
-                    self.options['rwgt_name'] = '%s_%s' % (reweight_name, i+1)
-                self.new_param_card = card
-                #card.write(pjoin(rw_dir, 'Cards', 'param_card.dat'))
-                self.exec_cmd("launch --keep_card", printcmd=False, precmd=True)
-    
-    def check_multicore(self):
-        """override check_multicore to overloading the CPU (we never want to run TREX in multicore mode)"""
-        return False
-    
-    def handle_param_card(self, model_line, args, type_rwgt):
-        """override handle_param_card to get rid of all the unnecessary checks and file writing
-        now simply loads the param_card and uses get_diff to translate into internal format"""
-
-        if self.rwgt_dir:
-            path_me =self.rwgt_dir
-        else:
-            path_me = self.me_dir 
-            
-        if self.second_model or self.second_process or self.dedicated_path:
-            rw_dir = pjoin(path_me, 'rw_me_%s' % self.nb_library)
-        else:
-            rw_dir = pjoin(path_me, 'rw_me')
-        if not '--keep_card' in args:
-            if self.has_nlo and self.rwgt_mode != "LO":
-                rwdir_virt = rw_dir.replace('rw_me', 'rw_mevirt')
-            with open(pjoin(rw_dir, 'Cards', 'param_card.dat'), 'w') as fsock:
-                fsock.write(self.banner['slha']) 
-            out, cmd = common_run_interface.CommonRunCmd.ask_edit_card_static(cards=['param_card.dat'],
-                                ask=self.ask, pwd=rw_dir, first_cmd=self.stored_line,
-                                write_file=False, return_instance=True
-                                )
-            self.stored_line = None
-            card = cmd.param_card
-            new_card = card.write()
-        elif self.new_param_card:
-            new_card = self.new_param_card.write()
-        else:
-            new_card = open(pjoin(rw_dir, 'Cards', 'param_card.dat')).read()
-        
-        # check for potential scan in the new card 
-        pattern_scan = re.compile(r'''^(decay)?[\s\d]*scan''', re.I+re.M) 
-        param_card_iterator = []
-        if pattern_scan.search(new_card):
-            import madgraph.interface.extended_cmd as extended_cmd
-            try:
-                import internal.extended_cmd as extended_internal
-                Shell_internal = extended_internal.CmdShell
-            except:
-                Shell_internal = None
-            if not isinstance(self.mother, (extended_cmd.CmdShell, Shell_internal)): 
-                raise Exception("scan are not allowed on the Web")
-            # at least one scan parameter found. create an iterator to go through the cards
-            main_card = check_param_card.ParamCardIterator(new_card)
-            if self.options['rwgt_name']:
-                self.options['rwgt_name'] = '%s_0' % self.options['rwgt_name']
-
-            param_card_iterator = main_card
-            first_card = param_card_iterator.next(autostart=True)
-            new_card = first_card.write()
-            self.new_param_card = first_card
-            #first_card.write(pjoin(rw_dir, 'Cards', 'param_card.dat'))  
-
-        # check if "Auto" is present for a width parameter)
-        if 'block' not in new_card.lower():
-            raise Exception(str(new_card))
-        tmp_card = new_card.lower().split('block',1)[1]
-        if "auto" in tmp_card: 
-            if param_card_iterator:
-                first_card.write(pjoin(rw_dir, 'Cards', 'param_card.dat'))
-            else:
-                ff = open(pjoin(rw_dir, 'Cards', 'param_card.dat'),'w')
-                ff.write(new_card)
-                ff.close()
-                
-            self.mother.check_param_card(pjoin(rw_dir, 'Cards', 'param_card.dat'))
-            new_card = open(pjoin(rw_dir, 'Cards', 'param_card.dat')).read()
-
-
-        # Find new tag in the banner and add information if needed
-        if 'initrwgt' in self.banner and self.output_type == 'default': 
-            if 'name=\'mg_reweighting\'' in self.banner['initrwgt']:
-                blockpat = re.compile(r'''<weightgroup name=\'mg_reweighting\'\s*weight_name_strategy=\'includeIdInWeightName\'>(?P<text>.*?)</weightgroup>''', re.I+re.M+re.S)
-                before, content, after = blockpat.split(self.banner['initrwgt'])
-                header_rwgt_other = before + after
-                pattern = re.compile('<weight id=\'(?:rwgt_(?P<id>\\d+)|(?P<id2>[_\\w\\-\\.]+))(?P<rwgttype>\\s*|_\\w+)\'>(?P<info>.*?)</weight>', re.S+re.I+re.M)
-                mg_rwgt_info = pattern.findall(content)
-                maxid = 0
-                for k,(i, fulltag, nlotype, diff) in enumerate(mg_rwgt_info):
-                    if i:
-                        if int(i) > maxid:
-                            maxid = int(i)
-                        mg_rwgt_info[k] = (i, nlotype, diff) # remove the pointless fulltag tag
-                    else:
-                        mg_rwgt_info[k] = (fulltag, nlotype, diff) # remove the pointless id tag
-                        
-                maxid += 1
-                rewgtid = maxid
-                if self.options['rwgt_name']:
-                    #ensure that the entry is not already define if so overwrites it
-                    for (i, nlotype, diff) in mg_rwgt_info[:]:
-                        for flag in type_rwgt:
-                            if 'rwgt_%s' % i == '%s%s' %(self.options['rwgt_name'],flag) or \
-                                i == '%s%s' % (self.options['rwgt_name'], flag):
-                                    logger.warning("tag %s%s already defines, will replace it", self.options['rwgt_name'],flag)
-                                    mg_rwgt_info.remove((i, nlotype, diff))
-                                                
-            else:
-                header_rwgt_other = self.banner['initrwgt'] 
-                mg_rwgt_info = []
-                rewgtid = 1
-        else:
-            self.banner['initrwgt']  = ''
-            header_rwgt_other = ''
-            mg_rwgt_info = []
-            rewgtid = 1
-
-        # add the reweighting in the banner information:
-        #starts by computing the difference in the cards.
-        #s_orig = self.banner['slha']
-        #self.orig_param_card_text = s_orig
-        s_new = new_card
-        self.new_param_card = check_param_card.ParamCard(s_new.splitlines())
-        
-        #define tag for the run
-        if self.options['rwgt_name']:
-            tag = self.options['rwgt_name']
-        else:
-            tag = str(rewgtid)
-
-        if 'rwgt_info' in self.options and self.options['rwgt_info']:
-            card_diff = self.options['rwgt_info']
-            for name in type_rwgt:
-                mg_rwgt_info.append((tag, name, self.options['rwgt_info']))
-        elif not self.second_model and not self.dedicated_path:
-            old_param = self.param_card
-            new_param =  self.new_param_card
-            card_diff = old_param.create_diff(new_param)
-            if card_diff == '' and not self.second_process:
-                    logger.warning(' REWEIGHTING: original card and new card are identical.')
-            try:
-                if old_param['sminputs'].get(3)- new_param['sminputs'].get(3) > 1e-3 * new_param['sminputs'].get(3):
-                    logger.warning("We found different value of alpha_s. Note that the value of alpha_s used is the one associate with the event and not the one from the cards.")
-            except Exception as error:
-                logger.debug("error in check of alphas: %s" % str(error))
-                pass #this is a security                
-            if not self.second_process:
-                for name in type_rwgt:
-                    mg_rwgt_info.append((tag, name, card_diff))
-            else:
-                str_proc = "\n change process  ".join([""]+self.second_process)
-                for name in type_rwgt:
-                    mg_rwgt_info.append((tag, name, str_proc + '\n'+ card_diff))
-        else:
-            if self.second_model:
-                str_info = "change model %s" % self.second_model
-            else:
-                str_info =''
-            if self.second_process:
-                str_info += "\n change process  ".join([""]+self.second_process)
-            if self.dedicated_path:
-                for k,v in self.dedicated_path.items():
-                    str_info += "\n change %s %s" % (k,v)
-            card_diff = str_info
-            str_info += '\n' + s_new
-            for name in type_rwgt:
-                mg_rwgt_info.append((tag, name, str_info))
-
-        # re-create the banner.
-        self.banner['initrwgt'] = header_rwgt_other
-        if self.output_type == 'default':
-            self.banner['initrwgt'] += '\n<weightgroup name=\'mg_reweighting\' weight_name_strategy=\'includeIdInWeightName\'>\n'
-        else:
-            self.banner['initrwgt'] += '\n<weightgroup name=\'main\'>\n'
-        for tag, rwgttype, diff in mg_rwgt_info:
-            if self.inc_sudakov:
-                try:
-                    sud_order = int(rwgttype[-1]) -1
-                    sud_order = '10' +rwgttype[-2:]
-                    self.banner['initrwgt'] += '<weight id=\'%s\'>%sscale_%s_sud</weight>\n' % \
-                            (rwgttype, diff, sud_order)
-                except IndexError:
-                    logger.critical('This is a reweighted event file! Do not reweight with ewsudakov twice')
-                    sys.exit(1)
-            else:
-                if tag.isdigit():
-                    self.banner['initrwgt'] += '<weight id=\'rwgt_%s%s\'>%s</weight>\n' % \
-                                    (tag, rwgttype, diff)
-                else:
-                    self.banner['initrwgt'] += '<weight id=\'%s%s\'>%s</weight>\n' % \
-                                    (tag, rwgttype, diff)
-        self.banner['initrwgt'] += '\n</weightgroup>\n'
-        self.banner['initrwgt'] = self.banner['initrwgt'].replace('\n\n', '\n')
-
-        #logger.info('starts to compute weight for events with the following modification to the param_card:')
-        #logger.info(card_diff.replace('\n','\nKEEP:'))
-        try:
-            self.run_card = banner.Banner(self.banner).charge_card('run_card')
-        except Exception:
-            logger.debug('no run card found -- reweight interface')
-            self.run_card = None
-
-        if self.options['rwgt_name']:
-            tag_name = self.options['rwgt_name']
-        else:
-            tag_name = 'rwgt_%s' % rewgtid
-
-        self.reweight_card.append(card_diff)
-        
-        return param_card_iterator, tag_name
-    
-    def write_reweight_card(self,rwgt_path):
-        """function for collecting all the reweight iterations from the parsed reweight card
-        and write it out with the explicit 'set BLOCK PARAM VALUE' format"""
-        if( len(self.reweight_names) != len(self.reweight_card) ):
-            raise Exception('Mismatch in number of reweight names and reweight cards')
-        
-        output_card = ''
-        
-        for i, card in enumerate(self.reweight_card):
-            output_card += 'launch --rwgt_name=%s\n' % self.reweight_names[i]
-            output_card += card + '\n'
-        
-        output_card = output_card.replace('param_card', '').replace('  ', ' ')
-        
-        with open(rwgt_path, 'w') as f:
-            f.write(output_card)
-        
-        return
-    
-    def do_quit(self, line):
-        if self.exitted:
-            return
-        try:
-            self.launch_actual_reweighting()
-        except:
-            raise Exception("Error in tRex reweighting. Exiting.")
-        
-        self.exitted = True
-        
-        if 'init' in self.banner:
-            cross = 0 
-            error = 0
-            for line in self.banner['init'].split('\n'):
-                split = line.split()
-                if len(split) == 4:
-                    cross += float(split[0])
-                    error += float(split[1])**2
-            error = error**0.5
-        if not self.multicore == 'create':
-            # No print of results for the multicore mode for the one printed on screen
-            if 'orig' not in self.all_cross_section:
-                logger.info('Original cross-section: %s +- %s pb' % (cross, error))
-            else: 
-                logger.info('Original cross-section: %s +- %s pb (cross-section from sum of weights: %s)' % (cross, error, self.all_cross_section['orig'][0]))
-            logger.info('Computed cross-section:')
-            keys = list(self.all_cross_section.keys())
-            keys.sort(key=lambda x: str(x))
-            for key in keys:
-                if key == 'orig':
-                    continue
-                logger.info('%s : %s +- %s pb' % (key[0] if not key[1] else '%s%s' % key,
-                    self.all_cross_section[key][0],self.all_cross_section[key][1] ))  
-        self.terminate_fortran_executables()
-
-        if self.rwgt_dir and self.multicore == False:
-            self.save_to_pickle()
-        
-        with misc.stdchannel_redirected(sys.stdout, os.devnull):
-            for run_id in self.calculator:
-                del self.calculator[run_id]
-            del self.calculator
diff --git a/madspace/CMakeLists.txt b/madspace/CMakeLists.txt
index c7989a423f..397ed42de4 100644
--- a/madspace/CMakeLists.txt
+++ b/madspace/CMakeLists.txt
@@ -337,7 +337,11 @@ install(TARGETS madspace DESTINATION madspace/lib)
 install(
     DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/include/"
     DESTINATION "madspace/include"
-    FILES_MATCHING PATTERN "*.hpp*"
+    # *.h (not just *.hpp*) so umami.h -- the public UMAMI ABI header needed by
+    # anyone implementing or consuming a MatrixElementApi-compatible shared
+    # library -- is installed too, and *.c for mixmax/mixmax_skip_N17.c, which
+    # mixmax.hpp #includes.
+    FILES_MATCHING PATTERN "*.hpp*" PATTERN "*.h" PATTERN "*.c"
 )
 # Generated public headers: installed alongside function.hpp/instruction.hpp
 # so a downstream #include "function_builder_mixin.inc" resolves the same
diff --git a/madspace/include/madspace/driver/channel_generator.hpp b/madspace/include/madspace/driver/channel_generator.hpp
index b8476a1e32..80b03861e6 100644
--- a/madspace/include/madspace/driver/channel_generator.hpp
+++ b/madspace/include/madspace/driver/channel_generator.hpp
@@ -226,6 +226,7 @@ class ChannelEventGenerator {
         std::array<int, 5> ladder;
         int subprocess_index;
         int bias;
+        int matrix_element;
         int random, rest;
     };
 
diff --git a/madspace/include/madspace/driver/context.hpp b/madspace/include/madspace/driver/context.hpp
index 0d1ed0eeb5..43eb5acab2 100644
--- a/madspace/include/madspace/driver/context.hpp
+++ b/madspace/include/madspace/driver/context.hpp
@@ -3,9 +3,11 @@
 #include <cstdint>
 #include <functional>
 #include <optional>
+#include <string>
 #include <type_traits>
 #include <typeindex>
 #include <unordered_map>
+#include <utility>
 
 #include "madspace/compgraphs.hpp"
 #include "madspace/driver/tensor.hpp"
@@ -34,6 +36,7 @@ struct UmamiFunctions {
         supported_outputs;
     std::function<std::remove_pointer_t<decltype(&umami_initialize)>> initialize;
     std::function<std::remove_pointer_t<decltype(&umami_set_parameter)>> set_parameter;
+    std::function<std::remove_pointer_t<decltype(&umami_get_parameter)>> get_parameter;
     std::function<std::remove_pointer_t<decltype(&umami_matrix_element)>>
         matrix_element;
     std::function<std::remove_pointer_t<decltype(&umami_free)>> free;
@@ -174,6 +177,26 @@ class MatrixElementApi {
     /// Opaque per-thread UMAMI process handle, initialized on first use.
     void* process_instance() const { return _instances.get().get(); }
 
+    // The Device this api's library was loaded on (ie the owning Context's
+    // device). Callers driving umami_matrix_element() need this to know
+    // whether the inputs/outputs it exchanges have to be staged through
+    // device-resident buffers (cuda/hip) or can be passed as plain host
+    // pointers (cpu).
+    DevicePtr device() const noexcept { return _device; }
+
+    // Whether the loaded library actually implements umami_set_parameter/
+    // umami_get_parameter (both optional symbols; a library built without
+    // them still loads fine, it just can't be parameter-reweighted in place
+    // and set_parameter()/get_parameter() will throw UMAMI_ERROR_NOT_IMPLEMENTED
+    // if called anyway).
+    bool supports_set_parameter() const noexcept { return bool(_functions.set_parameter); }
+    bool supports_get_parameter() const noexcept { return bool(_functions.get_parameter); }
+
+    // Directly sets/reads physics parameters according to whatever
+    // convention the library has implemented.
+    void set_parameter(const std::string& name, double real, double imag = 0.0) const;
+    std::pair<double, double> get_parameter(const std::string& name) const;
+
 private:
     MatrixElementApi(
         const std::string& file,
@@ -208,6 +231,7 @@ class MatrixElementApi {
     ThreadResource<InstanceType> _instances;
     std::string _file_name;
     std::size_t _index;
+    DevicePtr _device;
 
     friend class Context;
 };
diff --git a/madspace/include/madspace/driver/io.hpp b/madspace/include/madspace/driver/io.hpp
index 4aca6da092..8366a34729 100644
--- a/madspace/include/madspace/driver/io.hpp
+++ b/madspace/include/madspace/driver/io.hpp
@@ -20,6 +20,17 @@ Tensor load_tensor(TarReader& tar, const std::string& name);
 /// Append `tensor` as an npy file called `name` to the tar archive `tar`.
 void save_tensor(TarWriter& tar, const std::string& name, Tensor tensor);
 
+// Peeks a .npy event file's own field descriptor (its header's "descr"
+// entry) without needing to already know its layout -- returns the field
+// names in declaration order (particle-level fields duplicated per particle
+// slot with a "partN_" prefix, matching full_descr()'s own naming). Used
+// where a file's actual layout has caller-optional fields the reader needs
+// to auto-detect before committing to a DataLayout to open it with (eg
+// LHE event records, which may or may not carry
+// x1/fact_scale1/x2/fact_scale2/partial_weight_product depending on whether
+// [generation] systematics was enabled when the file was written).
+std::vector<std::string> peek_npy_field_names(const std::string& file);
+
 struct FieldLayout {
     enum Type { i32, f64 };
 
@@ -216,6 +227,7 @@ class EventRecord {
     static constexpr int f_partial_weights = 64;
     static constexpr int f_syst_weights = 128;
     static constexpr int f_bias = 256;
+    static constexpr int f_matrix_element = 512;
 
     // `syst_weight_count` columns of systematic variation weights are appended
     // when f_syst_weights is set; `syst_weight_ids` names them (rwgt_<id>).
@@ -282,6 +294,9 @@ class EventRecord {
         if (fields & f_bias) {
             ret.push_back({"bias", FieldLayout::f64, 8});
         }
+        if (fields & f_matrix_element) {
+            ret.push_back({"matrix_element", FieldLayout::f64, 9});
+        }
         return ret;
     }
 
@@ -328,6 +343,9 @@ class EventRecord {
     // event bias, divided out of the weight when the channels are combined
     UnalignedRef<double> bias() { return &_data[_offsets[7] + 0]; }
 
+    // squared matrix element the event was generated with, before PDFs and flux
+    UnalignedRef<double> matrix_element() { return &_data[_offsets[8] + 0]; }
+
     void from_lhe_event(const LHEEvent& event) {
         lhe_process_id() = event.process_id;
         lhe_weight() = event.weight;
@@ -486,6 +504,11 @@ class EventFile {
     void clear();
     std::size_t particle_count() const { return _particle_count; }
     std::size_t event_count() const { return _event_count; }
+    // The DataLayout this file was actually constructed with -- lets a
+    // reader that only has an already-open EventFile& (not the layout that
+    // was used to open it) build a matching EventBuffer, rather than having
+    // to independently know/recompute the same layout itself.
+    const DataLayout& layout() const { return _layout; }
     ~EventFile();
 
     void write(EventBuffer& buffer) {
@@ -533,6 +556,7 @@ class EventFile {
     std::size_t _event_size;
     Mode _mode;
     bool _delete_on_close;
+    DataLayout _layout;
 };
 
 } // namespace madspace
diff --git a/madspace/include/madspace/phasespace/cross_section.hpp b/madspace/include/madspace/phasespace/cross_section.hpp
index 6143b95f30..d30b7629c8 100644
--- a/madspace/include/madspace/phasespace/cross_section.hpp
+++ b/madspace/include/madspace/phasespace/cross_section.hpp
@@ -62,6 +62,9 @@ class DifferentialCrossSection : public FunctionGenerator {
      *                                 rather than reconstructed.
      * @param decay                    If true, describe the decay of a single
      *                                 incoming particle instead of a collision.
+     * @param matrix_element_output    If true, also return the squared matrix
+     *                                 element itself (before PDFs and flux) as
+     *                                 a last output, `raw_matrix_element`.
      */
     DifferentialCrossSection(
         const MatrixElement& matrix_element,
@@ -73,7 +76,8 @@ class DifferentialCrossSection : public FunctionGenerator {
         const std::variant<std::monostate, PdfGrid, CachedPdf>& pdf1 = std::monostate{},
         const std::variant<std::monostate, PdfGrid, CachedPdf>& pdf2 = std::monostate{},
         bool input_momentum_fraction = true,
-        bool decay = false
+        bool decay = false,
+        bool matrix_element_output = false
     );
 
     /// The allowed initial-parton flavour combinations.
@@ -86,6 +90,8 @@ class DifferentialCrossSection : public FunctionGenerator {
     const std::optional<RunningCoupling>& running_coupling() const {
         return _running_coupling;
     }
+    /// Whether the squared matrix element is returned as a last output.
+    bool matrix_element_output() const { return _matrix_element_output; }
 
 private:
     NamedVector<Value> build_function_impl(
@@ -101,6 +107,7 @@ class DifferentialCrossSection : public FunctionGenerator {
     std::variant<std::monostate, EnergyScale, CachedScale> _energy_scale;
     bool _input_momentum_fraction;
     bool _decay;
+    bool _matrix_element_output;
 };
 
 } // namespace madspace
diff --git a/madspace/include/madspace/umami.h b/madspace/include/madspace/umami.h
index 91f3adc413..8510845910 100644
--- a/madspace/include/madspace/umami.h
+++ b/madspace/include/madspace/umami.h
@@ -51,6 +51,8 @@ typedef enum {
     UMAMI_ERROR_UNSUPPORTED_META,
     /** a mandatory matrix element input was not provided */
     UMAMI_ERROR_MISSING_INPUT,
+    /** the provided parameter name is not a known, settable parameter */
+    UMAMI_ERROR_UNKNOWN_PARAMETER,
 } UmamiStatus;
 
 typedef enum {
@@ -222,7 +224,7 @@ UmamiStatus umami_supported_outputs(bool const** supported, int* count);
  *
  * @param handle
  *     pointer to an instance of the subprocess. Has to be cleaned up by
- *     the caller with `free_subprocess`.
+ *     the caller with `umami_free`.
  * @param param_card_path
  *     path to the parameter file
  * @return
@@ -242,7 +244,8 @@ UmamiStatus umami_initialize(UmamiHandle* handle, char const* param_card_path);
  * @param parameter_imag
  *     imaginary part of the parameter value. Ignored for real valued parameters.
  * @return
- *     UMAMI_SUCCESS on success, error code otherwise
+ *     UMAMI_SUCCESS on success, UMAMI_ERROR_UNKNOWN_PARAMETER if "name" is not
+ *     a settable parameter, error code otherwise
  */
 UmamiStatus umami_set_parameter(
     UmamiHandle handle, char const* name, double parameter_real, double parameter_imag
@@ -261,7 +264,8 @@ UmamiStatus umami_set_parameter(
  *     pointer to double to return imaginary part of the parameter value. Ignored
  *     for real-valued parameters (i.e. you may pass a null pointer)
  * @return
- *     UMAMI_SUCCESS on success, error code otherwise
+ *     UMAMI_SUCCESS on success, UMAMI_ERROR_UNKNOWN_PARAMETER if "name" is not
+ *     a known parameter, error code otherwise
  */
 UmamiStatus umami_get_parameter(
     UmamiHandle handle, char const* name, double* parameter_real, double* parameter_imag
diff --git a/madspace/src/driver/backend.cpp b/madspace/src/driver/backend.cpp
index 313dab2b61..4edc7ddb53 100644
--- a/madspace/src/driver/backend.cpp
+++ b/madspace/src/driver/backend.cpp
@@ -26,7 +26,14 @@ struct LoadedBackend {
 #endif
         shared_lib = std::shared_ptr<void>(
             dlopen(std::format("{}/{}.{}", lib_path, file, so_ext).c_str(), RTLD_NOW),
-            [](void* lib) { dlclose(lib); }
+            // Unlike std::unique_ptr, std::shared_ptr's deleter is invoked
+            // unconditionally on destruction, even when the stored pointer is
+            // null -- if dlopen() above fails, the throw a few lines down
+            // would otherwise unwind through a dlclose(nullptr), which
+            // segfaults instead of letting the runtime_error propagate.
+            [](void* lib) {
+                if (lib) dlclose(lib);
+            }
         );
         if (!shared_lib) {
             throw std::runtime_error(
diff --git a/madspace/src/driver/channel_generator.cpp b/madspace/src/driver/channel_generator.cpp
index fbdbf7ffb7..0006d2c968 100644
--- a/madspace/src/driver/channel_generator.cpp
+++ b/madspace/src/driver/channel_generator.cpp
@@ -24,6 +24,9 @@ int event_extra_flags(const std::unordered_map<std::string, std::size_t>& index_
     if (index_map.contains("bias")) {
         flags |= EventRecord::f_bias;
     }
+    if (index_map.contains("matrix_element")) {
+        flags |= EventRecord::f_matrix_element;
+    }
     return flags;
 }
 
@@ -357,6 +360,8 @@ void ChannelEventGenerator::init_field_indices() {
         _field_indices.subprocess_index = -1;
     }
     _field_indices.bias = index_map.contains("bias") ? index_map.at("bias") : -1;
+    _field_indices.matrix_element =
+        index_map.contains("matrix_element") ? index_map.at("matrix_element") : -1;
     _field_indices.random = index_map.at("random");
     _field_indices.rest = _field_indices.random + 1;
 }
@@ -907,6 +912,14 @@ void ChannelEventGenerator::write_events(
         }
     }
 
+    if (_field_indices.matrix_element != -1) {
+        auto me_view =
+            unweighted_events.at(_field_indices.matrix_element).view<double, 1>();
+        for (std::size_t i = 0; i < me_view.size(); ++i) {
+            event_buffer.event(i).matrix_element() = me_view[i];
+        }
+    }
+
     if (_field_indices.subprocess_index != -1) {
         auto subproc_view =
             unweighted_events.at(_field_indices.subprocess_index).view<me_int_t, 1>();
diff --git a/madspace/src/driver/context.cpp b/madspace/src/driver/context.cpp
index 398fe70cc5..2c25ae9975 100644
--- a/madspace/src/driver/context.cpp
+++ b/madspace/src/driver/context.cpp
@@ -36,7 +36,7 @@ MatrixElementApi::MatrixElementApi(
     std::size_t index,
     const std::unordered_map<std::string, double>& parameters
 ) :
-    _file_name(file), _index(index) {
+    _file_name(file), _index(index), _device(device) {
     _shared_lib = std::unique_ptr<void, std::function<void(void*)>>(
         dlopen(file.c_str(), RTLD_NOW), [file](void* lib) { dlclose(lib); }
     );
@@ -53,6 +53,7 @@ MatrixElementApi::MatrixElementApi(
     resolve_symbol(lib, "umami_supported_outputs", _functions.supported_outputs);
     resolve_symbol(lib, "umami_initialize", _functions.initialize);
     resolve_symbol(lib, "umami_set_parameter", _functions.set_parameter);
+    resolve_symbol(lib, "umami_get_parameter", _functions.get_parameter);
     resolve_symbol(lib, "umami_matrix_element", _functions.matrix_element);
     resolve_symbol(lib, "umami_free", _functions.free);
     init_functions(parameters);
@@ -148,6 +149,8 @@ void MatrixElementApi::check_umami_status(UmamiStatus status) const {
         throw_error("unsupported metadata key");
     case UMAMI_ERROR_MISSING_INPUT:
         throw_error("missing input");
+    case UMAMI_ERROR_UNKNOWN_PARAMETER:
+        throw_error("unknown or non-settable parameter");
     default:
         throw_error("unknown error");
     }
@@ -159,6 +162,22 @@ void MatrixElementApi::throw_error(const std::string& message) const {
     );
 }
 
+void MatrixElementApi::set_parameter(const std::string& name, double real, double imag) const {
+    if (!_functions.set_parameter) {
+        check_umami_status(UMAMI_ERROR_NOT_IMPLEMENTED);
+    }
+    check_umami_status(_functions.set_parameter(process_instance(), name.c_str(), real, imag));
+}
+
+std::pair<double, double> MatrixElementApi::get_parameter(const std::string& name) const {
+    if (!_functions.get_parameter) {
+        check_umami_status(UMAMI_ERROR_NOT_IMPLEMENTED);
+    }
+    double real = 0.0, imag = 0.0;
+    check_umami_status(_functions.get_parameter(process_instance(), name.c_str(), &real, &imag));
+    return {real, imag};
+}
+
 const MatrixElementApi& Context::load_matrix_element(
     const std::string& file,
     const std::string& param_card,
diff --git a/madspace/src/driver/event_generator.cpp b/madspace/src/driver/event_generator.cpp
index da22f8d241..44625e2558 100644
--- a/madspace/src/driver/event_generator.cpp
+++ b/madspace/src/driver/event_generator.cpp
@@ -1041,6 +1041,9 @@ void EventGenerator::copy_event_fields(
     if (contains("partial_weight_product")) {
         event_out.partial_weight_product() = event_in.partial_weight_product();
     }
+    if (contains("matrix_element")) {
+        event_out.matrix_element() = event_in.matrix_element();
+    }
     if (_systematics) {
         std::size_t count = _systematics->weight_count();
         for (std::size_t k = 0; k < count; ++k) {
@@ -1222,6 +1225,8 @@ void EventGenerator::read_and_combine(
     bool has_subproc_index =
         _channels.at(0)->event_layout_extra_flags() & EventRecord::f_subproc_index;
     bool has_bias = _channels.at(0)->event_layout_extra_flags() & EventRecord::f_bias;
+    bool has_matrix_element =
+        _channels.at(0)->event_layout_extra_flags() & EventRecord::f_matrix_element;
 
     for (std::size_t event_index = 0; event_index < event_count; ++event_index) {
         std::size_t random_index = rand_gen.generate_int(channel_data.back().cum_count);
@@ -1288,6 +1293,9 @@ void EventGenerator::read_and_combine(
         if (has_partial) {
             event_out.partial_weight_product() = event_in.partial_weight_product();
         }
+        if (has_matrix_element) {
+            event_out.matrix_element() = event_in.matrix_element();
+        }
 
         std::size_t i = 0;
         for (; i < sampled_chan->event_buffer.particle_count(); ++i) {
diff --git a/madspace/src/driver/io.cpp b/madspace/src/driver/io.cpp
index 87c5a79e5a..d4234ea0f7 100644
--- a/madspace/src/driver/io.cpp
+++ b/madspace/src/driver/io.cpp
@@ -2,6 +2,7 @@
 
 #include <algorithm>
 #include <cctype>
+#include <cstdint>
 #include <filesystem>
 
 #include <nlohmann/json.hpp>
@@ -36,17 +37,46 @@ DataType str_to_dtype(std::string dtype) {
     }
 }
 
+// The header-length field's width depends on the npy format version: 1.0
+// uses 2 bytes (uint16, so the header itself must fit in 65535 bytes), 2.0/3.0
+// use 4 bytes (uint32) -- write_event_header() switches to 2.0 whenever the
+// header would otherwise be too long to describe in 2 bytes (eg several "add
+// process" channels of different final-state multiplicity combined into one
+// binary EventFile can easily need a header longer than 65535 bytes' worth of
+// field descriptors). Only ever reading the 2-byte case silently truncated the
+// *declared* length for any such oversized header, cutting a field name off
+// mid-string for every reader (this one and numpy alike).
 constexpr std::size_t npy_preamble_size = 10;
+constexpr std::size_t npy_wide_preamble_size = 12;
 
-/// Check the magic number and version of the first 10 bytes of an npy file and
-/// return the length of the header that follows.
-std::size_t parse_preamble(const char* bytes) {
-    if (std::memcmp(bytes, "\x93NUMPY\x01\x00", 8) != 0) {
+/// Check the magic number and version (the first 8 bytes) of an npy file and
+/// return the size of its preamble: 10 bytes for format 1.0, 12 for 2.0/3.0.
+std::size_t preamble_size(const char* bytes) {
+    if (std::memcmp(bytes, "\x93NUMPY", 6) != 0) {
+        throw std::runtime_error("invalid header of npy file");
+    }
+    switch (static_cast<unsigned char>(bytes[6])) {
+    case 1:
+        return npy_preamble_size;
+    case 2:
+    case 3:
+        return npy_wide_preamble_size;
+    default:
         throw std::runtime_error("invalid header of npy file");
     }
-    // little-endian 16 bit length
-    return static_cast<std::size_t>(static_cast<unsigned char>(bytes[8])) |
-        (static_cast<std::size_t>(static_cast<unsigned char>(bytes[9])) << 8);
+}
+
+/// Return the length of the header that follows the preamble of an npy file.
+/// bytes must hold the full preamble (see preamble_size()).
+std::size_t parse_preamble(const char* bytes) {
+    // little-endian 16 or 32 bit length
+    std::size_t length_bytes = preamble_size(bytes) - 8;
+    std::size_t length = 0;
+    for (std::size_t i = 0; i < length_bytes; ++i) {
+        length |= static_cast<std::size_t>(static_cast<unsigned char>(bytes[8 + i]))
+            << (8 * i);
+    }
+    return length;
 }
 
 json parse_header_text(std::string header) {
@@ -82,8 +112,12 @@ json parse_header_text(std::string header) {
 }
 
 json parse_header(std::fstream& file_stream) {
-    char preamble[npy_preamble_size];
-    file_stream.read(preamble, npy_preamble_size);
+    char preamble[npy_wide_preamble_size];
+    file_stream.read(preamble, 8);
+    if (file_stream.fail()) {
+        throw std::runtime_error("invalid header of npy file");
+    }
+    file_stream.read(preamble + 8, preamble_size(preamble) - 8);
     if (file_stream.fail()) {
         throw std::runtime_error("invalid header of npy file");
     }
@@ -157,6 +191,25 @@ full_descr(std::size_t particle_count, const DataLayout& layout) {
     return field_descr;
 }
 
+} // namespace
+
+std::vector<std::string> madspace::peek_npy_field_names(const std::string& file) {
+    std::fstream file_stream(file, std::ios::binary | std::ios::in);
+    if (!file_stream) {
+        throw std::runtime_error(std::format("Could not open file {}", file));
+    }
+    json header = parse_header(file_stream);
+    json descr = header.at("descr");
+    std::vector<std::string> names;
+    names.reserve(descr.size());
+    for (auto& item : descr) {
+        names.push_back(item.at(0).get<std::string>());
+    }
+    return names;
+}
+
+namespace {
+
 std::tuple<std::size_t, std::size_t, std::size_t>
 read_event_header(std::fstream& file_stream, const DataLayout& layout) {
     json header = parse_header(file_stream);
@@ -208,28 +261,62 @@ std::tuple<std::size_t, std::size_t> write_event_header(
     std::size_t header_size = 0
 ) {
     using namespace std::string_literals;
-    file_stream << "\x93NUMPY\x01\x00\x00\x00{'descr':["s;
+    // Built into a string first (rather than written straight to
+    // file_stream, as this function used to) so its true length is known
+    // before deciding whether npy format 1.0's 2-byte header-length field
+    // (max 65535) is wide enough, or format 2.0's 4-byte field is needed
+    // instead -- see parse_header()'s matching version-selection logic for
+    // why: silently wrapping an oversized length into 16 bits (the old
+    // behavior here) corrupts the file for every reader, this project's own
+    // and numpy's alike. A large particle_count (eg several "add process"
+    // channels of different final-state multiplicity combined into one
+    // binary EventFile) easily needs a header longer than 65535 bytes'
+    // worth of field descriptors.
+    std::string content = "{'descr':["s;
     auto field_descr = full_descr(particle_count, layout);
     for (auto [field_name, field_type] : field_descr) {
-        file_stream << std::format("('{}','{}'),", field_name, field_type);
+        content += std::format("('{}','{}'),", field_name, field_type);
     }
-    file_stream << "],'fortran_order':False,'shape':(";
-    std::size_t shape_pos = file_stream.tellp();
+    content += "],'fortran_order':False,'shape':(";
+
+    // prefix_len: magic (6) + major/minor version (2) + header-length field
+    // (2 for v1.0, 4 for v2.0) -- header_size is always the *absolute* file
+    // offset where the header ends and event data begins (matching
+    // EventFile's own _header_size-based seek math), regardless of format.
+    bool wide = false;
+    std::size_t prefix_len = 10;
     if (header_size == 0) {
-        header_size = (shape_pos + 100) / 64 * 64;
+        header_size = (prefix_len + content.size() + 100) / 64 * 64;
+        if (header_size - prefix_len > 0xFFFF) {
+            wide = true;
+            prefix_len = 12;
+            header_size = (prefix_len + content.size() + 100) / 64 * 64;
+        }
+    } else {
+        wide = (header_size - prefix_len > 0xFFFF);
+        if (wide) prefix_len = 12;
     }
-    for (int i = shape_pos; i < header_size - 1; ++i) {
+
+    file_stream << (wide ? "\x93NUMPY\x02\x00"s : "\x93NUMPY\x01\x00"s);
+    if (wide) {
+        uint32_t zero32 = 0;
+        file_stream.write(reinterpret_cast<char*>(&zero32), 4);
+    } else {
+        uint16_t zero16 = 0;
+        file_stream.write(reinterpret_cast<char*>(&zero16), 2);
+    }
+    file_stream << content;
+    std::size_t shape_pos = file_stream.tellp();
+    for (std::size_t i = shape_pos; i < header_size - 1; ++i) {
         file_stream.put(' ');
     }
     file_stream.put('\n');
     file_stream.seekp(8);
-    // little-endian 16 bit length
-    std::size_t header_length = header_size - npy_preamble_size;
-    char length_bytes[2] = {
-        static_cast<char>(header_length & 0xff),
-        static_cast<char>((header_length >> 8) & 0xff)
-    };
-    file_stream.write(length_bytes, 2);
+    // little-endian 16 bit (v1.0) or 32 bit (v2.0) length
+    std::size_t header_length = header_size - prefix_len;
+    for (std::size_t i = 0; i < prefix_len - 8; ++i) {
+        file_stream.put(static_cast<char>((header_length >> (8 * i)) & 0xff));
+    }
     file_stream.seekp(header_size);
     return {header_size, shape_pos};
 }
@@ -262,16 +349,18 @@ void madspace::save_tensor(const std::string& file, Tensor tensor) {
 
 Tensor madspace::load_tensor(TarReader& tar, const std::string& name) {
     std::vector<char> data = tar.read(name);
-    if (data.size() < npy_preamble_size) {
+    if (data.size() < npy_preamble_size ||
+        data.size() < preamble_size(data.data())) {
         throw std::runtime_error(std::format("Invalid npy file '{}'", name));
     }
+    std::size_t preamble_length = preamble_size(data.data());
     std::size_t header_length = parse_preamble(data.data());
-    std::size_t data_offset = npy_preamble_size + header_length;
+    std::size_t data_offset = preamble_length + header_length;
     if (data.size() < data_offset) {
         throw std::runtime_error(std::format("Invalid npy file '{}'", name));
     }
     auto [dtype, shape] = parse_tensor_header(
-        parse_header_text(std::string(data.data() + npy_preamble_size, header_length))
+        parse_header_text(std::string(data.data() + preamble_length, header_length))
     );
     Tensor tensor(dtype, shape);
     if (data.size() - data_offset != tensor.byte_size()) {
@@ -304,7 +393,8 @@ EventFile::EventFile(
     _particle_count(particle_count),
     _event_size(layout.event_size() + particle_count * layout.particle_size()),
     _mode(mode),
-    _delete_on_close(delete_on_close) {
+    _delete_on_close(delete_on_close),
+    _layout(layout) {
     auto file_mode = std::ios::binary | std::ios::in;
     if (mode == EventFile::create) {
         file_mode |= std::ios::out | std::ios::trunc;
diff --git a/madspace/src/phasespace/cross_section.cpp b/madspace/src/phasespace/cross_section.cpp
index f5f46b31cd..bea2fbabbb 100644
--- a/madspace/src/phasespace/cross_section.cpp
+++ b/madspace/src/phasespace/cross_section.cpp
@@ -11,7 +11,8 @@ DifferentialCrossSection::DifferentialCrossSection(
     const std::variant<std::monostate, PdfGrid, CachedPdf>& pdf1,
     const std::variant<std::monostate, PdfGrid, CachedPdf>& pdf2,
     bool input_momentum_fraction,
-    bool decay
+    bool decay,
+    bool matrix_element_output
 ) :
     FunctionGenerator(
         "DifferentialCrossSection",
@@ -50,7 +51,13 @@ DifferentialCrossSection::DifferentialCrossSection(
             }
             return arg_types;
         }(),
-        matrix_element.return_types()
+        [&] {
+            auto ret_types = matrix_element.return_types();
+            if (matrix_element_output) {
+                ret_types.push_back("raw_matrix_element", batch_float);
+            }
+            return ret_types;
+        }()
     ),
     _pid_options(pid_options),
     _matrix_element(matrix_element),
@@ -62,7 +69,8 @@ DifferentialCrossSection::DifferentialCrossSection(
     _e_cm(cm_energy),
     _energy_scale(energy_scale),
     _input_momentum_fraction(input_momentum_fraction),
-    _decay(decay) {
+    _decay(decay),
+    _matrix_element_output(matrix_element_output) {
     if (decay) {
         if (_has_pdf.at(0) || _has_pdf.at(1)) {
             throw std::invalid_argument("a decay has no beams, so no PDFs");
@@ -157,6 +165,7 @@ NamedVector<Value> DifferentialCrossSection::build_function_impl(
         throw std::runtime_error("matrix element missing in return values");
     }
     std::size_t me_index = search - _matrix_element.outputs().begin();
+    Value matrix_element = me_result.at(me_index);
     if (_decay) {
         // Differential decay rate: dGamma = |M|^2 dPhi_n / (2 M), in GeV. The
         // flux is a compile-time constant here (M is fixed by the decaying
@@ -175,5 +184,8 @@ NamedVector<Value> DifferentialCrossSection::build_function_impl(
             _e_cm * _e_cm
         );
     }
+    if (_matrix_element_output) {
+        me_result.push_back("raw_matrix_element", matrix_element);
+    }
     return me_result;
 }
diff --git a/madspace/src/phasespace/integrand.cpp b/madspace/src/phasespace/integrand.cpp
index 5533fc48f1..5e2d6ac03d 100644
--- a/madspace/src/phasespace/integrand.cpp
+++ b/madspace/src/phasespace/integrand.cpp
@@ -230,6 +230,9 @@ Integrand::Integrand(
                         ret_types.push_back(name, batch_float_array(2));
                     }
                 }
+                if (diff_xs_first.matrix_element_output()) {
+                    ret_types.push_back("matrix_element", batch_float);
+                }
                 // before "random": the outputs after it feed the discrete histograms
                 if (custom_bias) {
                     ret_types.push_back("bias", batch_float);
@@ -1405,6 +1408,11 @@ NamedVector<Value> Integrand::build_common_part(
             for (int k = 0; k < _mlm_history_qcd_power; ++k) {
                 factor = fb.mul(factor, alpha_ratio);
             }
+            if (_diff_xs.at(0).matrix_element_output()) {
+                // the stored squared matrix element is at the event's alpha_s
+                // too, without the beam densities
+                dxs_vec.back() = fb.mul(dxs_vec.back(), factor);
+            }
             for (std::size_t i = 0; i < 2; ++i) {
                 if (_diff_xs.at(0).has_pdf(i)) {
                     // a first density of exactly zero gave a zero weight, which
@@ -1660,6 +1668,11 @@ NamedVector<Value> Integrand::build_common_part(
                 )
             );
         }
+        if (_diff_xs.at(0).matrix_element_output()) {
+            outputs.push_back(
+                "matrix_element", scatter_or_drop(zeros_float, dxs_vec.back())
+            );
+        }
         if (bias_acc) {
             outputs.push_back(
                 "bias", scatter_or_drop(fb.full({1., batch_size_val}), bias_acc)
diff --git a/madspace/src/python/madspace.cpp b/madspace/src/python/madspace.cpp
index 9a46c0c1a1..5f979924e0 100644
--- a/madspace/src/python/madspace.cpp
+++ b/madspace/src/python/madspace.cpp
@@ -337,7 +337,21 @@ PYBIND11_MODULE(_madspace_py, m) {
             &MatrixElementApi::helicity_count,
             pydoc::doc("MatrixElementApi::helicity_count")
         )
-        .def("index", &MatrixElementApi::index, pydoc::doc("MatrixElementApi::index"));
+        .def("index", &MatrixElementApi::index, pydoc::doc("MatrixElementApi::index"))
+        .def("file_name", &MatrixElementApi::file_name)
+        .def("supported_inputs", &MatrixElementApi::supported_inputs)
+        .def("required_inputs", &MatrixElementApi::required_inputs)
+        .def("supported_outputs", &MatrixElementApi::supported_outputs)
+        .def("supports_set_parameter", &MatrixElementApi::supports_set_parameter)
+        .def("supports_get_parameter", &MatrixElementApi::supports_get_parameter)
+        .def(
+            "set_parameter",
+            &MatrixElementApi::set_parameter,
+            py::arg("name"),
+            py::arg("real"),
+            py::arg("imag") = 0.0
+        )
+        .def("get_parameter", &MatrixElementApi::get_parameter, py::arg("name"));
 
     py::classh<Tensor>(m, "Tensor", py::dynamic_attr(), pydoc::doc("Tensor"))
         .def(
@@ -2363,6 +2377,7 @@ PYBIND11_MODULE(_madspace_py, m) {
                     PdfGrid,
                     DifferentialCrossSection::CachedPdf>&,
                 bool,
+                bool,
                 bool>(),
             py::arg("matrix_element"),
             py::arg("cm_energy"),
@@ -2373,6 +2388,7 @@ PYBIND11_MODULE(_madspace_py, m) {
             py::arg("pdf2") = std::monostate{},
             py::arg("input_momentum_fraction") = true,
             py::arg("decay") = false,
+            py::arg("matrix_element_output") = false,
             pydoc::doc("DifferentialCrossSection::DifferentialCrossSection")
         )
         .def(
@@ -2395,6 +2411,11 @@ PYBIND11_MODULE(_madspace_py, m) {
             "running_coupling",
             &DifferentialCrossSection::running_coupling,
             pydoc::doc("DifferentialCrossSection::running_coupling")
+        )
+        .def(
+            "matrix_element_output",
+            &DifferentialCrossSection::matrix_element_output,
+            pydoc::doc("DifferentialCrossSection::matrix_element_output")
         );
 
     py::classh<Unweighter, FunctionGenerator>(m, "Unweighter", pydoc::doc("Unweighter"))
diff --git a/madtrex/.gitignore b/madtrex/.gitignore
new file mode 100644
index 0000000000..ef8efef4c7
--- /dev/null
+++ b/madtrex/.gitignore
@@ -0,0 +1,5 @@
+build/
+install/
+install_settings.json
+*.egg-info/
+__pycache__/
diff --git a/madtrex/CMakeLists.txt b/madtrex/CMakeLists.txt
new file mode 100644
index 0000000000..138c43ddc9
--- /dev/null
+++ b/madtrex/CMakeLists.txt
@@ -0,0 +1,151 @@
+cmake_minimum_required (VERSION 3.15...3.27)
+project (
+    madtrex
+    VERSION 0.1
+    LANGUAGES CXX
+)
+
+# MadtRex is layered strictly on top of madspace: it builds against an
+# already-installed madspace (headers + libmadspace) and the bundled Rex fork
+# in Rex/, and madspace knows nothing about it.
+
+set(CMAKE_POSITION_INDEPENDENT_CODE 1)
+
+if (CMAKE_BUILD_TYPE STREQUAL "Release")
+    include(CheckIPOSupported)
+    check_ipo_supported(RESULT ipo_supported OUTPUT ipo_error)
+    if(ipo_supported)
+        set(CMAKE_INTERPROCEDURAL_OPTIMIZATION TRUE)
+        message(STATUS "Enabling LTO")
+    endif()
+endif()
+
+add_compile_options(-fno-omit-frame-pointer)
+
+########################################################################################
+# Load dependencies                                                                    #
+########################################################################################
+
+find_package(Python REQUIRED COMPONENTS Interpreter Development.Module)
+include(FetchContent)
+
+# Must stay identical to madspace/CMakeLists.txt (and Rex/CMakeLists.txt):
+# _madtrex_py takes and returns types registered by _madspace_py and pyrex,
+# and pybind11 only shares registered types between modules built against the
+# same internals version.
+FetchContent_Declare(
+    pybind11
+    GIT_REPOSITORY https://github.com/pybind/pybind11
+    GIT_TAG        v3.0.1
+)
+FetchContent_MakeAvailable(pybind11)
+
+FetchContent_Declare(
+    json
+    URL https://github.com/nlohmann/json/releases/download/v3.11.3/json.tar.xz
+)
+FetchContent_MakeAvailable(json)
+
+find_package(Threads REQUIRED)
+
+# The installed madspace package (the directory holding its __init__.py, with
+# include/ and lib/ inside it). Found by importing madspace with the build's
+# Python unless given explicitly with -DMADSPACE_DIR=...
+if(NOT MADSPACE_DIR)
+    execute_process(
+        COMMAND "${Python_EXECUTABLE}" -c
+            "import madspace, os; print(os.path.dirname(os.path.abspath(madspace.__file__)))"
+        OUTPUT_VARIABLE MADSPACE_DIR
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+        RESULT_VARIABLE madspace_import_result
+        ERROR_VARIABLE madspace_import_error
+    )
+    if(NOT madspace_import_result EQUAL 0)
+        message(FATAL_ERROR
+            "Could not import madspace with ${Python_EXECUTABLE}:\n"
+            "${madspace_import_error}\n"
+            "Install madspace first (python madspace/install.py) and make it "
+            "importable (e.g. PYTHONPATH=madspace/install), or pass "
+            "-DMADSPACE_DIR=/path/to/site-packages/madspace."
+        )
+    endif()
+endif()
+set(MADSPACE_DIR "${MADSPACE_DIR}" CACHE PATH "Installed madspace package directory")
+message(STATUS "Using madspace from ${MADSPACE_DIR}")
+
+find_library(
+    MADSPACE_LIBRARY
+    NAMES madspace
+    PATHS "${MADSPACE_DIR}/lib"
+    NO_DEFAULT_PATH
+    REQUIRED
+)
+add_library(madspace::madspace SHARED IMPORTED)
+set_target_properties(
+    madspace::madspace
+    PROPERTIES
+        IMPORTED_LOCATION "${MADSPACE_LIBRARY}"
+        INTERFACE_INCLUDE_DIRECTORIES "${MADSPACE_DIR}/include;${MADSPACE_DIR}/include/madspace/compgraphs"
+        INTERFACE_COMPILE_FEATURES cxx_std_20
+)
+
+########################################################################################
+# Python bindings                                                                      #
+########################################################################################
+
+python_add_library(
+    _madtrex_py
+    MODULE
+    src/python/madtrex.cpp
+    src/tupperware.cpp
+    src/madtrex.cpp
+    src/binary_io.cpp
+    src/param_handler.cpp
+    src/pdfware.cpp
+    Rex/src/Rex.cc
+    Rex/src/teaRex.cc
+    WITH_SOABI
+)
+target_include_directories(_madtrex_py PRIVATE include Rex/include)
+target_compile_features(_madtrex_py PRIVATE cxx_std_20)
+target_link_libraries(_madtrex_py PRIVATE pybind11::headers)
+target_link_libraries(_madtrex_py PRIVATE madspace::madspace)
+target_link_libraries(_madtrex_py PRIVATE nlohmann_json::nlohmann_json)
+target_link_libraries(_madtrex_py PRIVATE Threads::Threads)
+# madtrex/__init__.py imports madspace (which preloads libmadspace) before
+# _madtrex_py, but point the RPATH at libmadspace as well, so the module also
+# loads on its own: first relative to the module, for the side-by-side
+# madtrex/install/madtrex <-> madspace/install/madspace layout that MadGraph
+# and its gridpacks use (and which survives moving the whole tree), then at
+# the very libmadspace this was built against.
+if(APPLE)
+    set(madtrex_origin "@loader_path")
+else()
+    set(madtrex_origin "$ORIGIN")
+endif()
+set_target_properties(
+    _madtrex_py
+    PROPERTIES
+        INSTALL_RPATH "${madtrex_origin}/../../../madspace/install/madspace/lib;${MADSPACE_DIR}/lib"
+        BUILD_RPATH "${MADSPACE_DIR}/lib"
+)
+install(TARGETS _madtrex_py DESTINATION madtrex)
+
+# The madspace headers _madtrex_py is compiled against (its classes and their
+# inline methods are baked in): record a digest of them, which
+# madtrex/_compat.py compares with the madspace found at import time, so that
+# a madspace reinstalled since is refused with a clear message instead of
+# failing as a binary mismatch. Same digest as _compat.headers_digest(): one
+# "<relative path> <sha256>" line per file, in sorted path order.
+file(GLOB_RECURSE madspace_headers LIST_DIRECTORIES false
+     RELATIVE "${MADSPACE_DIR}/include" "${MADSPACE_DIR}/include/*")
+list(SORT madspace_headers)
+set(madspace_headers_lines "")
+foreach(header IN LISTS madspace_headers)
+    file(SHA256 "${MADSPACE_DIR}/include/${header}" header_sha)
+    string(APPEND madspace_headers_lines "${header} ${header_sha}\n")
+endforeach()
+string(SHA256 MADSPACE_HEADERS_DIGEST "${madspace_headers_lines}")
+file(WRITE "${CMAKE_CURRENT_BINARY_DIR}/_build_info.json"
+     "{\"madspace_headers\": \"${MADSPACE_HEADERS_DIGEST}\", \"madspace_dir\": \"${MADSPACE_DIR}\"}\n")
+install(FILES "${CMAKE_CURRENT_BINARY_DIR}/_build_info.json" DESTINATION madtrex)
diff --git a/madtrex/README.md b/madtrex/README.md
new file mode 100644
index 0000000000..901c27798d
--- /dev/null
+++ b/madtrex/README.md
@@ -0,0 +1,320 @@
+# MadtRex
+
+MadtRex is MadGraph7's event-reweighting layer. It connects madspace's
+UMAMI matrix-element libraries (`madspace.MatrixElementApi`) to the
+Rex/teaRex LHE parsing and reweighting library, and provides:
+
+- `TupperWare` / `WareHouse`: turn loaded `MatrixElementApi` objects into
+  `REX::tea` weightors and build `procReweightor`s from them;
+- `Driver`: a JSON/card-driven front end for parameter reweighting
+  (`ParamHandler` reads the reweight cards);
+- `PdfDriver`: PDF and scale-variation reweighting;
+- readers/writers for madspace's binary (`.npy`) event format.
+
+MadGraph7 uses it for the reweighting step after event generation
+(`madgraph/iolibs/template_files/mg7/madtrex.py`), as one of two
+reweighting backends (see "Choosing a backend").
+
+## Layout
+
+```
+madtrex/
+  include/madtrex/   C++ headers
+  src/               C++ sources
+  src/python/        pybind11 bindings (the _madtrex_py module)
+  madtrex/           Python package
+  Rex/               bundled fork of Rex (Python package: pyrex)
+  install.py         installer
+```
+
+`Rex/` is a plain copy of https://github.com/zeniheisser/Rex (commit
+94b4ed435a4bddce2f041c210e0455f2eff01379), tracked as ordinary files in this
+repository; it no longer follows upstream. Rex is dual licensed: LGPL v3 (or
+later) as upstream, or, as part of MadGraph7, the MadGraph7 license (`LICENSE`
+at the top of MadGraph7).
+
+## Requirements
+
+- **madspace**, installed first (normally `python madspace/install.py`, which
+  installs into `madspace/install/`). MadtRex builds against its installed
+  headers and `libmadspace`, and imports it at runtime.
+- **pyrex**, the Python package of the bundled `Rex/` fork. `install.py`
+  builds and installs it for you.
+- A C++20 compiler, CMake >= 3.15 and Python >= 3.11. pybind11 and
+  nlohmann_json are fetched by CMake at the same versions madspace uses
+  (required for the two extension modules to share types).
+
+MadSpace does **not** depend on MadtRex: madspace builds, installs and
+imports without it, and without Rex/pyrex.
+
+## Installing
+
+```bash
+python madspace/install.py      # if madspace is not installed yet
+python madtrex/install.py
+```
+
+or, from the MadGraph prompt, `install madspace` followed by
+`install madtrex` (which forwards `nb_core` as the job count).
+
+This installs `pyrex` and `madtrex` into `madtrex/install/`. Options:
+`--yes` (no prompts), `-j N` (parallel jobs), `--clean` (rebuild from
+scratch), `--system` (install into the current Python environment, where
+madspace must then also be importable). To use it outside MadGraph:
+
+```bash
+export PYTHONPATH=madtrex/install:madspace/install
+python -c "import madtrex; print(madtrex.TupperWare)"
+```
+
+## Reweighting drivers
+
+- `madgraph/iolibs/template_files/mg7/madtrex_driver.py` is the standalone
+  driver: it needs only madspace and madtrex (and its three
+  `*_reweight_card.py` modules next to it), not MadGraph. It runs against an
+  MG7 process directory or a gridpack:
+
+  ```bash
+  python madgraph/iolibs/template_files/mg7/madtrex_driver.py \
+      --process-dir PROC --rwgt-card PROC/Cards/reweight_card.dat
+  ```
+
+  It adds the weights to the events file itself (see "Choosing a backend"),
+  or writes them to `--output PATH`. It finds `madspace/install` and
+  `madtrex/install` by walking up from its own location, or uses whatever
+  `madspace`/`madtrex` are already importable.
+  It may need `make`, to compile a matrix element at another precision when
+  the sources are there. It cannot handle `change model` cards.
+- `madgraph/iolibs/template_files/mg7/madtrex.py` is the MadGraph side, used
+  by MG7's reweighting step: the same driver, plus `change model` support
+  (regenerating the process in the new model with `bin/madgraph`). Run it
+  as `python -m madgraph.iolibs.template_files.mg7.madtrex`.
+
+## Choosing a backend
+
+An MG7 output can be reweighted by MadtRex or by MadGraph's classic
+reweighting module (`madgraph/interface/reweight_interface.py`, f2py-compiled
+matrix elements). Both read the same `Cards/reweight_card.dat`, with one
+vocabulary (the classic one): `change <option> <value>` for options of the
+reweighting, `set ...` for physics parameters. Options both backends have use
+the classic name and default.
+
+Who reweights is decided in this order (`madgraph/various/reweight_backend.py`):
+
+1. `generate_events --reweight-backend=classic|madtrex|auto`;
+2. the reweight switch of the launch question: `classic`, `madtrex`,
+   `density` or a plugin (classic). `ON` is auto;
+3. the card's `change backend classic|madtrex|auto` line;
+4. auto: a card with directives only the classic module understands
+   (`change mode NLO`, `change output`, density-matrix options, ...) goes to
+   the classic module, one with MadtRex-only directives (`change mode pdf`,
+   `change fptype`, ...) to MadtRex (both kinds at once is an error). An
+   output other than MG7, or NLO events, use the classic module. Otherwise
+   MadtRex reweights when it is installed and the run's `output_format` is
+   `lhe` or `lhe_npy`, else the classic module (when f2py is available).
+
+When two explicit choices disagree, the higher one wins, with a warning
+naming both. A choice that cannot run (MadtRex on a madevent output or on
+`compact_npy` events, the classic module without f2py, a card with the other
+backend's directives) stops the launch before the events are generated, with
+every reason. The run logs one line, e.g.
+`Reweighting with madtrex (auto: MG7 output, madtrex installed, no
+classic-only directives in reweight_card.dat)`, and records the decision in
+`Events/<run>/reweighting.json`.
+
+Both backends add their weights to the run's own event file, after the
+weights already there (the systematics', an earlier reweighting's):
+`Events/<run>/events.lhe.gz` (or `events.lhe`), or `Events/<run>/events.npy`
+for an `lhe_npy` run reweighted by MadtRex. The file is rewritten next to
+itself and replaced atomically (gzipped again if it was).
+
+- In an LHE file the weights are `<wgt>` entries of the `<rwgt>` block, in
+  the `mg_reweighting` weight group (classic module) or `tearex_reweighting`
+  (MadtRex) of `<initrwgt>`.
+- In `events.npy` each weight is a column `rwgt_<name>`, after the
+  systematics' `rwgt_<n>` columns, and `header.lhe` lists it in a
+  `tearex_reweighting` weight group. `npy_to_lhe` writes those weights into
+  the LHE file it makes.
+- The madevent tools' copy of the events (`unweighted_events.lhe.gz`, which
+  MadSpin and the shower read) is kept identical to the run's event file.
+- An `output_format = "lhe"` run that MadtRex reweights (also when `auto`
+  chooses it, and when a shower or MadSpin switched the run to `lhe`) is
+  generated as `lhe_npy`: MadtRex reads and extends `events.npy` much faster
+  than it parses and rewrites an LHE file. Once reweighted, `npy_to_lhe`
+  writes the requested `events.lhe.gz` from it, with every weight (the
+  systematics', MadtRex's), and `events.npy`/`header.lhe` are removed; the
+  user gets the same LHE file as before, in the format of an LHE run that is
+  not reweighted. When the reweighting fails, the LHE file is still written,
+  without its weights (and the failure reported, as for any run). A run
+  with an LHE-only step before the reweighting (`[postprocessing]`
+  `time_of_flight`, or the legacy `systematics`) is generated as LHE, as
+  before.
+- `reweighting.json` records the file's path (`events`).
+
+The standalone driver does the same unless `--output PATH` asks for the
+reweighted events (or, for npy input, the weight files) to be written
+elsewhere.
+
+Reweighting runs once per generation, after the systematics.
+`generate_events --reweight=CARD` is a shortcut for switching reweighting on
+with CARD as `Cards/reweight_card.dat` (a different card already there is
+kept as `reweight_card.dat.bak`).
+
+The card vocabulary:
+
+| line | backends |
+|---|---|
+| `change backend classic\|madtrex\|auto` | both (default auto) |
+| `change helicity True\|False` | both (default True) |
+| `change mode LO` | both: parameter reweighting, the default |
+| `change model NAME`, `change process DEF` | both (MadtRex needs both) |
+| `launch [--rwgt_name=NAME]`, `set BLOCK ID VALUE`, `set NAME VALUE`, `scan:[...]` | both |
+| `change mode pdf`, then `set PDF_SET\|BEAM1_PDF_SET\|BEAM2_PDF_SET\|MUF\|MUR\|MUF_FACTOR\|MUR_FACTOR` | MadtRex |
+| `change fptype`, `change verbose\|max_batch\|max_deferred_iterations\|streaming\|stream_batch_size`, `change nb_core\|min_batch` (see "Threads"), `change alpha_s_power\|scale_variation_output` | MadtRex |
+| `change mode NLO\|NLO_tree\|LO+NLO`, `change output`, `rwgt_dir`, `tree_path`, `virtual_path`, `systematics`, `multicore`, `boost`, `change process ... --add`, a param_card path, density-matrix options, plugins | classic |
+
+A card section starts with `change mode LO` or `change mode pdf`; lines before
+the first one are LO. The former MadtRex spellings (`set
+helicity_reweighting`, `set use_helicity`, `set FPTYPE`, `set mode
+parameter|pdf`, `set VERBOSE`, ...) are still accepted, with a deprecation
+warning. The standalone driver (`bin/reweight`) is MadtRex only: it accepts
+`change backend madtrex|auto` and stops on a card for the classic module,
+which needs MadGraph.
+
+## Threads
+
+MadtRex reweights on as many threads as the event generation uses: the
+`[run] cpu_thread_pool_size` of the run card (all available cores for -1,
+the default). The events are classified into their subprocesses in parallel,
+and each batch of events (one per subprocess, or a part of one) is
+reweighted on its own thread, the normalisation as well as each launch.
+
+A sample dominated by one subprocess would keep a single thread busy, so a
+subprocess is split into batches when it is large: with `N` events and `T`
+threads, every subprocess with more than `max(min_batch, N / T)` events is cut
+into batches of that size. A small sample, or a small subprocess, is not
+spread over all threads: below `min_batch` events (1000 by default) a batch
+costs more to set up -- its own matrix element instance, loaded from the
+library and the param_card, its own sorting -- than another thread saves,
+and a sample of fewer than `2 * min_batch` events is not split at all.
+
+Both are reweight-card options of the LO part, before the first `launch`:
+
+```
+change nb_core 4        # threads (auto: as the generation, the default)
+change min_batch 5000   # fewest events per batch when splitting
+```
+
+An explicit `change max_batch N` splits at N events instead, and `change
+nb_core 1` runs on one thread. From Python, `Driver.set_threads()`/
+`set_min_batch()` (or the `WareHouse` ones, and `WareHouse.batch_size_for()`
+for the resulting batch size); PDF reweighting (`PdfDriver`) runs as one
+batch.
+
+A batch split off a subprocess evaluates its own instance of the matrix
+element library. MadMatrix libraries keep the model parameters in one
+singleton per process directory (in the common library), which loading an
+instance resets to the param_card: MadtRex therefore loads them one at a
+time, while no matrix element is evaluated, and sets the parameters of the
+current launch on them again.
+
+## Helicity reweighting
+
+By default MadtRex reweights each event at its own helicity: the matrix
+elements (normaliser and reweighted alike) are evaluated for the helicity
+combination in the event's LHE `SPINUP` column, through MadMatrix's
+per-event helicity input (`UMAMI_IN_HELICITY_INDEX`), rather than summed
+over helicities. Set `change helicity False` (the classic reweighting
+module's option, with its default True) for helicity-summed matrix elements
+instead:
+
+- in a reweight card (in its LO part, before the first `launch`):
+  ```
+  change helicity False
+  ```
+  (`set helicity_reweighting false` and `set use_helicity false` are its
+  deprecated former spellings). It has no meaning for PDF reweighting, and a
+  PDF section that sets it is an error;
+- on the command line of either driver (and `bin/reweight` in a gridpack):
+  `--helicity-reweighting` / `--no-helicity-reweighting`, which override the
+  card;
+- from Python: `run_parameter_reweighting(..., helicity_reweighting=False)`,
+  `Driver.set_helicity_reweighting(False)` or
+  `TupperWare.set_helicity_reweighting(False)`.
+
+Either way an unchanged parameter point reweights every event by exactly 1.
+
+How an event gets its helicity index: the library enumerates its helicity
+combinations in the order of the subprocess's helicity table in
+`subprocesses.json` (the table MG7 also writes `SPINUP` from; UMAMI exposes
+the library's helicity count, not the table itself). MadtRex matches each
+event to a flavor channel of its subprocess by the positions of its legs
+(incoming, then outgoing, in LHE order), reads the legs' `SPINUP` values in
+that order and looks the combination up in the table. MG7 writes the events
+of a mirrored flavor channel with their two incoming legs swapped (LHE puts
+the +z beam first); those are put back into the library's leg order, the
+momenta passed to the matrix element as well as the helicities.
+
+Helicity reweighting needs all of this to work:
+
+- a **matrix element library** that accepts per-event helicities
+  (`UMAMI_IN_HELICITY_INDEX`), and a **helicity map** for its subprocess: a
+  helicity table in `subprocesses.json` with one helicity per leg and as many
+  combinations as the library reports (`UMAMI_META_HELICITY_COUNT`). If
+  either is missing, that subprocess is reweighted with helicity summation,
+  with a one-time warning on stderr;
+- a **sample** with helicity information: a `SPINUP` of -1, 0 or +1 for
+  every particle of every event, in a combination found in the table. An
+  event with any other value (9, "unknown/summed", as MG7 writes for an
+  interference sample generated with helicity-summed matrix elements), or
+  with a combination not in the table, is an **error** naming the matrix
+  element and the event; reweight such a sample with
+  `change helicity False`.
+
+`madtrex/tools/check_helicity_map.py PROC_DIR EVENTS.lhe` checks the map for
+a sample against the matrix element libraries directly (through UMAMI,
+independently of MadtRex): the per-helicity |M|^2 sum to the summed one, no
+event's own helicity has |M|^2 = 0, and the events' helicities are
+distributed as the |M|^2 says they should be.
+
+## Gridpacks
+
+With `[gridpack] include_madtrex = true` (the default) in `run_card.toml`, an
+installed madtrex is copied into the gridpack as `madtrex/install/` (just
+`madtrex`, `pyrex` and the Rex libraries), next to `madspace/install/`, and
+the gridpack's `bin/generate_events` puts it on `sys.path`. If madtrex is not
+installed it is skipped with a message. With madtrex included, the gridpack
+also gets the standalone driver as `bin/reweight` (plus the reweight-card
+modules and `SubProcesses/subprocesses.json`), so the events it generates can
+be reweighted without MadGraph:
+
+```bash
+bin/generate_events --events 10000
+bin/reweight --rwgt-card my_reweight_card.dat
+```
+
+Without `include_source`, only the precision the libraries were built with
+is available, and `change model` cards need MadGraph.
+
+`include_madtrex_source = true` copies the madtrex sources (including `Rex/`)
+as well, and also gives the gridpack `bin/reweight`. On its first use,
+`bin/reweight` builds madtrex from those sources (`madtrex/install.py`) when
+the gridpack has no madtrex install, and rebuilds it when the install was
+compiled against another madspace than the gridpack's (madtrex records the
+madspace headers it was built against and refuses any other madspace). This
+needs a madspace with its headers, i.e. built from source in the gridpack:
+use it with `include_madspace = false` and `include_madspace_source = true`
+when the gridpack runs on another machine. The madtrex module finds
+`libmadspace` relative to itself in the gridpack layout, so the gridpack can
+be moved.
+
+Outside a gridpack, `import madtrex` likewise refuses a madspace other than
+the one it was built against (e.g. after `install madspace`), with the
+command to rebuild it, and MadGraph's automatic backend choice then uses the
+classic module.
+
+## Building by hand
+
+When building by hand, CMake finds madspace by importing it with the build's
+Python; pass `-DMADSPACE_DIR=/path/to/madspace` (the package directory, which
+contains `include/` and `lib/`) to override.
diff --git a/madtrex/Rex/.gitignore b/madtrex/Rex/.gitignore
new file mode 100644
index 0000000000..20e4060016
--- /dev/null
+++ b/madtrex/Rex/.gitignore
@@ -0,0 +1,16 @@
+# makefile build output
+*.o
+*.d
+librex.so
+libtearex.so
+
+# CMake build output
+build/
+build-*/
+
+# Python packaging output
+*.egg-info/
+__pycache__/
+dist/
+wheelhouse/
+rex/_rex_py*.so
diff --git a/madtrex/Rex/CMakeLists.txt b/madtrex/Rex/CMakeLists.txt
new file mode 100644
index 0000000000..29f53f1a83
--- /dev/null
+++ b/madtrex/Rex/CMakeLists.txt
@@ -0,0 +1,77 @@
+cmake_minimum_required(VERSION 3.15...3.27)
+project(Rex VERSION 1.1.0 LANGUAGES CXX)
+
+# Building the Python bindings is strictly opt-in: everything that touches
+# Python or pybind11 lives behind this flag, so a bare `cmake -S Rex -B build`
+# never needs a Python interpreter, dev headers, or network access to fetch
+# pybind11. C++-only consumers of Rex/teaRex (including the plain `makefile`
+# build, left untouched alongside this file) are unaffected either way.
+option(REX_BUILD_PYTHON "Build the pybind11 Python bindings for Rex/teaRex" OFF)
+
+set(CMAKE_POSITION_INDEPENDENT_CODE ON)
+if(NOT CMAKE_CXX_STANDARD)
+    set(CMAKE_CXX_STANDARD 17)
+endif()
+
+add_library(rex SHARED src/Rex.cc)
+target_include_directories(
+    rex PUBLIC $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>
+               $<INSTALL_INTERFACE:include>
+)
+target_compile_features(rex PUBLIC cxx_std_17)
+
+find_package(Threads REQUIRED)
+add_library(tearex SHARED src/teaRex.cc)
+target_include_directories(
+    tearex PUBLIC $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>
+                  $<INSTALL_INTERFACE:include>
+)
+target_compile_features(tearex PUBLIC cxx_std_17)
+target_link_libraries(tearex PUBLIC rex Threads::Threads)
+# librex/libtearex always land in the same directory as each other (both
+# plain `cmake --install` and the wheel layout put them side by side), so
+# tearex only needs to look next to itself to find rex at runtime. Without
+# this, install(TARGETS ...) strips the build-tree RPATH and libtearex.so
+# can't locate librex.so post-install.
+# ($ORIGIN is the ELF spelling; macOS's dynamic loader wants @loader_path.)
+if(APPLE)
+    set(REX_RPATH_ORIGIN "@loader_path")
+else()
+    set(REX_RPATH_ORIGIN "$ORIGIN")
+endif()
+set_target_properties(tearex PROPERTIES INSTALL_RPATH "${REX_RPATH_ORIGIN}")
+
+install(TARGETS rex tearex)
+install(FILES include/Rex.h include/teaRex.h DESTINATION include)
+
+########################################################################################
+# Python bindings (optional)                                                          #
+########################################################################################
+
+if(REX_BUILD_PYTHON)
+    find_package(Python REQUIRED COMPONENTS Interpreter Development.Module)
+    include(FetchContent)
+
+    FetchContent_Declare(
+        pybind11
+        GIT_REPOSITORY https://github.com/pybind/pybind11
+        GIT_TAG        v3.0.1
+    )
+    FetchContent_MakeAvailable(pybind11)
+
+    python_add_library(
+        _rex_py
+        MODULE
+        src/python/rex.cpp
+        src/python/tearex.cpp
+        src/python/tearex.hpp
+        WITH_SOABI
+    )
+    target_link_libraries(_rex_py PRIVATE pybind11::headers)
+    target_link_libraries(_rex_py PUBLIC rex tearex)
+    # _rex_py is installed one level below librex/libtearex (pyrex/_rex_py.so
+    # vs lib/librex.so) in both the wheel layout and a plain `cmake
+    # --install`, so it needs to look up one directory then into lib/.
+    set_target_properties(_rex_py PROPERTIES INSTALL_RPATH "${REX_RPATH_ORIGIN}/../lib")
+    install(TARGETS _rex_py DESTINATION pyrex)
+endif()
diff --git a/madtrex/Rex/README.md b/madtrex/Rex/README.md
new file mode 100644
index 0000000000..9d5f59adf4
--- /dev/null
+++ b/madtrex/Rex/README.md
@@ -0,0 +1,53 @@
+# Rex and teaRex
+
+**R**apid **e**vent e**x**traction Version 1.1.0
+
+Rex is a C++ library for parsing and manipulating Les Houches Event-format (LHE) files. It is designed to fast and lightweight, and generic, allowing usage as an LHE parser and parton-level event manipulator for whatever purpose the user desires. Rex functionality is generalised through std::function-based interfaces which can be generated from customisable constructors for full generality and flexibility, while still providing a simple interface for typical use cases.
+
+**t**ensorial **e**vent **a**daption with **Rex** Version 1.1.0
+
+teaRex is an extension to the Rex library for the generic reweighting of parton-level events. It provides a flexible framework for applying weight modifications to events based on user-defined criteria, using the underlying Rex formats to sort, extract, and rewrite event-level information, and extending it to allow for generic reweighting using any information stored in an LHE file as input for a user-provided reweighting function acting on REX::process objects, which are SoA (Structure of Arrays) objects for storing event information. Users can either provide the REX::process objects themselves, or use the flexible Rex sorting architecture to extract the necessary information from an LHE file.
+
+
+
+Copyright © 2023-2025 CERN, CERN Author Zenny Wettersten.
+Copyright © 2025-2026 Zenny Wettersten.
+
+Dual licensed: under the GNU Lesser General Public License (version 3 or
+later), as upstream Rex (https://github.com/zeniheisser/Rex), or, as part of
+MadGraph7, under the MadGraph7 license (the LICENSE file of the MadGraph7
+distribution).
+
+All rights not expressly granted are reserved.
+
+
+----
+
+
+## Installation guide
+
+Rex and teaRex are made up of two files each --- an implementation provided in the /src/ directory, and a header provided in the /include/ directory. Any program using Rex needs to include the Rex.h file (i.e. ```#include "Rex.h"```), and at compilation needs to be linked against the compiled library by adding its path to the compile-time ```LDFLAGS```. teaRex is used identically, although note that Rex.h is included in teaRex.h, making it unnecessary to include both (although both compiled libraries still need to be linked against).
+
+A makefile supporting compilation as well as installation (both local and global) are provided with this repository, as the file ```makefile``` in the main directory. Both libraries are compiled with ```-O3``` optimisation by running the ```make``` command here, and can be installed to a specific path by running the commands
+```
+make
+make install PREFIX=/PATH/TO/LIBRARY
+```
+at which point the headers and compiled shared libraries will be installed at ```/PATH/TO/LIBRARY/include``` and ```/PATH/TO/LIBRARY/lib```, respectively. This command can be used to compile Rex and teaRex system-wide as
+```
+make
+sudo make install PREFIX=/usr/local
+sudo ldconfig
+```
+
+After installation, compile your program using Rex and teaRex as
+```
+c++ ... -I /PATH/TO/LIBRARY/include -L /PATH/TO/LIBRARY/lib -lrex -ltearex
+```
+noting that when using teaRex, your binary must be compiled with multithreading support e.g. for ```g++``` by using the compilation flag ```-pthread```. Your program can now be run as
+```
+LD_LIBRARY_PATH=/PATH/TO/LIBRARY/lib:$LD_LIBRARY_PATH ./your_binary
+```
+and your binary should run using Rex without any issues.
+
+Alternatively, Rex and teaRex can be installed through CMake using the standard ```cmake``` command; using CMake, the Python bindings provided through pyrex can also be installed by setting ```DREX_BUILD_PYTHON=ON```.
\ No newline at end of file
diff --git a/madmatrix/MadtRex/Rex.h b/madtrex/Rex/include/Rex.h
similarity index 85%
rename from madmatrix/MadtRex/Rex.h
rename to madtrex/Rex/include/Rex.h
index 081317d51c..500ee314aa 100644
--- a/madmatrix/MadtRex/Rex.h
+++ b/madtrex/Rex/include/Rex.h
@@ -8,12 +8,16 @@
  *
  ***/
 //
-// *R*apid *e*vent e*x*traction Version 1.0.0
+// *R*apid *e*vent e*x*traction Version 1.1.0
 // Rex is a C++ library for parsing and manipulating Les Houches Event-format (LHE) files.
-// It is designed to fast and lightweight, in comparison to internal parsers in programs like MadGraph.
-// Currently, Rex is in development and may not contain all features necessary for full LHE parsing.
+// It is designed to fast and lightweight, and generic, allowing usage as an LHE parser
+// and parton-level event manipulator for whatever purpose the user desires.
+// Rex functionality is generalised through std::function-based interfaces
+// which can be generated from customisable constructors for full generality
+// and flexibility, while still providing a simple interface for typical use cases.
 //
-// Copyright © 2023-2026 CERN, CERN Author Zenny Wettersten.
+// Copyright © 2023-2025 CERN, CERN Author Zenny Wettersten.
+// Copyright © 2025-2026 Zenny Wettersten.
 // Licensed under the GNU Lesser General Public License (version 3 or later).
 // All rights not expressly granted are reserved.
 //
@@ -52,6 +56,8 @@
 #include <type_traits>
 #include <cassert>
 #include <variant>
+#include <atomic>
+#include <thread>
 
 // Define pi
 constexpr double pi = 3.141592653589793;
@@ -95,6 +101,41 @@ namespace REX
     // generic warning function for printing warnings without throwing
     void warning(std::string message);
 
+    // Generic verbosity-gated output helper. Objects that want an opt-out
+    // "quiet mode" hold one of these (typically as a member) and route their
+    // diagnostic output through it
+    class verbosity
+    {
+    public:
+        verbosity() : verbose_(std::make_shared<std::atomic<bool>>(true)) {}
+        verbosity(bool verbose) : verbose_(std::make_shared<std::atomic<bool>>(verbose)) {}
+        verbosity(const verbosity &) = default;
+        verbosity(verbosity &&) = default;
+        verbosity &operator=(const verbosity &) = default;
+        verbosity &operator=(verbosity &&) = default;
+
+        bool is_verbose() const noexcept { return this->verbose_->load(std::memory_order_relaxed); }
+        void set_verbose(bool v) noexcept { this->verbose_->store(v, std::memory_order_relaxed); }
+        operator bool() const noexcept { return this->is_verbose(); }
+        verbosity &operator=(bool v) noexcept
+        {
+            this->set_verbose(v);
+            return *this;
+        }
+
+        // Shares this object's underlying flag with another verbosity, so
+        // toggling either one toggles both
+        void link(const verbosity &parent) { this->verbose_ = parent.verbose_; }
+        // Detaches from whatever's currently shared, keeping the current value
+        void unlink() { this->verbose_ = std::make_shared<std::atomic<bool>>(this->is_verbose()); }
+
+        void print(const std::string &msg, bool force = false) const;
+        void warn(const std::string &msg, bool force = false) const;
+
+    private:
+        std::shared_ptr<std::atomic<bool>> verbose_;
+    };
+
     // free functions / callables that take (ostream&, args...), map to string
     template <class F, class... Args,
               std::enable_if_t<!std::is_member_pointer_v<std::decay_t<F>>, int> = 0>
@@ -1120,6 +1161,11 @@ namespace REX
         double m2() const;    // mass squared
     };
 
+    struct event; // forward declaration for hash function type
+
+    using event_hash_fn = std::function<size_t(event &)>;        // hash function type
+    using cevent_hash_fn = std::function<size_t(const event &)>; // hash function type
+
     struct event
     {
     public:
@@ -1149,6 +1195,13 @@ namespace REX
         vecArr2<short int> mother_ = {}, icol_ = {};            // mother and color indices
         std::vector<double> wgts_ = {};                         // additional weights, if any; note that wgt ids are not stored at the event level, so custom writers need to handle this at the LHEF level
 
+        // Not part of the LHEF standard: event-level indices identifying which
+        // helicity/flavor (channel) configuration this event was generated with
+        // or should be evaluated at. Left at -1 (unset) unless explicitly assigned,
+        // eg by a generator/reweighter that fixes helicities or flavor channels
+        int helicity_ = -1;
+        int flavor_ = -1;
+
         // Self-returning setters
         event &set_n(size_t n);
         event &set_proc_id(long int id);
@@ -1172,6 +1225,10 @@ namespace REX
         event &set_icol(const std::vector<std::array<short int, 2>> &c);
         event &set_wgts(const std::vector<double> &w);
         event &add_wgt(double w, const std::string &id = "");
+        event &set_helicity(int h);
+        event &set_helicity(event_hash_fn hash_fn);
+        event &set_flavor(int f);
+        event &set_flavor(event_hash_fn hash_fn);
 
         std::vector<size_t> indices = {};                                  // indices of particles for ordered views without modifying underlying data
         event &set_indices();                                              // Default indexing is sequential by storage order
@@ -1272,6 +1329,10 @@ namespace REX
         std::vector<double> &wgts();
         const std::vector<double> &wgts() const;
         size_t n_wgts() const;
+        int &helicity();
+        const int &helicity() const;
+        int &flavor();
+        const int &flavor() const;
 
         // IDs for various additional weights, shared between events (and the LHE struct)
         std::shared_ptr<std::vector<std::string>> weight_ids = nullptr;
@@ -1606,7 +1667,7 @@ namespace REX
             }
             if (it->second.type() != typeid(T))
             {
-                throw std::bad_any_cast("event::get: Parameter '" + name + "' is not of requested type");
+                throw std::runtime_error("event::get: Parameter '" + name + "' is not of requested type");
             }
             return std::any_cast<const T &>(it->second);
         }
@@ -1741,7 +1802,7 @@ namespace REX
     // sort particle-level fields using exclusively the fields specified in the configuration.
     // Additionally, the status_filter variable allows for defining which particle statuses to extract for comparison.
     // For doubles, relative tolerances can be set independently for each field by the user.
-    // However, for integers, only exact equality is supported.
+    // However, for integers, only exact equality is supported
     struct eventComparatorConfig
     {
         // Status filter: only compare particles with one of these statuses
@@ -2161,6 +2222,8 @@ namespace REX
         std::vector<std::shared_ptr<event>> events = {};
         event_equal_fn comparator = external_legs_comparator;
         cevent_equal_fn const_comparator = external_legs_const_comparator;
+        event_hash_fn helicity_index = nullptr;
+        event_hash_fn flavor_index = nullptr;
         // Default constructors
         eventBelongs() = default;
         eventBelongs(const eventBelongs &) = default;
@@ -2220,8 +2283,7 @@ namespace REX
 
     eventBelongs all_events_belong();
 
-    using event_hash_fn = std::function<size_t(event &)>;        // hash function type
-    using cevent_hash_fn = std::function<size_t(const event &)>; // hash function type
+
 
     // Struct to sort events by their belonging to sets of events
     struct eventSorter
@@ -2279,9 +2341,76 @@ namespace REX
         cevent_hash_fn get_const_hash() const;
     };
 
-    eventSorter make_sample_sorter(std::vector<event> sample, event_equal_fn comp = external_legs_comparator);
+    eventSorter make_sample_sorter(const std::vector<event> &sample, event_equal_fn comp = external_legs_comparator);
     eventSorter make_sample_sorter(std::vector<std::shared_ptr<event>> sample, event_equal_fn comp = external_legs_comparator);
 
+    struct initNode; // forward declaration -- histogram::normalise() only needs the type by reference
+
+    double weight_norm(const initNode &init, const std::vector<std::shared_ptr<event>> &evts, verbosity *verbose = nullptr);
+
+        struct histogram
+    {
+        struct entry
+        {
+            std::shared_ptr<event> evt = nullptr;
+            size_t index = npos;
+            double base_weight = 1.0;
+        };
+
+        std::vector<entry> entries = {};
+        std::vector<size_t> nominal = {};
+        std::vector<std::vector<double>> mult = {};
+        size_t n_wgts_synced = 0;
+        double norm = 1.0;
+        size_t n_bins = npos;
+        verbosity verbose;
+
+        // Default constructors
+        histogram() = default;
+        histogram(const histogram &) = default;
+        histogram(histogram &&) noexcept = default;
+        histogram &operator=(const histogram &) = default;
+        histogram &operator=(histogram &&) noexcept = default;
+
+        // Construct from events + explicit bin indices
+        histogram(const std::vector<std::shared_ptr<event>> &evts, const std::vector<size_t> &indices);
+
+        // Construct from events + an eventSorter/event_hash_fn classifying them into bins
+        histogram(const std::vector<std::shared_ptr<event>> &evts, eventSorter sorter);
+        histogram(const std::vector<std::shared_ptr<event>> &evts, event_hash_fn hash);
+        histogram(const std::vector<std::shared_ptr<event>> &evts, std::vector<eventBelongs> e_sets);
+        histogram(const std::vector<std::shared_ptr<event>> &evts, std::vector<event_bool_fn> comps);
+
+        histogram &add_event(std::shared_ptr<event> e, size_t index);
+        histogram &add_event(std::shared_ptr<event> e, eventSorter sorter);
+        histogram &add_event(std::shared_ptr<event> e, event_hash_fn hash);
+
+        histogram &add_events(const std::vector<std::shared_ptr<event>> &evts, const std::vector<size_t> &indices);
+        histogram &add_events(const std::vector<std::shared_ptr<event>> &evts, eventSorter sorter);
+        histogram &add_events(const std::vector<std::shared_ptr<event>> &evts, event_hash_fn hash);
+        histogram &add_events(const std::vector<std::shared_ptr<event>> &evts, const eventBelongs &e_set);
+        histogram &add_events(const std::vector<std::shared_ptr<event>> &evts, event_bool_fn comp);
+        histogram &add_events(const std::vector<std::shared_ptr<event>> &evts, cevent_bool_fn comp);
+        histogram &add_events(const std::vector<std::shared_ptr<event>> &evts, std::vector<eventBelongs> e_sets);
+        histogram &add_events(const std::vector<std::shared_ptr<event>> &evts, std::vector<event_bool_fn> comps);
+
+        void rebuild();
+        histogram &normalise(const initNode &init);
+        histogram &set_norm(double norm);
+
+        bool sync_wgts();
+
+        size_t size() const;                               // number of events held (entries.size())
+        size_t bins() const;                               // number of bins (nominal.size())
+        size_t rows() const;                               // number of weight sets currently held (mult.size(), >= 1)
+        const std::vector<size_t> &counts() const;         // alias for nominal
+        std::vector<double> heights(size_t row) const;     // nominal[b]*mult[row][b] per bin -- weight set `row`'s bin heights, on row 0's scale (see mult above)
+        std::vector<std::vector<double>> heights() const;  // heights(row) for every row -- the full, immediately-plottable 2D array
+
+    private:
+        void place(std::shared_ptr<event> e, size_t index);
+    };
+
     struct process
     {
         // Default constructors
@@ -2292,6 +2421,8 @@ namespace REX
         process &operator=(process &&) noexcept = default;
         explicit process(std::vector<std::shared_ptr<event>> evts, bool filter_partons = false);
         explicit process(std::vector<event> evts, bool filter_partons = false);
+        explicit process(std::vector<std::shared_ptr<event>> evts, bool filter_partons = false, bool column_major = false);
+        explicit process(std::vector<event> evts, bool filter_partons = false, bool column_major = false);
 
         process &add_event(const event &e);
         process &add_event(std::shared_ptr<event> e);
@@ -2299,9 +2430,20 @@ namespace REX
         process &add_event_raw(std::shared_ptr<event> e);
         process &add_event_filtered(const event &e);
         process &add_event_filtered(std::shared_ptr<event> e);
+        process &add_event_umami_filtered(const std::vector<event> &evts);
+        process &add_event_umami_filtered(std::vector<std::shared_ptr<event>> evts);
         process &add_event(const std::vector<event> &evts);
         process &add_event(std::vector<std::shared_ptr<event>> evts);
 
+        // Transpose parton-level data between the row-major LHEF layout
+        // (momenta_ etc., stored as [event][parton][E,px,py,pz]) and the
+        // column-major UMAMI layout (umami_momenta_ etc., stored as
+        // [E,px,py,pz][parton][event]). Requires every event to have the
+        // same number of partons. Updates column_major to reflect the
+        // resulting layout
+        process &to_umami();
+        process &from_umami();
+
         // LHEF data, vectorised
         // Note that vecors of vectors are not contiguous in memory
         std::vector<size_t> n_ = {};         // number of partons per event
@@ -2322,6 +2464,42 @@ namespace REX
         std::vector<std::vector<double>> wgts_ = {}; // additional weights, if any; note that wgt ids are not stored at the event level, so custom
         std::unordered_map<std::string, std::vector<std::any>> extra;
 
+        // Not part of the LHEF standard: event-level indices identifying which
+        // helicity/flavor (channel) configuration each event has
+        std::vector<int> helicity_ = {};
+        std::vector<int> flavor_ = {};
+
+        // Alternate representation of particle-level data
+        // matching the UMAMI column-major format to simplify
+        // interfacing with UMAMI scattering amplitudes;
+        // if  "column_major" is true, data is stored here in
+        // column-major order, otherwise it is stored in the
+        // standard LHEF row-major order in the above vectors.
+        // Compare: Above, momenta_ is stored as [event][parton][E,px,py,pz],
+        // while below it is stored as [E,px,py,pz][parton][event]
+        bool column_major = false;
+        std::vector<double> umami_momenta_ = {}; // column-major momenta, if column_major is true
+        std::vector<double> umami_mass_ = {}, umami_vtim_ = {}, umami_spin_ = {};
+        std::vector<long int> umami_pdg_ = {};
+        std::vector<short int> umami_status_ = {}, umami_mother_ = {}, umami_icol_ = {};
+
+        // Generic, opaque handle to an externally-owned, process-wide buffer.
+        // Rex never allocates, interprets, or frees the pointee: this exists
+        // purely so that off-CPU (eg GPU-resident) functions have a standard
+        // place to stash and retrieve raw pointers/handles (device buffers,
+        // streams, contexts, ...) alongside the SoA process data
+        struct devptr
+        {
+            void *ptr = nullptr;   // opaque pointer, meaning defined by the owner
+            std::size_t stride = 0; // caller-defined stride between per-event slots
+            std::any tag;          // opaque context the owner can use to interpret/manage lifetime
+        };
+        std::unordered_map<std::string, devptr> device_data = {};
+        process &set_device_data(const std::string &name, void *ptr, std::size_t stride = 0, std::any tag = {});
+        devptr &get_device_data(const std::string &name);
+        const devptr &get_device_data(const std::string &name) const;
+        bool has_device_data(const std::string &name) const;
+
         bool filter = false; // whether to extract data using raw data or event_view
 
         // Accessors
@@ -2376,36 +2554,50 @@ namespace REX
         const vecArr4<double> &momentum() const { return momenta_; }
         vecArr4<double> &momenta() { return momenta_; }
         const vecArr4<double> &momenta() const { return momenta_; }
+        std::vector<double> &umami_momenta() { return umami_momenta_; }
+        const std::vector<double> &umami_momenta() const { return umami_momenta_; }
         std::vector<double> &mUP() { return mass_; }
         const std::vector<double> &mUP() const { return mass_; }
         std::vector<double> &m() { return mass_; }
         const std::vector<double> &m() const { return mass_; }
         std::vector<double> &mass() { return mass_; }
         const std::vector<double> &mass() const { return mass_; }
+        std::vector<double> &umami_mass() { return umami_mass_; }
+        const std::vector<double> &umami_mass() const { return umami_mass_; }
         std::vector<double> &vtim() { return vtim_; }
         const std::vector<double> &vtim() const { return vtim_; }
+        std::vector<double> &umami_vtim() { return umami_vtim_; }
+        const std::vector<double> &umami_vtim() const { return umami_vtim_; }
         std::vector<double> &vTimUP() { return vtim_; }
         const std::vector<double> &vTimUP() const { return vtim_; }
         std::vector<double> &vTim() { return vtim_; }
         const std::vector<double> &vTim() const { return vtim_; }
         std::vector<double> &spin() { return spin_; }
         const std::vector<double> &spin() const { return spin_; }
+        std::vector<double> &umami_spin() { return umami_spin_; }
+        const std::vector<double> &umami_spin() const { return umami_spin_; }
         std::vector<double> &spinUP() { return spin_; }
         const std::vector<double> &spinUP() const { return spin_; }
         std::vector<long int> &idUP() { return pdg_; }
         const std::vector<long int> &idUP() const { return pdg_; }
         std::vector<long int> &pdg() { return pdg_; }
         const std::vector<long int> &pdg() const { return pdg_; }
+        std::vector<long int> &umami_pdg() { return umami_pdg_; }
+        const std::vector<long int> &umami_pdg() const { return umami_pdg_; }
         std::vector<long int> &id() { return pdg_; }
         const std::vector<long int> &id() const { return pdg_; }
         std::vector<short int> &status() { return status_; }
         const std::vector<short int> &status() const { return status_; }
+        std::vector<short int> &umami_status() { return umami_status_; }
+        const std::vector<short int> &umami_status() const { return umami_status_; }
         std::vector<short int> &iSt() { return status_; }
         const std::vector<short int> &iSt() const { return status_; }
         std::vector<short int> &iStUP() { return status_; }
         const std::vector<short int> &iStUP() const { return status_; }
         vecArr2<short int> &mother() { return mother_; }
         const vecArr2<short int> &mother() const { return mother_; }
+        std::vector<short int> &umami_mother() { return umami_mother_; }
+        const std::vector<short int> &umami_mother() const { return umami_mother_; }
         vecArr2<short int> &moth() { return mother_; }
         const vecArr2<short int> &moth() const { return mother_; }
         vecArr2<short int> &mothUP() { return mother_; }
@@ -2416,12 +2608,21 @@ namespace REX
         const vecArr2<short int> &iCol() const { return icol_; }
         vecArr2<short int> &icol() { return icol_; }
         const vecArr2<short int> &icol() const { return icol_; }
+        std::vector<short int> &umami_icol() { return umami_icol_; }
+        const std::vector<short int> &umami_icol() const { return umami_icol_; }
 
         std::vector<std::vector<double>> &wgtUP() { return wgts_; }
         const std::vector<std::vector<double>> &wgtUP() const { return wgts_; }
         std::vector<std::vector<double>> &wgts() { return wgts_; }
         const std::vector<std::vector<double>> &wgts() const { return wgts_; }
 
+        std::vector<int> &helicity() { return helicity_; }
+        const std::vector<int> &helicity() const { return helicity_; }
+        std::vector<int> &flavor() { return flavor_; }
+        const std::vector<int> &flavor() const { return flavor_; }
+
+        std::size_t size() const { return n_.size(); }
+
         std::vector<double> &get_muF()
         {
             // if muF empty, set it equal to scale before returning
@@ -2496,6 +2697,8 @@ namespace REX
         process &set_icol(const vecArr2<short int> &icol);
         process &set_wgts(const std::vector<std::vector<double>> &wgts);
         process &append_wgts(const std::vector<double> &wgts);
+        process &set_helicity(const std::vector<int> &helicity);
+        process &set_flavor(const std::vector<int> &flavor);
         process &add_extra(const std::string &name, const std::any &value);
         process &add_extra(const std::unordered_map<std::string, std::any> &values);
         process &add_extra(const std::string &name, const std::vector<std::any> &values);
@@ -2556,6 +2759,8 @@ namespace REX
         process &transpose_iColUP();
         process &transpose_wgts();
         process &transpose_extra();
+        process &transpose_helicity();
+        process &transpose_flavor();
 
         // Specific momentum component transpositions
         process &transpose_E();
@@ -2705,6 +2910,63 @@ namespace REX
         }
     };
 
+    struct eventSource
+    {
+        std::function<initNode()> read_init = nullptr;   // called once, eagerly, before any pull()
+        std::function<std::any()> read_header = nullptr; // optional
+        // Pulls up to n events, advancing the source; returns fewer (or an
+        // empty vector) once exhausted. Not required to be safe to call
+        // again after returning fewer than n events
+        std::function<std::vector<std::shared_ptr<event>>(size_t)> pull = nullptr;
+        // Present only if the source supports rewinding to the start of the
+        // event stream (init/header are not re-read) for a second pass.
+        // Left null for forward-only sources
+        std::function<void()> rewind = nullptr;
+        bool seekable() const noexcept { return static_cast<bool>(rewind); }
+    };
+
+    // Opens filename for streaming, reusing load_lhef()'s header/init scan
+    // and per-event xml_to_event() parse, one <event> block at a time
+    eventSource open_lhe_source(const std::string &filename);
+    eventSource open_lhe_source(std::istream &in); // caller retains ownership; must outlive the source
+
+    struct eventSink
+    {
+        // Opening tag + header + init, written exactly once before any
+        // write_event() call
+        std::function<void(const initNode &, const std::any &)> write_open = nullptr;
+        std::function<void(std::shared_ptr<event>)> write_event = nullptr;
+        std::function<void()> write_close = nullptr; // closing tag, exactly once
+    };
+
+    eventSink open_lhe_sink(const std::string &filename, bool include_ids = false);
+
+    // Splits every event pulled from src across n_shards independent
+    // destinations, each obtained once, up front, via make_sink(shard_index)
+    // -- so a caller wanting eg per-shard binary EventFile output can defer
+    // even opening the underlying file until it actually receives its first
+    // event, if it wants shard file "shape" (eg a fixed particle count) to
+    // fall out of whatever landed there rather than being fixed in advance.
+    // src's init (and header, if it provides one) is written to every shard
+    // via its own write_open(), so each resulting shard is independently a
+    // complete, valid event source in its own right -- not just a fragment
+    // that only makes sense read back alongside its siblings. Events are
+    // distributed round-robin, in pulled order (shard 0, 1, ..., n_shards-1,
+    // 0, 1, ...), by streaming src in batch_size chunks via its own pull(),
+    // so memory use stays O(batch_size) regardless of the source's total
+    // size. Every shard's write_close() is called exactly once, after src is
+    // exhausted (or immediately, with zero events written, if src already
+    // was). Returns the number of events written to each shard, in shard
+    // order. Throws if n_shards == 0, if src is missing read_init/pull, or
+    // if any eventSink returned by make_sink is missing write_open/
+    // write_event/write_close
+    std::vector<size_t> split_source(
+        eventSource &src,
+        size_t n_shards,
+        const std::function<eventSink(size_t)> &make_sink,
+        size_t batch_size = 10000
+    );
+
     // Data-driven LHE struct
     // Contains both the object oriented event representation
     // and the data-oriented SoA process representation,
@@ -2731,6 +2993,23 @@ namespace REX
         std::any header;
 
         bool filter_processes = true;
+        bool column_major = false; // whether to store the process data in column-major order (for UMAMI compatibility)
+
+        // Maximum number of events per REX::process built by events_to_processes().
+        // Left at npos (unset) by default. If set, sort_events() stops filling a
+        // sorter class's bucket once it reaches max_batch events and spills any
+        // further events of that same class into a freshly appended bucket instead
+        // NOTE: BATCH_SIZE HAS TO DO WITH STREAMING, MAX_BATCH HAS TO DO WITH PROCESS SPLITTING,
+        // THEY HAVE NO RELATION TO EACH OTHER
+        std::size_t max_batch = npos;
+
+        // Threads sort_events() classifies the events with (each a chunk of at
+        // least sort_chunk events); 1, the default, classifies them in order on
+        // the calling thread. The sorter's classifiers must then be safe to
+        // call concurrently on different events, as eventBelongs' are (they
+        // only write to the event they classify).
+        std::size_t sort_threads = 1;
+        static constexpr std::size_t sort_chunk = 4096;
 
         eventSorter sorter;
         event_hash_fn event_hash = nullptr; // hash function for events (ie sorter)
@@ -2740,11 +3019,42 @@ namespace REX
         std::vector<std::shared_ptr<event>> events = {};                // vector of events
         std::vector<std::vector<std::shared_ptr<event>>> sorted_events; // vector of vectors of events, sorted according to the processes scheme
         std::vector<std::shared_ptr<process>> processes = {};           // vector of processes
+        std::vector<size_t> process_group = {};
 
         std::shared_ptr<std::vector<std::string>> weight_ids = std::make_shared<std::vector<std::string>>(); // weight ids for the <rwgt> block, if any
 
         std::vector<std::string> weight_context = {}; // context strings for each weight, if any
 
+        // Streaming mode: when source is set, events/sorted_events/processes
+        // hold only the current batch, refreshed by pull_batch(). All other
+        // lhe machinery (sort_events, events_to_processes, transpose, ...)
+        // is unaware of and unaffected by streaming -- it always just
+        // operates on whatever is currently in `events`, batch or not.
+        // NOTE: BATCH_SIZE HAS TO DO WITH STREAMING, MAX_BATCH HAS TO DO WITH PROCESS SPLITTING,
+        // THEY HAVE NO RELATION TO EACH OTHER
+        std::shared_ptr<eventSource> source = nullptr;
+        std::size_t batch_size = npos;         // events requested per pull_batch(); npos is invalid for streaming use
+        std::size_t events_pulled_total = 0;   // running count across all batches pulled so far
+        bool streaming() const noexcept { return static_cast<bool>(source); }
+
+        // Builds a streaming lhe: reads init (and header, if the source
+        // provides one) eagerly, but pulls no events yet -- call
+        // pull_batch() to fetch the first batch
+        static lhe open_streaming(std::shared_ptr<eventSource> src, std::size_t batch_size = 10000);
+
+        // Discards whatever events/sorted_events/processes/process_group the
+        // lhe currently holds and replaces them with up to batch_size fresh
+        // events pulled from source. Returns the number of events pulled
+        // (0 means the source is exhausted). Throws if not in streaming mode
+        std::size_t pull_batch();
+
+        // Named histograms owned by this lhe (composition, not inheritance:
+        // an lhe *has* histograms of its events, it isn't one). Keyed by a
+        // user-chosen name (eg an observable), built against this->events
+        // via make_histogram() and re-synced to newly appended reweighted
+        // weights via sync_histogram(s)()
+        std::unordered_map<std::string, std::shared_ptr<histogram>> histograms = {};
+
         // Self-returning setting functions
         lhe &set_events(std::vector<std::shared_ptr<event>> evts);
         lhe &set_processes(std::vector<std::shared_ptr<process>> procs);
@@ -2758,6 +3068,7 @@ namespace REX
         void extract_hash();
         lhe &set_hash(event_hash_fn hash);
         lhe &set_filter(bool v);
+        lhe &set_max_batch(std::size_t n);
         lhe &set_weight_ids(const std::vector<std::string> &ids);
         lhe &set_weight_ids(std::vector<std::string> &&ids);
         lhe &set_weight_ids(std::shared_ptr<std::vector<std::string>> ids);
@@ -2778,12 +3089,62 @@ namespace REX
         void sync_weight_ids();
         void append_weight_ids(bool include = false);
 
+        // Histogram management, built against this->events. Each
+        // make_histogram() overload builds the histogram, then calls
+        // histogram::normalise(*this) on it
+        lhe &add_histogram(const std::string &name, std::shared_ptr<histogram> h);
+        lhe &add_histogram(const std::string &name, histogram h);
+        lhe &make_histogram(const std::string &name, const std::vector<size_t> &indices);
+        lhe &make_histogram(const std::string &name, eventSorter sorter);
+        lhe &make_histogram(const std::string &name, event_hash_fn hash);
+        lhe &make_histogram(const std::string &name, const eventBelongs &e_set);
+        lhe &make_histogram(const std::string &name, event_bool_fn comp);
+        lhe &make_histogram(const std::string &name, cevent_bool_fn comp);
+        lhe &make_histogram(const std::string &name, std::vector<eventBelongs> e_sets);
+        lhe &make_histogram(const std::string &name, std::vector<event_bool_fn> comps);
+        bool has_histogram(const std::string &name) const;
+        histogram &hist(const std::string &name);
+        const histogram &hist(const std::string &name) const;
+        lhe &remove_histogram(const std::string &name);
+        bool sync_histograms();
+        bool sync_histogram(const std::string &name);
+
         void print_header(std::ostream &os = std::cout) const;
         void print(std::ostream &os = std::cout, bool include_ids = false);
+
+        // Called by every lhe constructor that derives nProc_/lProc_ from
+        // the actual distinct event::proc_id_ values present in events (see
+        // eg lhe(const initNode&, vector<shared_ptr<event>>)) -- pads
+        // xSec_/xSecErr_/xMax_ up to nProc_ with 0.0 wherever they're
+        // currently shorter than that, so initNode::validate_init() (which
+        // requires all four the same length) doesn't fail the moment such
+        // an lhe is printed. Needed whenever the source initNode's own
+        // literal header declared fewer processes than events actually
+        // carry -- in particular NPRUP=0 with an otherwise-empty init block
+        // (eg MadGraph7's own events.lhe, which always leaves its <init>
+        // block a zeroed-out stub and stores real per-process xSec/xMax
+        // information nowhere in the XML at all): nProc_/lProc_ end up
+        // correctly derived from events here, but xSec_/xSecErr_/xMax_ have
+        // nothing to derive from and would otherwise stay at the stub's
+        // original (too-short) length. A padded 0.0 is a safe placeholder:
+        // downstream code already treats an all-zero xSec_ as "no real
+        // header cross section available" and estimates one from event
+        // weights instead (see reweightor::calc_xSec_from_weights()) --
+        // this only makes that fallback reachable/consistent for xMax_ too,
+        // it doesn't change what value ends up used for xSec_/xSecErr_.
+        // Never *shortens* any of the three, so a genuinely-populated init
+        // block (real per-process data already the right length) is left
+        // untouched
+        void pad_proc_arrays();
     };
 
+    // Free-function core of lhe::print_header(), factored out so eventSink
+    // implementations (eg open_lhe_sink()) can print a std::any header
+    // without needing an lhe object to hang it off of
+    void print_lhe_header(const std::any &header, std::ostream &os = std::cout);
+
     // Classes for XML handling
-    // xmlDoc: Owns the entire XML text buffer. Nodes keep this alive via shared_ptr.
+    // xmlDoc: Owns the entire XML text buffer. Nodes keep this alive via shared_ptr
     class xmlDoc
     {
     public:
@@ -2821,7 +3182,7 @@ namespace REX
         xmlNode();
         ~xmlNode();
 
-        // Parse from an owning string (keeps only one owned copy).
+        // Parse from an owning string (keeps only one owned copy)
         static std::shared_ptr<xmlNode> parse(std::string xml);
 
         // Parse from an already-shared buffer.
@@ -2869,6 +3230,20 @@ namespace REX
         bool remove_child(const xmlNode *child) noexcept;
         bool remove_child(std::string_view name) noexcept;
 
+        // Actually erases a child previously added via add_child()/
+        // insert_child_*() (as opposed to remove_child()'s suppressed_
+        // flag, which only ever affects a genuinely in-doc child -- one
+        // write_impl() reaches via the doc-buffer-walking loop over
+        // children_, checking is_in_doc()/suppressed_ as it goes. A child
+        // added via add_child()/insert_child_*() is instead always emitted
+        // through the separate inserts_ mechanism, which has no notion of
+        // suppressed_ at all, so remove_child() is a silent no-op on it --
+        // this is the one that actually works for such a child, dropping
+        // its entry from both inserts_ and children_. Returns false (no-op)
+        // if child isn't found in inserts_ at all (eg it's a genuinely
+        // in-doc child -- use remove_child() for those instead)
+        bool remove_inserted_child(const xmlNode *child) noexcept;
+
         // Insert new children relative to existing ones
         bool insert_child_before(size_t anchor_in_doc_ordinal, std::shared_ptr<xmlNode> child) noexcept;
         bool insert_child_after(size_t anchor_in_doc_ordinal, std::shared_ptr<xmlNode> child) noexcept;
@@ -2878,7 +3253,7 @@ namespace REX
         bool replace_child(std::string_view anchor_name, std::shared_ptr<xmlNode> child, bool add_nl = false) noexcept;
 
         // Insert at specific location in this node's content
-        // (rel_offset is in bytes relative to content_start(); clamped to [0, content_len]).
+        // (rel_offset is in bytes relative to content_start(); clamped to [0, content_len])
         bool insert_child_at_content_offset(size_t rel_offset, std::shared_ptr<xmlNode> child) noexcept;
 
         // Convenience: explicit start/end of content
@@ -2893,13 +3268,18 @@ namespace REX
         // Internal constructor used by parser.
         explicit xmlNode(std::shared_ptr<const std::string> doc);
 
-        // Recursive element parser. Expects 'pos' at '<' for a normal element.
+        // Recursive element parser. Expects 'pos' at '<' for a normal element
         static std::shared_ptr<xmlNode> parse_element(const std::shared_ptr<const std::string> &doc, size_t &pos);
+        // ancestors: names of the elements enclosing this one, innermost
+        // last; returns nullptr for an element that is never closed (see the
+        // definition), so the caller can treat its start tag as text.
+        static std::shared_ptr<xmlNode> parse_element(const std::shared_ptr<const std::string> &doc, size_t &pos,
+                                                      std::vector<std::string_view> &ancestors);
 
-        // Top-level scanner to locate the first element (skips XML decl, comments, etc.).
+        // Top-level scanner to locate the first element (skips XML decl, comments, etc.)
         static size_t find_first_element_start(const std::string &s, size_t pos);
 
-        // Helpers for parsing attributes and skipping markup we don't turn into nodes.
+        // Helpers for parsing attributes and skipping markup we don't turn into nodes
         static void parse_attributes(xmlNode &node, size_t &cur);
         static bool skip_comment(const std::string &s, size_t &pos); // <!-- ... -->
         static bool skip_pi(const std::string &s, size_t &pos);      // <? ... ?>
@@ -2925,7 +3305,7 @@ namespace REX
         size_t prolog_end_ = 0;   // byte offset of first '<' of the root element
 
         // Child nodes may be “suppressed” for serialization by a parent;
-        // this flag is *only* consulted by the parent’s writer loop.
+        // this flag is *only* consulted by the parent’s writer loop
         bool suppressed_ = false;
 
         bool self_closing_ = false;
@@ -3055,7 +3435,7 @@ namespace REX
         template <class F>
         lheReader &set_event_translator(F &&f)
         {
-            // Accepts any callable returning event-like; adapt to shared_ptr<event>.
+            // Accepts any callable returning event-like; adapt to shared_ptr<event>
             event_tx_ = [fn = std::forward<F>(f)](const EventRaw &r) -> std::shared_ptr<event>
             {
                 auto out = fn(r);
@@ -3350,7 +3730,7 @@ namespace REX
     std::shared_ptr<xmlNode> load_xml(const std::string &filename);
 
     /// Minimal SLHA container: BLOCK entries with integer indices -> double values,
-    /// and DECAY widths keyed by PDG id. Comments are discarded.
+    /// and DECAY widths keyed by PDG id. Comments are discarded
     class slha
     {
     public:
@@ -3374,6 +3754,10 @@ namespace REX
         double get(const std::string &block,
                    int index,
                    double fallback = 0.0) const;
+        // indices known only at run time (multi-index blocks of a reweight card)
+        double get(const std::string &block,
+                   const std::vector<int> &indices,
+                   double fallback = 0.0) const;
 
         // Set
         void set(const std::string &block,
@@ -3382,6 +3766,9 @@ namespace REX
         void set(const std::string &block,
                  int index,
                  double value);
+        void set(const std::string &block,
+                 const std::vector<int> &indices,
+                 double value);
 
         // -------- DECAY: Get / Set --------
         // With fallback
@@ -3423,4 +3810,4 @@ namespace REX
 
 } // namespace REX
 
-#endif // DEFINING _REX_H_ FUNCTIONALITY
+#endif // DEFINING _REX_H_ FUNCTIONALITY
\ No newline at end of file
diff --git a/madmatrix/MadtRex/teaRex.h b/madtrex/Rex/include/teaRex.h
similarity index 50%
rename from madmatrix/MadtRex/teaRex.h
rename to madtrex/Rex/include/teaRex.h
index 11699ba440..f676a09340 100644
--- a/madmatrix/MadtRex/teaRex.h
+++ b/madtrex/Rex/include/teaRex.h
@@ -8,7 +8,7 @@
  *
  ***/
 //
-// *t*ensorial *e*vent *a*daption with *R*e*x* Version 1.0.0
+// *t*ensorial *e*vent *a*daption with *R*e*x* Version 1.1.0
 // teaRex is an extension to the Rex library for the generic reweighting of parton-level events.
 // It provides a flexible framework for applying weight modifications to events based on user-defined criteria,
 // using the underlying Rex formats to sort, extract, and rewrite event-level information,
@@ -17,7 +17,8 @@
 // objects for storing event information. Users can either provide the REX::process objects themselves,
 // or use the flexible Rex sorting architecture to extract the necessary information from an LHE file.
 //
-// Copyright © 2023-2026 CERN, CERN Author Zenny Wettersten.
+// Copyright © 2023-2025 CERN, CERN Author Zenny Wettersten.
+// Copyright © 2025-2026 Zenny Wettersten.
 // Licensed under the GNU Lesser General Public License (version 3 or later).
 // All rights not expressly granted are reserved.
 //
@@ -45,11 +46,34 @@ namespace REX::tea
     using lhe = REX::lhe;
     using process = REX::process;
     using slha = REX::slha;
+    using event_bool_fn = REX::event_bool_fn;
+    using verbosity = REX::verbosity;
+    using eventSource = REX::eventSource;
+    using eventSink = REX::eventSink;
     using iterator = std::function<bool()>;
     using weightor = std::function<std::shared_ptr<std::vector<double>>(process &)>;
 
+    // A puller is a zero-argument callable that materializes a previously
+    // computed weight vector on demand. deferred_weightor is a
+    // weightor-like callable that, rather than returning the weights
+    // directly, returns a puller for them -- letting an implementation
+    // whose actual result currently lives somewhere other than host memory
+    // (eg process::device_data, see Rex.h) defer materializing it until
+    // genuinely needed, instead of paying for that on every reweighting
+    // iteration
+    using weight_puller = std::function<std::shared_ptr<std::vector<double>>()>;
+    using deferred_weightor = std::function<weight_puller(process &)>;
+
     bool true_function();
 
+    // Convenience predicates for building eventBelongs/eventSorter selectors,
+    // key off of the event-level helicity_/flavor_ indices in REX::event
+    // and REX::process. A negative index matches "any" (ie unset selection),
+    // mirroring the -1 "unset" sentinel used by those fields
+    event_bool_fn helicity_selector(int helicity);
+    event_bool_fn flavor_selector(int flavor);
+    event_bool_fn helicity_flavor_selector(int helicity, int flavor);
+
     // Type for handling SLHA card modifictions using cards of the form
     // # launch rwgt_name=OPTIONAL_NAME
     // # set BLOCK_NAME PARAM_ID PARAM_VALUE
@@ -89,6 +113,8 @@ namespace REX::tea
         {
             std::string name;
             std::vector<std::pair<int, double>> params;
+            // entries with several indices ("set NMIX 1 2 0.5")
+            std::vector<std::pair<std::vector<int>, double>> multi_params = {};
         };
 
         struct rwgt_card
@@ -98,6 +124,7 @@ namespace REX::tea
             std::unordered_map<std::string, rwgt_block> blocks;
             void add_param(const std::string &block_name, std::pair<int, double> param);
             void add_param(const std::string &block_name, int param_id, double param_value);
+            void add_param(const std::string &block_name, std::vector<int> param_ids, double param_value);
         };
 
         std::vector<rwgt_card> cards = {};
@@ -166,21 +193,97 @@ namespace REX::tea
         procReweightor(std::vector<weightor> rwgts, std::shared_ptr<eventBelongs> selector);
         procReweightor(std::vector<weightor> rwgts, eventBelongs selector, weightor normaliser);
         procReweightor(std::vector<weightor> rwgts, std::shared_ptr<eventBelongs> selector, weightor normaliser);
+        // Mirror set of constructors for deferred_weightor
+        procReweightor(deferred_weightor reweight_function);
+        procReweightor(deferred_weightor reweight_function, eventBelongs selector);
+        procReweightor(deferred_weightor reweight_function, std::shared_ptr<eventBelongs> selector);
+        procReweightor(std::vector<deferred_weightor> rwgts);
+        procReweightor(std::vector<deferred_weightor> rwgts, eventBelongs selector);
+        procReweightor(std::vector<deferred_weightor> rwgts, std::shared_ptr<eventBelongs> selector);
+        procReweightor(std::vector<deferred_weightor> rwgts, eventBelongs selector, weightor normaliser);
+        procReweightor(std::vector<deferred_weightor> rwgts, std::shared_ptr<eventBelongs> selector, weightor normaliser);
+
+        // A uniform reweighting slot. Registering a plain weightor wraps it
+        // in a trivial deferred_weightor -- computed immediately, right
+        // where the weightor would have been, with a puller that just
+        // returns the already-computed value -- so eager and genuinely
+        // deferred registrations share one list, one evaluate(amp)
+        // dispatch, and one backlog/pull pipeline instead of two parallel
+        // ones. `immediate` records which case a slot came from, since
+        // that -- not its type -- is what determines whether pulling it is
+        // safe/cheap right away: used both by append_backlog()'s
+        // per-iteration flush and by initialise()'s implicit-normaliser
+        // default below
+        struct rwgt_slot
+        {
+            deferred_weightor fn = nullptr;
+            bool immediate = false;
+            explicit operator bool() const noexcept { return static_cast<bool>(fn); }
+        };
 
         std::shared_ptr<eventBelongs> event_checker;
         REX::event_bool_fn event_checker_fn = nullptr;
         weightor normaliser = nullptr;
-        std::vector<weightor> reweight_functions = {};
+        std::vector<rwgt_slot> reweight_functions = {};
         std::vector<double> normalisation = {};
         std::shared_ptr<process> proc = nullptr;
-        std::vector<std::vector<double>> backlog = {};
+        
+        verbosity verbose;
+
+        // Every evaluate(amp) call produces exactly one backlog_entry. An
+        // immediate slot's result is already fully computed by the time
+        // it's pushed here, so its puller is just a trivial wrapper around
+        // that value
+        struct backlog_entry
+        {
+            weight_puller pull;
+            // true: value already computed (came from an immediate slot),
+            // flushed every iteration by append_backlog() -- same timing
+            // the old eager-only backlog always had.
+            // false: actual result may not exist yet (eg still on device),
+            // so it accumulates here and is only pulled once, in
+            // pull_deferred_backlog() -- called from
+            // reweightor::finalise_reweighting() after the whole run
+            bool immediate;
+        };
+        std::vector<backlog_entry> backlog = {};
+
+        // Caps how many outer reweighting iterations' worth of deferred
+        // (eg GPU/device-resident) results are allowed to accumulate in
+        // backlog before append_backlog() pulls and flushes them itself --
+        // ie copies them back to the host and clears backlog -- rather
+        // than letting every iteration's deferred results pile up
+        // unflushed until the whole run (or, in streaming mode, the whole
+        // batch) finishes, as REX::npos (the default) does. Lower values
+        // trade some throughput (more, smaller device->host copies) for
+        // bounded device-side memory: make_deferred_weightor()'s
+        // n_iterations-wide device buffer only ever needs to hold
+        // max_deferred_iterations iterations' worth of results at a time,
+        // not the whole run's, once its caller actually resizes for that
+        // (see madtrex::TupperWare -- this field alone only changes when
+        // pull_deferred_backlog() gets called, not how large a buffer a
+        // given deferred_weightor was built to assume). Immediate (eager,
+        // eg cpu) results are unaffected either way: append_backlog()
+        // already flushes those every iteration, independent of this
+        size_t max_deferred_iterations = npos;
 
         procReweightor &set_event_checker(eventBelongs checker);
         procReweightor &set_event_checker(REX::event_bool_fn checker);
         procReweightor &set_normaliser(weightor normaliser);
+        procReweightor &calc_normalisation(std::vector<double> wgts);
+        procReweightor &set_initial_weights(std::vector<double> normalisation);
+        procReweightor &set_normalisation(std::vector<double> normalisation);
+        // set_reweight_functions/set_deferred_reweight_functions both
+        // replace the whole of reweight_functions.
+        // To build a slot-by-slot mix of eager and deferred
+        // functions (eg one per amp), use add_reweight_function/
+        // add_deferred_reweight_function in the desired order instead
         procReweightor &set_reweight_functions(weightor rwgt);
         procReweightor &set_reweight_functions(std::vector<weightor> rwgts);
         procReweightor &add_reweight_function(weightor rwgt);
+        procReweightor &set_deferred_reweight_functions(deferred_weightor rwgt);
+        procReweightor &set_deferred_reweight_functions(std::vector<deferred_weightor> rwgts);
+        procReweightor &add_deferred_reweight_function(deferred_weightor rwgt);
         procReweightor &set_process(std::shared_ptr<process> p);
 
         // Member functions for handling reweighting
@@ -189,7 +292,48 @@ namespace REX::tea
         void evaluate();
         void evaluate(size_t amp);
         void append_zero_weights();
+        // Pulls and appends every immediate (already-computed) entry in
+        // backlog into proc->wgts_, in accumulation order, leaving any
+        // non-immediate ones to pick up later. Called once per
+        // iteration from reweightor::run_all_iterations(); also the one
+        // place that advances/checks max_deferred_iterations, so it may
+        // itself end up calling pull_deferred_backlog() before returning
         void append_backlog();
+        // Pulls and appends every remaining entry in backlog into proc->wgts_.
+        // A no-op if backlog is empty. Called automatically by
+        // append_backlog() once max_deferred_iterations is reached; always
+        // called at least once more, unconditionally, from
+        // reweightor::finalise_reweighting() (or, in streaming mode, once
+        // per batch) after the whole iteration loop, to drain whatever's
+        // left
+        void pull_deferred_backlog();
+
+    private:
+        static rwgt_slot wrap(weightor w);
+        static rwgt_slot wrap(deferred_weightor w);
+        bool explicitly_set_norm = false;
+        // Outer iterations since backlog was last drained (by either
+        // pull_deferred_backlog() itself, or its own reset at the top of
+        // that function); see max_deferred_iterations
+        size_t deferred_iterations_since_flush = 0;
+    };
+
+    // Running-reduction accumulator standing in for the cross section/error
+    // calculations reweightor::calc_xSec_from_weights()/calc_norm()/
+    // calc_xSecs()/calc_xErrs() otherwise compute in one pass over a fully
+    // memory-resident events vector
+    struct xSecAccumulator
+    {
+        size_t n_events = 0;
+        double weight_sum = 0.0, weight_sumSq = 0.0;
+        std::vector<double> wgt_sum = {};     // per weight column: sum(ev->wgts_[k])
+        std::vector<double> ratio_sum = {};   // per weight column: sum(wgts_[k]/weight_)
+        std::vector<double> ratio_sumSq = {}; // per weight column: sum((wgts_[k]/weight_)^2)
+
+        void absorb(const std::vector<std::shared_ptr<REX::event>> &events);
+        // Sets xSec_/xSecErr_ (only if no header cross section was already
+        // present), rwgt_xSec, rwgt_xErr and norm_factor on r.
+        void finalise_into(struct reweightor &r) const;
     };
 
     // The reweightor object is an extension to REX::lhe
@@ -221,11 +365,13 @@ namespace REX::tea
 
         size_t curr_iter = 0;
         size_t n_amps = 0;
+        verbosity verbose = true;
 
         double norm_factor = 0.0;
 
         void calc_norm();
         void set_norm(double norm);
+        void calc_xSec_from_weights();
 
         std::vector<double> rwgt_xSec = {};
         std::vector<double> rwgt_xErr = {};
@@ -250,6 +396,11 @@ namespace REX::tea
         void extract_sorter();
         void initialise_reweightors();
         void finalise_reweighting();
+        // Recomputes n_amps (the max reweight_functions size across
+        // this->reweightors) from their current contents, throwing if it
+        // comes out to 0. Factored out of setup() since run_streaming() needs
+        // recompute once per batch, not just once at the start of a run
+        void sync_n_amps();
         void setup();
         void run_iteration();
         void run_all_iterations();
@@ -257,6 +408,26 @@ namespace REX::tea
 
         void calc_xSecs();
         void calc_xErrs();
+
+        // Streaming counterpart to setup()/run_all_iterations()/
+        // finalise_reweighting()/run(), for when this->streaming() (see
+        // REX::lhe::source). Repeatedly pull_batch()es from the source, runs
+        // every registered procReweightor's full iterator sequence against
+        // just that batch's processes, writes the finished batch to sink,
+        // and frees it -- so only O(batch_size) events are ever resident
+        void run_streaming(eventSink &sink, bool include_ids = false);
+
+        // Per-batch counterpart to extract_sorter(): rebuilds this->sorted_events/
+        // processes/reweightors/process_group from the current batch in
+        // this->events, matched against a caller-held, never-mutated `base`
+        // (the procReweightors registered before streaming began). Unlike
+        // extract_sorter() -- which is only safe to call once, since it
+        // rebuilds the sorter from whatever this->reweightors currently
+        // holds and may bind a base procReweightor directly into
+        // this->reweightors -- this always clones a fresh procReweightor
+        // per matched process, so `base` stays an indefinitely-reusable
+        // template across batches and this is safe to call once per batch
+        void extract_sorter_stream(const std::vector<std::shared_ptr<procReweightor>> &base);
     };
 
     struct param_rwgt : public reweightor
@@ -280,4 +451,4 @@ namespace REX::tea
 
 } // namespace REX::tea
 
-#endif // _TEAREX_H_
+#endif // _TEAREX_H_
\ No newline at end of file
diff --git a/madtrex/Rex/makefile b/madtrex/Rex/makefile
new file mode 100644
index 0000000000..92ff23f571
--- /dev/null
+++ b/madtrex/Rex/makefile
@@ -0,0 +1,44 @@
+# Tools and flags
+CXX      ?= g++
+CPPFLAGS += -Iinclude
+CXXFLAGS ?= -O3 -std=c++17 -fPIC -pthread
+LDFLAGS  ?= -shared
+
+# Install dirs
+PREFIX ?= /usr/local
+LIBDIR ?= $(PREFIX)/lib
+INCDIR ?= $(PREFIX)/include
+
+# Sources / objects
+SRCS := src/Rex.cc src/teaRex.cc
+OBJS := $(SRCS:.cc=.o)
+DEPS := $(OBJS:.o=.d)
+
+# Default target
+all: librex.so libtearex.so
+
+# Shared libs
+librex.so: src/Rex.o
+	$(CXX) $(LDFLAGS) $^ -o $@
+
+libtearex.so: src/teaRex.o
+	$(CXX) $(LDFLAGS) $^ -o $@
+
+# Compile with automatic header dependency generation
+# -MMD: write .d files, -MP: add phony targets to avoid issues on rm
+src/%.o: src/%.cc
+	$(CXX) $(CPPFLAGS) $(CXXFLAGS) -MMD -MP -c $< -o $@
+
+# Install headers + libs
+install: all
+	install -d $(DESTDIR)$(LIBDIR) $(DESTDIR)$(INCDIR)
+	install -m 0644 librex.so libtearex.so $(DESTDIR)$(LIBDIR)/
+	install -m 0644 include/Rex.h include/teaRex.h $(DESTDIR)$(INCDIR)/
+
+clean:
+	rm -f $(OBJS) $(DEPS) librex.so libtearex.so
+
+.PHONY: all install clean
+
+# Include generated dependency files so header changes trigger rebuilds
+-include $(DEPS)
diff --git a/madtrex/Rex/pyproject.toml b/madtrex/Rex/pyproject.toml
new file mode 100644
index 0000000000..aebaf80897
--- /dev/null
+++ b/madtrex/Rex/pyproject.toml
@@ -0,0 +1,23 @@
+[build-system]
+requires = ["scikit-build-core>=0.10", "packaging>=22"]
+build-backend = "scikit_build_core.build"
+
+[project]
+name = "pyrex"
+version = "1.0.0"
+description = "Python bindings for Rex/teaRex LHE parsing and event reweighting"
+readme = "README.md"
+requires-python = ">=3.9"
+# dual licensed: upstream Rex is LGPL-3.0-or-later; this copy is distributed
+# with MadGraph7 under the MadGraph7 license (LICENSE at the top of MadGraph7)
+license = { text = "LGPL-3.0-or-later, or the MadGraph7 license (LICENSE in the MadGraph7 distribution)" }
+authors = [
+  { name = "Zenny Wettersten" },
+]
+
+[tool.scikit-build]
+minimum-version = "build-system.requires"
+cmake.build-type = "Release"
+# pip install always builds the Python bindings, regardless of the CMake
+# option's own OFF default (which only matters for bare `cmake` builds).
+cmake.define.REX_BUILD_PYTHON = "ON"
diff --git a/madtrex/Rex/pyrex/__init__.py b/madtrex/Rex/pyrex/__init__.py
new file mode 100644
index 0000000000..c4d43e401f
--- /dev/null
+++ b/madtrex/Rex/pyrex/__init__.py
@@ -0,0 +1,14 @@
+import sys as _sys
+
+from . import _rex_py
+from ._rex_py import *
+
+# tea is a pybind11 submodule (m.def_submodule("tea")) living inside _rex_py,
+# not a separate file -- expose it explicitly as pyrex.tea rather than relying
+# on wildcard-import semantics to carry it over. Also register it under
+# sys.modules so `import pyrex.tea` works as a normal dotted import, not just
+# attribute access after `import pyrex` -- pybind11 doesn't do this
+# automatically since the submodule's own __name__ is "_rex_py.tea", not
+# "pyrex.tea"
+tea = _rex_py.tea
+_sys.modules[__name__ + ".tea"] = tea
diff --git a/madmatrix/MadtRex/Rex.cc b/madtrex/Rex/src/Rex.cc
similarity index 78%
rename from madmatrix/MadtRex/Rex.cc
rename to madtrex/Rex/src/Rex.cc
index 3e90686108..ea610e2265 100644
--- a/madmatrix/MadtRex/Rex.cc
+++ b/madtrex/Rex/src/Rex.cc
@@ -8,12 +8,16 @@
  *
  ***/
 //
-// *R*apid *e*vent e*x*traction Version 1.0.0
+// *R*apid *e*vent e*x*traction Version 1.1.0
 // Rex is a C++ library for parsing and manipulating Les Houches Event-format (LHE) files.
-// It is designed to fast and lightweight, in comparison to internal parsers in programs like MadGraph.
-// Currently, Rex is in development and may not contain all features necessary for full LHE parsing.
+// It is designed to fast and lightweight, and generic, allowing usage as an LHE parser
+// and parton-level event manipulator for whatever purpose the user desires.
+// Rex functionality is generalised through std::function-based interfaces
+// which can be generated from customisable constructors for full generality
+// and flexibility, while still providing a simple interface for typical use cases.
 //
-// Copyright © 2023-2026 CERN, CERN Author Zenny Wettersten.
+// Copyright © 2023-2025 CERN, CERN Author Zenny Wettersten.
+// Copyright © 2025-2026 Zenny Wettersten.
 // Licensed under the GNU Lesser General Public License (version 3 or later).
 // All rights not expressly granted are reserved.
 //
@@ -41,6 +45,20 @@ namespace REX
         std::cout << "\033[0m\n";
     }
 
+    void verbosity::print(const std::string &msg, bool force) const
+    {
+        if (!force && !this->is_verbose())
+            return;
+        std::cout << msg << std::endl;
+    }
+
+    void verbosity::warn(const std::string &msg, bool force) const
+    {
+        if (!force && !this->is_verbose())
+            return;
+        REX::warning(msg);
+    }
+
     // Explicit instantiation of templated functions and structs/classes
     template std::shared_ptr<std::vector<size_t>> ind_sort<int>(const std::vector<int> &vector, std::function<bool(const int &, const int &)> comp);
     template std::shared_ptr<std::vector<size_t>> ind_sort<double>(const std::vector<double> &vector, std::function<bool(const double &, const double &)> comp);
@@ -684,6 +702,11 @@ namespace REX
 
     size_t event::n_wgts() const { return wgts_.size(); }
 
+    int &event::helicity() { return helicity_; }
+    const int &event::helicity() const { return helicity_; }
+    int &event::flavor() { return flavor_; }
+    const int &event::flavor() const { return flavor_; }
+
     event &event::set_n(size_t n_particles)
     {
         if (n_particles != this->n_)
@@ -846,6 +869,30 @@ namespace REX
         return *this;
     }
 
+    event &event::set_helicity(int h)
+    {
+        this->helicity_ = h;
+        return *this;
+    }
+
+    event &event::set_helicity(event_hash_fn hash_fn)
+    {
+        this->helicity_ = hash_fn(*this);
+        return *this;
+    }
+
+    event &event::set_flavor(int f)
+    {
+        this->flavor_ = f;
+        return *this;
+    }
+
+    event &event::set_flavor(event_hash_fn hash_fn)
+    {
+        this->flavor_ = hash_fn(*this);
+        return *this;
+    }
+
     event &event::add_particle(const parton &p)
     {
         momenta_.push_back(p.momenta_);
@@ -2227,6 +2274,8 @@ namespace REX
             if (this->comparator(*ev, e))
             {
                 e.set_indices(*ev);
+                if(this->helicity_index) e.set_helicity(this->helicity_index);
+                if(this->flavor_index) e.set_flavor(this->flavor_index);
                 return true;
             }
         }
@@ -2663,6 +2712,350 @@ namespace REX
         return sorter;
     }
 
+    // Mirrors REX::tea::reweightor::calc_norm()'s IDWTUP-driven branches
+    // (see teaRex.cc), generalised to work from a bare initNode + event
+    // list instead of a reweightor's own accumulated processes/xSec_
+    // members. |idWgt_| == 3/4 are the "properly unweighted" LHEF
+    // conventions, where the physical per-event weight isn't event::weight()
+    // itself but is derived from the header cross section (or, for ==4,
+    // implicitly uniform); anything else is treated as a weighted sample
+    // and calibrated by matching the summed raw weight to the header xSec_
+    double weight_norm(const initNode &init, const std::vector<std::shared_ptr<event>> &evts, verbosity *verbose)
+    {
+        double accumulated = 0.0;
+        size_t n = 0;
+        for (const auto &ev : evts)
+        {
+            if (!ev)
+                continue;
+            accumulated += ev->weight_;
+            ++n;
+        }
+        if (n == 0)
+            throw std::runtime_error("weight_norm: no valid events given, cannot calculate normalisation");
+
+        double xSec = std::accumulate(init.xSec_.begin(), init.xSec_.end(), 0.0);
+        if (xSec == 0.0)
+            xSec = accumulated / double(n); // mirrors reweightor::calc_xSec_from_weights()'s fallback estimate
+
+        if (std::abs(init.idWgt_) == 3)
+            return xSec / double(n);
+        if (std::abs(init.idWgt_) == 4)
+            return 1.0 / double(n);
+
+        if (verbose && (std::abs(init.idWgt_) > 2 || init.idWgt_ == 0))
+            verbose->warn("weight_norm: idWgt is not set to a value defined in the LHE standard. Assuming weighted events.");
+        if (accumulated == 0.0)
+            throw std::runtime_error("weight_norm: total weight is zero, cannot calculate normalisation");
+        return xSec / accumulated;
+    }
+
+    void histogram::place(std::shared_ptr<event> e, size_t index)
+    {
+        double bw = e ? e->weight() : 1.0;
+        this->entries.push_back(entry{std::move(e), index, bw});
+    }
+
+    histogram::histogram(const std::vector<std::shared_ptr<event>> &evts, const std::vector<size_t> &indices)
+    {
+        if (evts.size() != indices.size())
+            throw std::invalid_argument("histogram::histogram: events and indices must be the same size");
+        this->entries.reserve(evts.size());
+        for (size_t i = 0; i < evts.size(); ++i)
+            this->place(evts[i], indices[i]);
+        this->sync_wgts();
+    }
+
+    histogram::histogram(const std::vector<std::shared_ptr<event>> &evts, eventSorter sorter)
+    {
+        this->entries.reserve(evts.size());
+        bool unmatched = false;
+        for (size_t i = 0; i < evts.size(); ++i)
+        {
+            size_t idx = sorter.position(evts[i]);
+            if (idx == npos)
+            {
+                idx = sorter.size();
+                unmatched = true;
+            }
+            this->place(evts[i], idx);
+        }
+        this->n_bins = sorter.size() + (unmatched ? 1 : 0);
+        this->sync_wgts();
+    }
+
+    histogram::histogram(const std::vector<std::shared_ptr<event>> &evts, event_hash_fn hash)
+    {
+        if (!hash)
+            throw std::invalid_argument("histogram::histogram: hash function is not callable");
+        this->entries.reserve(evts.size());
+        for (size_t i = 0; i < evts.size(); ++i)
+        {
+            size_t idx = evts[i] ? hash(*evts[i]) : npos;
+            this->place(evts[i], idx);
+        }
+        this->sync_wgts();
+    }
+
+    histogram::histogram(const std::vector<std::shared_ptr<event>> &evts, std::vector<eventBelongs> e_sets)
+        : histogram(evts, eventSorter(std::move(e_sets))) {}
+
+    histogram::histogram(const std::vector<std::shared_ptr<event>> &evts, std::vector<event_bool_fn> comps)
+        : histogram(evts, eventSorter(std::move(comps))) {}
+
+    histogram &histogram::add_event(std::shared_ptr<event> e, size_t index)
+    {
+        this->place(std::move(e), index);
+        return *this;
+    }
+
+    histogram &histogram::add_event(std::shared_ptr<event> e, eventSorter sorter)
+    {
+        size_t idx = sorter.position(e);
+        bool unmatched = (idx == npos);
+        if (unmatched)
+            idx = sorter.size();
+        this->place(std::move(e), idx);
+        size_t want = sorter.size() + (unmatched ? 1 : 0);
+        if (this->n_bins == npos || want > this->n_bins)
+            this->n_bins = want;
+        return *this;
+    }
+
+    histogram &histogram::add_event(std::shared_ptr<event> e, event_hash_fn hash)
+    {
+        if (!hash)
+            throw std::invalid_argument("histogram::add_event: hash function is not callable");
+        size_t idx = e ? hash(*e) : npos;
+        this->place(std::move(e), idx);
+        return *this;
+    }
+
+    histogram &histogram::add_events(const std::vector<std::shared_ptr<event>> &evts, const std::vector<size_t> &indices)
+    {
+        if (evts.size() != indices.size())
+            throw std::invalid_argument("histogram::add_events: events and indices must be the same size");
+        this->entries.reserve(this->entries.size() + evts.size());
+        for (size_t i = 0; i < evts.size(); ++i)
+            this->place(evts[i], indices[i]);
+        this->rebuild();
+        return *this;
+    }
+
+    histogram &histogram::add_events(const std::vector<std::shared_ptr<event>> &evts, eventSorter sorter)
+    {
+        this->entries.reserve(this->entries.size() + evts.size());
+        bool unmatched = false;
+        for (size_t i = 0; i < evts.size(); ++i)
+        {
+            size_t idx = sorter.position(evts[i]);
+            if (idx == npos)
+            {
+                idx = sorter.size();
+                unmatched = true;
+            }
+            this->place(evts[i], idx);
+        }
+        size_t want = sorter.size() + (unmatched ? 1 : 0);
+        if (this->n_bins == npos || want > this->n_bins)
+            this->n_bins = want;
+        this->rebuild();
+        return *this;
+    }
+
+    histogram &histogram::add_events(const std::vector<std::shared_ptr<event>> &evts, event_hash_fn hash)
+    {
+        if (!hash)
+            throw std::invalid_argument("histogram::add_events: hash function is not callable");
+        this->entries.reserve(this->entries.size() + evts.size());
+        for (size_t i = 0; i < evts.size(); ++i)
+        {
+            size_t idx = evts[i] ? hash(*evts[i]) : npos;
+            this->place(evts[i], idx);
+        }
+        this->rebuild();
+        return *this;
+    }
+
+    histogram &histogram::add_events(const std::vector<std::shared_ptr<event>> &evts, const eventBelongs &e_set)
+    {
+        return this->add_events(evts, eventSorter(e_set));
+    }
+
+    histogram &histogram::add_events(const std::vector<std::shared_ptr<event>> &evts, event_bool_fn comp)
+    {
+        return this->add_events(evts, eventSorter(comp));
+    }
+
+    histogram &histogram::add_events(const std::vector<std::shared_ptr<event>> &evts, cevent_bool_fn comp)
+    {
+        return this->add_events(evts, eventSorter(comp));
+    }
+
+    histogram &histogram::add_events(const std::vector<std::shared_ptr<event>> &evts, std::vector<eventBelongs> e_sets)
+    {
+        return this->add_events(evts, eventSorter(std::move(e_sets)));
+    }
+
+    histogram &histogram::add_events(const std::vector<std::shared_ptr<event>> &evts, std::vector<event_bool_fn> comps)
+    {
+        return this->add_events(evts, eventSorter(std::move(comps)));
+    }
+
+    // Recomputes nominal (per-bin unweighted counts) and mult[0] (the base
+    // sample's own deviation from that unweighted assumption) from the
+    // current entries. Already-synced mult[k>=1] rows are left untouched,
+    // only grown (never shrunk) to match a bin count that has grown since
+    // they were last computed.
+    void histogram::rebuild()
+    {
+        size_t nb = 0;
+        if (this->n_bins != npos)
+        {
+            nb = this->n_bins;
+        }
+        else
+        {
+            for (const auto &en : this->entries)
+                if (en.index != npos)
+                    nb = std::max(nb, en.index + 1);
+        }
+
+        this->nominal.assign(nb, 0);
+        std::vector<double> sum_base(nb, 0.0);
+        double total_base = 0.0;
+        size_t n_valid = 0;
+        for (const auto &en : this->entries)
+        {
+            if (en.index == npos || en.index >= nb)
+                continue;
+            ++this->nominal[en.index];
+            sum_base[en.index] += en.base_weight;
+            total_base += en.base_weight;
+            ++n_valid;
+        }
+
+        if (this->mult.empty())
+            this->mult.emplace_back();
+        this->mult[0].assign(nb, 0.0);
+        for (size_t k = 1; k < this->mult.size(); ++k)
+            if (this->mult[k].size() < nb)
+                this->mult[k].resize(nb, 0.0);
+
+        if (n_valid == 0)
+            return;
+        double w_bar = total_base / double(n_valid);
+        if (w_bar == 0.0)
+            return;
+        for (size_t b = 0; b < nb; ++b)
+            if (this->nominal[b] > 0)
+                this->mult[0][b] = (sum_base[b] / double(this->nominal[b])) / w_bar;
+    }
+
+    histogram &histogram::normalise(const initNode &init)
+    {
+        std::vector<std::shared_ptr<event>> evts;
+        evts.reserve(this->entries.size());
+        for (const auto &en : this->entries)
+            if (en.evt)
+                evts.push_back(en.evt);
+        return this->set_norm(weight_norm(init, evts, &this->verbose));
+    }
+
+    histogram &histogram::set_norm(double norm)
+    {
+        if (norm != this->norm)
+        {
+            this->norm = norm;
+            // Every already-synced row was computed with the old norm, and
+            // there's no way to retroactively rescale it without redoing
+            // the per-event division -- so just redo it: drop back to the
+            // base-sample row and re-sync from scratch.
+            this->n_wgts_synced = 0;
+            this->mult.resize(1);
+            this->sync_wgts();
+        }
+        return *this;
+    }
+
+    // Appends, per synced-past bin, one mult row for every event::wgts_
+    // column beyond n_wgts_synced. All events are assumed to share the
+    // same wgts_ length at any given time (as they do after a normal
+    // reweighting iteration appends one weight per event); the minimum
+    // length actually observed is used as the new sync point, so a
+    // straggler with fewer weights than the rest simply caps how far this
+    // call advances rather than being skipped
+    bool histogram::sync_wgts()
+    {
+        this->rebuild();
+        if (this->entries.empty())
+            return true;
+        size_t min_n = npos;
+        for (const auto &en : this->entries)
+        {
+            if (!en.evt)
+                continue;
+            min_n = std::min(min_n, en.evt->n_wgts());
+        }
+        if (min_n == npos) // no valid (non-null) events to sync from
+            return true;
+        if (min_n < this->n_wgts_synced)
+        {
+            this->verbose.warn("histogram::sync_wgts: an event now has fewer weights (" + std::to_string(min_n) +
+                               ") than were already synced into this histogram (" + std::to_string(this->n_wgts_synced) +
+                               "). Refusing to sync -- histogram is left unmodified.");
+            return false;
+        }
+        if (min_n == this->n_wgts_synced)
+            return true; // nothing new to append
+
+        size_t nb = this->bins();
+        for (size_t k = this->n_wgts_synced; k < min_n; ++k)
+        {
+            std::vector<double> sum(nb, 0.0);
+            for (const auto &en : this->entries)
+            {
+                if (!en.evt || en.index == npos || en.index >= nb)
+                    continue;
+                if (en.base_weight == 0.0)
+                    continue; // nothing to normalise this event's ratio against
+                double w_new = en.evt->wgts()[k] * this->norm;
+                sum[en.index] += w_new / en.base_weight;
+            }
+            std::vector<double> row(nb, 0.0);
+            for (size_t b = 0; b < nb; ++b)
+                if (this->nominal[b] > 0)
+                    row[b] = sum[b] / double(this->nominal[b]);
+            this->mult.push_back(std::move(row));
+        }
+        this->n_wgts_synced = min_n;
+        return true;
+    }
+
+    size_t histogram::size() const { return this->entries.size(); }
+    size_t histogram::bins() const { return this->nominal.size(); }
+    size_t histogram::rows() const { return this->mult.size(); }
+    const std::vector<size_t> &histogram::counts() const { return this->nominal; }
+
+    std::vector<double> histogram::heights(size_t row) const
+    {
+        if (row >= this->mult.size())
+            throw std::out_of_range("histogram::heights: no such weight set");
+        std::vector<double> out(this->nominal.size(), 0.0);
+        for (size_t b = 0; b < out.size(); ++b)
+            out[b] = double(this->nominal[b]) * this->mult[row][b];
+        return out;
+    }
+
+    std::vector<std::vector<double>> histogram::heights() const
+    {
+        std::vector<std::vector<double>> out;
+        out.reserve(this->mult.size());
+        for (size_t row = 0; row < this->mult.size(); ++row)
+            out.push_back(this->heights(row));
+        return out;
+    }
+
     process::process(std::vector<std::shared_ptr<event>> evs, bool filter_partons)
     {
         this->filter = filter_partons;
@@ -2675,6 +3068,20 @@ namespace REX
         this->add_event(evs);
     }
 
+    process::process(std::vector<std::shared_ptr<event>> evs, bool filter_partons, bool column_major)
+    {
+        this->filter = filter_partons;
+        this->column_major = column_major;
+        this->add_event(evs);
+    }
+
+    process::process(std::vector<event> evs, bool filter_partons, bool column_major)
+    {
+        this->filter = filter_partons;
+        this->column_major = column_major;
+        this->add_event(evs);
+    }
+
     process &process::add_event_raw(const event &ev)
     {
         auto summed_n = (this->n_summed.empty() ? 0 : this->n_summed.back()) + ev.n_;
@@ -2703,6 +3110,11 @@ namespace REX
             this->icol_.push_back(prtcl.icol_);
         }
         this->wgts_.push_back(ev.wgts_);
+        // -1 ("no fixed helicity") is kept too: skipping such events would
+        // leave helicity_ shorter than the event count for a process mixing
+        // events with and without helicities, failing validate()
+        this->helicity_.push_back(ev.helicity_);
+        if(ev.flavor_ != -1) this->flavor_.push_back(ev.flavor_);
         this->add_extra(ev.extra);
         this->events.push_back(std::make_shared<event>(ev));
         return *this;
@@ -2736,6 +3148,8 @@ namespace REX
             this->icol_.push_back(prtcl.icol_);
         }
         this->wgts_.push_back(ev->wgts_);
+        this->helicity_.push_back(ev->helicity_);
+        if(ev->flavor_ != -1) this->flavor_.push_back(ev->flavor_);
         this->add_extra(ev->extra);
         this->events.push_back(ev);
         return *this;
@@ -2770,6 +3184,8 @@ namespace REX
             this->icol_.push_back(prtcl.icol_);
         }
         this->wgts_.push_back(ev.wgts_);
+        this->helicity_.push_back(ev.helicity_);
+        if(ev.flavor_ != -1) this->flavor_.push_back(ev.flavor_);
         this->add_extra(ev.extra);
         this->events.push_back(std::make_shared<event>(ev));
         return *this;
@@ -2804,11 +3220,271 @@ namespace REX
             this->icol_.push_back(prtcl.icol_);
         }
         this->wgts_.push_back(ev->wgts_);
+        this->helicity_.push_back(ev->helicity_);
+        if(ev->flavor_ != -1) this->flavor_.push_back(ev->flavor_);
         this->add_extra(ev->extra);
         this->events.push_back(ev);
         return *this;
     }
 
+    process &process::add_event_umami_filtered(const std::vector<event> &evs)
+    {
+        for (const auto &ev : evs)
+        {
+            auto ev_view = ev.view();
+            auto summed_n = (this->n_summed.empty() ? 0 : this->n_summed.back()) + ev_view.size();
+            this->n_.push_back(ev_view.size());
+            this->n_summed.push_back(summed_n);
+            this->proc_id_.push_back(ev.proc_id_);
+            this->weight_.push_back(ev.weight_);
+            this->scale_.push_back(ev.scale_);
+            auto muF = (ev.muF_ == 0.0) ? ev.scale_ : ev.muF_;
+            auto muR = (ev.muR_ == 0.0) ? ev.scale_ : ev.muR_;
+            auto muPS = (ev.muPS_ == 0.0) ? ev.scale_ : ev.muPS_;
+            this->muF_.push_back(muF);
+            this->muR_.push_back(muR);
+            this->muPS_.push_back(muPS);
+            this->alphaEW_.push_back(ev.alphaEW_);
+            this->alphaS_.push_back(ev.alphaS_);
+            this->wgts_.push_back(ev.wgts_);
+            this->helicity_.push_back(ev.helicity_);
+            if(ev.flavor_ != -1) this->flavor_.push_back(ev.flavor_);
+            this->add_extra(ev.extra);
+            this->events.push_back(std::make_shared<event>(ev));
+        }
+        auto n_sum = this->n_summed.empty() ? 0 : this->n_summed.back();
+        auto n_prt = this->n_.empty() ? 0 : this->n_[0];
+        auto n_evt = this->n_.size();
+        if(n_sum == 0) return *this;
+        if(n_prt == 0) return *this;
+        if (n_evt == 0) return *this;
+        for (size_t i = 0; i < this->n_.size(); ++i)
+        {
+            if (this->n_[i] != n_prt)
+            {
+                throw std::runtime_error("process::add_event_umami_filtered: all events must have the same number of particles");
+            }
+        }
+        this->umami_momenta_ = std::vector<double>(n_sum * 4, 0.0);
+        this->umami_mass_ = std::vector<double>(n_sum, 0.0);
+        this->umami_vtim_ = std::vector<double>(n_sum, 0.0);
+        this->umami_spin_ = std::vector<double>(n_sum, 0);
+        this->umami_pdg_ = std::vector<long int>(n_sum, 0);
+        this->umami_status_ = std::vector<short int>(n_sum, 0);
+        this->umami_mother_ = std::vector<short int>(n_sum * 2, 0);
+        this->umami_icol_ = std::vector<short int>(n_sum * 2, 0);
+        for (size_t i = 0; i < evs.size(); ++i)
+        {
+            auto ev_view = evs[i].view();
+            for(size_t j = 0; j < ev_view.size(); ++j)
+            {
+                this->umami_momenta_[j * n_evt + i] = ev_view[j].E();
+                this->umami_momenta_[n_sum + j * n_evt + i] = ev_view[j].px();
+                this->umami_momenta_[2 * n_sum + j * n_evt + i] = ev_view[j].py();
+                this->umami_momenta_[3 * n_sum + j * n_evt + i] = ev_view[j].pz();
+                this->umami_mass_[j * n_evt + i] = ev_view[j].mass();
+                this->umami_vtim_[j * n_evt + i] = ev_view[j].vtim();
+                this->umami_spin_[j * n_evt + i] = ev_view[j].spin();
+                this->umami_pdg_[j * n_evt + i] = ev_view[j].pdg();
+                this->umami_status_[j * n_evt + i] = ev_view[j].status();
+                this->umami_mother_[j * n_evt + i] = ev_view[j].mother()[0];
+                this->umami_mother_[n_sum + j * n_evt + i] = ev_view[j].mother()[1];
+                this->umami_icol_[j * n_evt + i] = ev_view[j].icol()[0];
+                this->umami_icol_[n_sum + j * n_evt + i] = ev_view[j].icol()[1];
+            }
+        }
+        return *this;
+    }
+
+    process &process::add_event_umami_filtered(std::vector<std::shared_ptr<event>> evs)
+    {
+        for (const auto &ev : evs)
+        {
+            auto ev_view = ev->view();
+            auto summed_n = (this->n_summed.empty() ? 0 : this->n_summed.back()) + ev_view.size();
+            this->n_.push_back(ev_view.size());
+            this->n_summed.push_back(summed_n);
+            this->proc_id_.push_back(ev->proc_id_);
+            this->weight_.push_back(ev->weight_);
+            this->scale_.push_back(ev->scale_);
+            auto muF = (ev->muF_ == 0.0) ? ev->scale_ : ev->muF_;
+            auto muR = (ev->muR_ == 0.0) ? ev->scale_ : ev->muR_;
+            auto muPS = (ev->muPS_ == 0.0) ? ev->scale_ : ev->muPS_;
+            this->muF_.push_back(muF);
+            this->muR_.push_back(muR);
+            this->muPS_.push_back(muPS);
+            this->alphaEW_.push_back(ev->alphaEW_);
+            this->alphaS_.push_back(ev->alphaS_);
+            this->wgts_.push_back(ev->wgts_);
+            this->helicity_.push_back(ev->helicity_);
+            if(ev->flavor_ != -1) this->flavor_.push_back(ev->flavor_);
+            this->add_extra(ev->extra);
+            this->events.push_back(ev);
+        }
+        auto n_sum = this->n_summed.empty() ? 0 : this->n_summed.back();
+        auto n_prt = this->n_.empty() ? 0 : this->n_[0];
+        auto n_evt = this->n_.size();
+        if(n_sum == 0) return *this;
+        if(n_prt == 0) return *this;
+        if (n_evt == 0) return *this;
+        for (size_t i = 0; i < this->n_.size(); ++i)
+        {
+            if (this->n_[i] != n_prt)
+            {
+                throw std::runtime_error("process::add_event_umami_filtered: all events must have the same number of particles");
+            }
+        }
+        this->umami_momenta_ = std::vector<double>(n_sum * 4, 0.0);
+        this->umami_mass_ = std::vector<double>(n_sum, 0.0);
+        this->umami_vtim_ = std::vector<double>(n_sum, 0.0);
+        this->umami_spin_ = std::vector<double>(n_sum, 0);
+        this->umami_pdg_ = std::vector<long int>(n_sum, 0);
+        this->umami_status_ = std::vector<short int>(n_sum, 0);
+        this->umami_mother_ = std::vector<short int>(n_sum * 2, 0);
+        this->umami_icol_ = std::vector<short int>(n_sum * 2, 0);
+        for (size_t i = 0; i < evs.size(); ++i)
+        {
+            auto ev_view = evs[i]->view();
+            for(size_t j = 0; j < ev_view.size(); ++j)
+            {
+                this->umami_momenta_[j * n_evt + i] = ev_view[j].E();
+                this->umami_momenta_[n_sum + j * n_evt + i] = ev_view[j].px();
+                this->umami_momenta_[2 * n_sum + j * n_evt + i] = ev_view[j].py();
+                this->umami_momenta_[3 * n_sum + j * n_evt + i] = ev_view[j].pz();
+                this->umami_mass_[j * n_evt + i] = ev_view[j].mass();
+                this->umami_vtim_[j * n_evt + i] = ev_view[j].vtim();
+                this->umami_spin_[j * n_evt + i] = ev_view[j].spin();
+                this->umami_pdg_[j * n_evt + i] = ev_view[j].pdg();
+                this->umami_status_[j * n_evt + i] = ev_view[j].status();
+                this->umami_mother_[j * n_evt + i] = ev_view[j].mother()[0];
+                this->umami_mother_[n_sum + j * n_evt + i] = ev_view[j].mother()[1];
+                this->umami_icol_[j * n_evt + i] = ev_view[j].icol()[0];
+                this->umami_icol_[n_sum + j * n_evt + i] = ev_view[j].icol()[1];
+            }
+        }
+        return *this;
+    }
+
+    // Transpose the row-major parton data (momenta_ etc., laid out as
+    // [event][parton][E,px,py,pz]) into the column-major UMAMI layout
+    // (umami_momenta_ etc., laid out as [E,px,py,pz][parton][event])
+    process &process::to_umami()
+    {
+        this->validate();
+        if (this->column_major)
+            return *this;
+
+        size_t n_evt = this->n_.size();
+        if (n_evt == 0)
+        {
+            this->column_major = true;
+            return *this;
+        }
+
+        size_t n_prt = this->n_[0];
+        for (size_t i = 0; i < n_evt; ++i)
+        {
+            if (this->n_[i] != n_prt)
+                throw std::runtime_error("process::to_umami(): all events must have the same number of partons to convert to the UMAMI column-major layout");
+        }
+
+        size_t n_sum = this->n_summed.back();
+        this->umami_momenta_ = std::vector<double>(n_sum * 4, 0.0);
+        this->umami_mass_ = std::vector<double>(n_sum, 0.0);
+        this->umami_vtim_ = std::vector<double>(n_sum, 0.0);
+        this->umami_spin_ = std::vector<double>(n_sum, 0.0);
+        this->umami_pdg_ = std::vector<long int>(n_sum, 0);
+        this->umami_status_ = std::vector<short int>(n_sum, 0);
+        this->umami_mother_ = std::vector<short int>(n_sum * 2, 0);
+        this->umami_icol_ = std::vector<short int>(n_sum * 2, 0);
+
+        for (size_t i = 0; i < n_evt; ++i)
+        {
+            size_t begin = (i == 0) ? 0 : this->n_summed[i - 1];
+            for (size_t j = 0; j < n_prt; ++j)
+            {
+                size_t idx = begin + j;  // row-major index: [event][parton]
+                size_t k = j * n_evt + i; // column-major index: [parton][event]
+                this->umami_momenta_[k] = this->momenta_[idx][0];
+                this->umami_momenta_[n_sum + k] = this->momenta_[idx][1];
+                this->umami_momenta_[2 * n_sum + k] = this->momenta_[idx][2];
+                this->umami_momenta_[3 * n_sum + k] = this->momenta_[idx][3];
+                this->umami_mass_[k] = this->mass_[idx];
+                this->umami_vtim_[k] = this->vtim_[idx];
+                this->umami_spin_[k] = this->spin_[idx];
+                this->umami_pdg_[k] = this->pdg_[idx];
+                this->umami_status_[k] = this->status_[idx];
+                this->umami_mother_[k] = this->mother_[idx][0];
+                this->umami_mother_[n_sum + k] = this->mother_[idx][1];
+                this->umami_icol_[k] = this->icol_[idx][0];
+                this->umami_icol_[n_sum + k] = this->icol_[idx][1];
+            }
+        }
+
+        this->column_major = true;
+        return *this;
+    }
+
+    // Transpose the column-major UMAMI parton data (umami_momenta_ etc.,
+    // laid out as [E,px,py,pz][parton][event]) back into the row-major
+    // LHEF layout (momenta_ etc., laid out as [event][parton][E,px,py,pz])
+    process &process::from_umami()
+    {
+        this->validate();
+        if (!this->column_major)
+            return *this;
+
+        size_t n_evt = this->n_.size();
+        if (n_evt == 0)
+        {
+            this->column_major = false;
+            return *this;
+        }
+
+        size_t n_prt = this->n_[0];
+        for (size_t i = 0; i < n_evt; ++i)
+        {
+            if (this->n_[i] != n_prt)
+                throw std::runtime_error("process::from_umami(): all events must have the same number of partons to convert from the UMAMI column-major layout");
+        }
+
+        size_t n_sum = this->n_summed.back();
+        this->momenta_ = vecArr4<double>(n_sum);
+        this->mass_ = std::vector<double>(n_sum, 0.0);
+        this->vtim_ = std::vector<double>(n_sum, 0.0);
+        this->spin_ = std::vector<double>(n_sum, 0.0);
+        this->pdg_ = std::vector<long int>(n_sum, 0);
+        this->status_ = std::vector<short int>(n_sum, 0);
+        this->mother_ = vecArr2<short int>(n_sum);
+        this->icol_ = vecArr2<short int>(n_sum);
+
+        for (size_t i = 0; i < n_evt; ++i)
+        {
+            size_t begin = (i == 0) ? 0 : this->n_summed[i - 1];
+            for (size_t j = 0; j < n_prt; ++j)
+            {
+                size_t idx = begin + j;  // row-major index: [event][parton]
+                size_t k = j * n_evt + i; // column-major index: [parton][event]
+                this->momenta_[idx][0] = this->umami_momenta_[k];
+                this->momenta_[idx][1] = this->umami_momenta_[n_sum + k];
+                this->momenta_[idx][2] = this->umami_momenta_[2 * n_sum + k];
+                this->momenta_[idx][3] = this->umami_momenta_[3 * n_sum + k];
+                this->mass_[idx] = this->umami_mass_[k];
+                this->vtim_[idx] = this->umami_vtim_[k];
+                this->spin_[idx] = this->umami_spin_[k];
+                this->pdg_[idx] = this->umami_pdg_[k];
+                this->status_[idx] = this->umami_status_[k];
+                this->mother_[idx][0] = this->umami_mother_[k];
+                this->mother_[idx][1] = this->umami_mother_[n_sum + k];
+                this->icol_[idx][0] = this->umami_icol_[k];
+                this->icol_[idx][1] = this->umami_icol_[n_sum + k];
+            }
+        }
+
+        this->column_major = false;
+        return *this;
+    }
+
     process &process::add_event(const event &ev)
     {
         if (this->filter)
@@ -2835,6 +3511,10 @@ namespace REX
 
     process &process::add_event(const std::vector<event> &evs)
     {
+        if (this->column_major)
+        {
+            return this->add_event_umami_filtered(evs);
+        }
         for (const auto &ev : evs)
         {
             this->add_event(ev);
@@ -2844,6 +3524,10 @@ namespace REX
 
     process &process::add_event(std::vector<std::shared_ptr<event>> evs)
     {
+        if (this->column_major)
+        {
+            return this->add_event_umami_filtered(evs);
+        }
         for (auto ev : evs)
         {
             this->add_event(ev);
@@ -3123,6 +3807,45 @@ namespace REX
         return *this;
     }
 
+    process &process::set_helicity(const std::vector<int> &helicity)
+    {
+        this->helicity_ = helicity;
+        return *this;
+    }
+
+    process &process::set_flavor(const std::vector<int> &flavor)
+    {
+        this->flavor_ = flavor;
+        return *this;
+    }
+
+    process &process::set_device_data(const std::string &name, void *ptr, std::size_t stride, std::any tag)
+    {
+        this->device_data[name] = devptr{ptr, stride, std::move(tag)};
+        return *this;
+    }
+
+    process::devptr &process::get_device_data(const std::string &name)
+    {
+        auto it = this->device_data.find(name);
+        if (it == this->device_data.end())
+            throw std::out_of_range("process::get_device_data: no device data registered under name '" + name + "'");
+        return it->second;
+    }
+
+    const process::devptr &process::get_device_data(const std::string &name) const
+    {
+        auto it = this->device_data.find(name);
+        if (it == this->device_data.end())
+            throw std::out_of_range("process::get_device_data: no device data registered under name '" + name + "'");
+        return it->second;
+    }
+
+    bool process::has_device_data(const std::string &name) const
+    {
+        return this->device_data.find(name) != this->device_data.end();
+    }
+
     process &process::add_extra(const std::string &key, const std::any &value)
     {
         auto it = this->extra.find(key);
@@ -3192,6 +3915,8 @@ namespace REX
         check_event_size(this->alphaEW_, "alphaEW");
         check_event_size(this->alphaS_, "alphaS");
         check_event_size(this->wgts_, "wgts");
+        check_mu_size(this->helicity_, "helicity");
+        check_mu_size(this->flavor_, "flavor");
 
         size_t summed_n = this->n_summed.empty() ? 0 : this->n_summed.back();
 
@@ -3213,14 +3938,50 @@ namespace REX
             }
         };
 
-        check_particle_size(this->momenta_, "momenta");
-        check_particle_size(this->mass_, "mass");
-        check_particle_size(this->vtim_, "vtim");
-        check_particle_size(this->spin_, "spin");
-        check_particle_size(this->pdg_, "pdg");
-        check_particle_size(this->status_, "status");
-        check_particle_size(this->mother_, "mother");
-        check_particle_size(this->icol_, "icol");
+        if(this->column_major)
+        {
+            // umami_momenta_/umami_mother_/umami_icol_ are flat std::vectors
+            // (not vecArrN, whose size() divides by the stride automatically),
+            // storing 4 and 2 components per particle respectively -- their
+            // raw .size() is summed_n times that stride, not summed_n itself
+            auto check_strided_particle_size = [summed_n](const auto &vec, const char *name, size_t stride)
+            {
+                if (vec.size() != summed_n * stride)
+                {
+                    std::ostringstream oss;
+                    oss << "process::validate() failed: '" << name
+                        << "' has size " << vec.size() << ", expected " << (summed_n * stride);
+                    throw std::runtime_error(oss.str());
+                }
+            };
+            check_strided_particle_size(this->umami_momenta_, "umami_momenta", 4);
+            check_particle_size(this->umami_mass_, "umami_mass");
+            check_particle_size(this->umami_vtim_, "umami_vtim");
+            check_particle_size(this->umami_spin_, "umami_spin");
+            check_particle_size(this->umami_pdg_, "umami_pdg");
+            check_particle_size(this->umami_status_, "umami_status");
+            check_strided_particle_size(this->umami_mother_, "umami_mother", 2);
+            check_strided_particle_size(this->umami_icol_, "umami_icol", 2);
+        }
+        else
+        {
+            check_particle_size(this->momenta_, "momenta");
+            check_particle_size(this->mass_, "mass");
+            check_particle_size(this->vtim_, "vtim");
+            check_particle_size(this->spin_, "spin");
+            check_particle_size(this->pdg_, "pdg");
+            check_particle_size(this->status_, "status");
+            check_particle_size(this->mother_, "mother");
+            check_particle_size(this->icol_, "icol");
+        }
+        // check_particle_size(this->momenta_, "momenta");
+        // check_particle_size(this->mass_, "mass");
+        // check_particle_size(this->vtim_, "vtim");
+        // check_particle_size(this->spin_, "spin");
+        // check_particle_size(this->pdg_, "pdg");
+        // check_particle_size(this->status_, "status");
+        // check_particle_size(this->mother_, "mother");
+        // check_particle_size(this->icol_, "icol");
     }
 
     process &process::add_extra(const std::unordered_map<std::string, std::any> &extras)
@@ -3255,6 +4016,10 @@ namespace REX
             {
                 ev->set_muPS(this->muPS_[idx]);
             }
+        if (this->helicity_.size() > idx)
+            ev->set_helicity(this->helicity_[idx]);
+        if (this->flavor_.size() > idx)
+            ev->set_flavor(this->flavor_[idx]);
         size_t begin;
         if (idx == 0)
         {
@@ -3811,6 +4576,44 @@ namespace REX
         return *this;
     }
 
+    process &process::transpose_helicity()
+    {
+        this->validate();
+        if (this->helicity_.empty())
+            return *this;
+        if (this->helicity_.size() != this->events.size())
+        {
+            if (!this->events.empty())
+                warning("process::transpose_helicity() - Number of events does not match number of helicity values. Overwriting process::events vector.");
+            this->transpose();
+            return *this;
+        }
+        for (size_t i = 0; i < this->events.size(); ++i)
+        {
+            this->events[i]->set_helicity(this->helicity_[i]);
+        }
+        return *this;
+    }
+
+    process &process::transpose_flavor()
+    {
+        this->validate();
+        if (this->flavor_.empty())
+            return *this;
+        if (this->flavor_.size() != this->events.size())
+        {
+            if (!this->events.empty())
+                warning("process::transpose_flavor() - Number of events does not match number of flavor values. Overwriting process::events vector.");
+            this->transpose();
+            return *this;
+        }
+        for (size_t i = 0; i < this->events.size(); ++i)
+        {
+            this->events[i]->set_flavor(this->flavor_[i]);
+        }
+        return *this;
+    }
+
     process &process::transpose_E()
     {
         this->validate();
@@ -4084,6 +4887,7 @@ namespace REX
         this->nProc_ = proc_ids.size();
         std::sort(proc_ids.begin(), proc_ids.end());
         this->lProc_ = proc_ids;
+        this->pad_proc_arrays();
     }
 
     lhe::lhe(std::vector<event> evs)
@@ -4105,6 +4909,7 @@ namespace REX
         this->nProc_ = proc_ids.size();
         std::sort(proc_ids.begin(), proc_ids.end());
         this->lProc_ = proc_ids;
+        this->pad_proc_arrays();
     }
 
     lhe::lhe(const initNode &i, std::vector<std::shared_ptr<event>> evts) : initNode(i), events(std::move(evts))
@@ -4121,6 +4926,7 @@ namespace REX
         this->nProc_ = proc_ids.size();
         std::sort(proc_ids.begin(), proc_ids.end());
         this->lProc_ = proc_ids;
+        this->pad_proc_arrays();
     }
 
     lhe::lhe(const initNode &i, std::vector<event> evts) : initNode(i)
@@ -4142,6 +4948,7 @@ namespace REX
         this->nProc_ = proc_ids.size();
         std::sort(proc_ids.begin(), proc_ids.end());
         this->lProc_ = proc_ids;
+        this->pad_proc_arrays();
     }
 
     lhe &lhe::set_events(std::vector<std::shared_ptr<event>> evs)
@@ -4201,7 +5008,12 @@ namespace REX
         }
         if (this->header.type() != typeid(std::shared_ptr<xmlNode>))
         {
-            throw std::runtime_error("lhe::extract_weight_ids() - Header is not of type std::shared_ptr<xmlNode>");
+            // No XML header to pull pre-existing <initrwgt> weight ids from
+            // (eg an lhe sourced from madtrex's internal binary EventFile
+            // format, which carries no header at all) -- nothing to extract,
+            // not an error: weight_ids is left as the (possibly just-created)
+            // empty vector ensured above
+            return;
         }
         auto xml_header = std::any_cast<std::shared_ptr<xmlNode>>(this->header);
         auto id_puller = [&](std::shared_ptr<xmlNode> node)
@@ -4442,6 +5254,12 @@ namespace REX
         return *this;
     }
 
+    lhe &lhe::set_max_batch(std::size_t n)
+    {
+        this->max_batch = n;
+        return *this;
+    }
+
     void lhe::sort_events()
     {
         if (this->events.empty())
@@ -4452,33 +5270,87 @@ namespace REX
         {
             this->set_sorter();
         }
-        this->process_order.clear();
-        this->process_order.reserve(this->events.size());
-        for (auto ev : this->events)
+        this->process_order.assign(this->events.size(), npos);
+        // The first event on this thread, so that a classifier initialising
+        // itself lazily does so once, then the rest in chunks across
+        // sort_threads threads
+        this->process_order[0] = this->event_hash(*this->events[0]);
+        size_t n_rest = this->events.size() - 1;
+        size_t n_threads = std::max<size_t>(1, std::min(this->sort_threads, n_rest / sort_chunk));
+        if (n_threads <= 1)
         {
-            this->process_order.push_back(this->event_hash(*ev));
+            for (size_t ind = 1; ind < this->events.size(); ++ind)
+                this->process_order[ind] = this->event_hash(*this->events[ind]);
+        }
+        else
+        {
+            std::vector<std::thread> workers;
+            std::vector<std::exception_ptr> errors(n_threads);
+            size_t per_thread = (n_rest + n_threads - 1) / n_threads;
+            for (size_t t = 0; t < n_threads; ++t)
+            {
+                size_t begin = 1 + t * per_thread;
+                size_t end = std::min(this->events.size(), begin + per_thread);
+                workers.emplace_back([this, begin, end, &errors, t]
+                                     {
+                    try {
+                        for (size_t ind = begin; ind < end; ++ind)
+                            this->process_order[ind] = this->event_hash(*this->events[ind]);
+                    } catch (...) {
+                        errors[t] = std::current_exception();
+                    } });
+            }
+            for (auto &worker : workers)
+                worker.join();
+            for (auto &error : errors)
+                if (error)
+                    std::rethrow_exception(error);
         }
         if (this->process_order.size() != this->events.size())
         {
             throw std::runtime_error("lhe::sort_events() failed: process_order size does not match events size");
         }
         this->sorted_events.clear();
+        this->process_group.clear();
         for (size_t j = 0; j < this->sorter.size(); ++j)
         {
             this->sorted_events.push_back({});
+            this->process_group.push_back(j);
+        }
+        this->sorted_events.push_back({}); // catch-all "unsorted" bucket
+        this->process_group.push_back(npos);
+
+        // active_bucket[k] = index into sorted_events currently receiving events
+        // whose sorter class is k (k == npos is the catch-all bucket). Only grows
+        // beyond the identity mapping seeded below when max_batch is set and a
+        // bucket has filled up: the class's active bucket is then switched to a
+        // freshly appended one, which process_group records as still belonging to
+        // class k, so downstream consumers can recognise/recombine the batches
+        std::unordered_map<size_t, size_t> active_bucket;
+        for (size_t j = 0; j < this->sorter.size(); ++j)
+        {
+            active_bucket[j] = j;
         }
-        this->sorted_events.push_back({});
+        active_bucket[npos] = this->sorter.size();
+
         for (size_t ind = 0; ind < this->events.size(); ++ind)
         {
-            size_t sort_ind = this->process_order[ind];
-            if (sort_ind == npos)
-                sort_ind = this->sorted_events.size() - 1; // If an event does not belong to any set in the sorter, returns npos; compensate by adding an additional "unsorted" vector at the end
-            this->sorted_events[sort_ind].push_back(this->events[ind]);
+            size_t class_id = this->process_order[ind];
+            size_t &bucket = active_bucket[class_id];
+            if (this->max_batch != npos && this->sorted_events[bucket].size() >= this->max_batch)
+            {
+                this->sorted_events.push_back({});
+                this->process_group.push_back(class_id);
+                bucket = this->sorted_events.size() - 1;
+            }
+            this->sorted_events[bucket].push_back(this->events[ind]);
         }
-        // If all events were successfully sorted, remove the empty "unsorted" vector
-        if (this->sorted_events.back().empty())
+        // If all events were successfully sorted, remove the empty catch-all bucket
+        // (which, if unused, is still at its original seeded position)
+        if (this->sorted_events[this->sorter.size()].empty())
         {
-            this->sorted_events.pop_back(); // Remove empty last set
+            this->sorted_events.erase(this->sorted_events.begin() + this->sorter.size());
+            this->process_group.erase(this->process_group.begin() + this->sorter.size());
         }
     }
 
@@ -4511,11 +5383,11 @@ namespace REX
         {
             throw std::runtime_error("lhe::events_to_processes() called with no events");
         }
-        this->sort_events(); // Overrides previous sorting (if one exists)
+        this->sort_events(); // Overrides previous sorting (if one exists); also (re)populates process_group
         this->processes.clear();
-        for (auto ev_set : this->sorted_events)
+        for (const auto &ev_set : this->sorted_events)
         {
-            this->processes.push_back(std::make_shared<process>(ev_set, this->filter_processes));
+            this->processes.push_back(std::make_shared<process>(ev_set, this->filter_processes, this->column_major));
         }
     }
 
@@ -4623,6 +5495,30 @@ namespace REX
         {
             initrwgt = xmlNode::parse("<initrwgt>\n</initrwgt>\n");
         }
+        // Replace, not accumulate: a re-run of append_weight_ids() (eg
+        // write_weights() called a second time after more weight_ids were
+        // added -- would otherwise leave the *previous* call's <weightgroup name=
+        // 'tearex_reweighting'> node sitting alongside the new one, stale
+        // and listing an incomplete/outdated set of ids. Only removes a
+        // weightgroup matching this exact name -- any other pre-existing
+        // <weightgroup> under <initrwgt> (eg from an upstream tool, already
+        // present in the loaded file's own header) is left untouched.
+        // Uses remove_inserted_child(), not remove_child(): the previous
+        // weightgroup was itself added via add_child() (right below, on
+        // the prior call), so it's only ever emitted through the inserts_
+        // mechanism, which remove_child()'s suppressed_ flag has no effect
+        // on -- see remove_inserted_child()'s own doc comment
+        for (auto &child : initrwgt->get_children("weightgroup"))
+        {
+            for (auto &attr : child->attrs())
+            {
+                if (attr.name() == "name" && attr.value() == "tearex_reweighting")
+                {
+                    initrwgt->remove_inserted_child(child.get());
+                    break;
+                }
+            }
+        }
         initrwgt->add_child(wgt_node, true);
         if (this->header.type() == typeid(REX::xmlNode))
         {
@@ -4634,26 +5530,163 @@ namespace REX
         }
     }
 
-    void lhe::print_header(std::ostream &os) const
+    lhe &lhe::add_histogram(const std::string &name, std::shared_ptr<histogram> h)
+    {
+        this->histograms[name] = std::move(h);
+        return *this;
+    }
+
+    lhe &lhe::add_histogram(const std::string &name, histogram h)
+    {
+        this->histograms[name] = std::make_shared<histogram>(std::move(h));
+        return *this;
+    }
+
+    lhe &lhe::make_histogram(const std::string &name, const std::vector<size_t> &indices)
+    {
+        auto h = std::make_shared<histogram>(this->events, indices);
+        h->normalise(*this);
+        this->histograms[name] = std::move(h);
+        return *this;
+    }
+
+    lhe &lhe::make_histogram(const std::string &name, eventSorter sorter)
+    {
+        auto h = std::make_shared<histogram>(this->events, std::move(sorter));
+        h->normalise(*this);
+        this->histograms[name] = std::move(h);
+        return *this;
+    }
+
+    lhe &lhe::make_histogram(const std::string &name, event_hash_fn hash)
+    {
+        auto h = std::make_shared<histogram>(this->events, hash);
+        h->normalise(*this);
+        this->histograms[name] = std::move(h);
+        return *this;
+    }
+
+    lhe &lhe::make_histogram(const std::string &name, const eventBelongs &e_set)
+    {
+        auto h = std::make_shared<histogram>(this->events, e_set);
+        h->normalise(*this);
+        this->histograms[name] = std::move(h);
+        return *this;
+    }
+
+    lhe &lhe::make_histogram(const std::string &name, event_bool_fn comp)
+    {
+        auto h = std::make_shared<histogram>(this->events, comp);
+        h->normalise(*this);
+        this->histograms[name] = std::move(h);
+        return *this;
+    }
+
+    lhe &lhe::make_histogram(const std::string &name, cevent_bool_fn comp)
     {
-        if (this->header.type() == typeid(std::string))
+        auto h = std::make_shared<histogram>(this->events, comp);
+        h->normalise(*this);
+        this->histograms[name] = std::move(h);
+        return *this;
+    }
+
+    lhe &lhe::make_histogram(const std::string &name, std::vector<eventBelongs> e_sets)
+    {
+        auto h = std::make_shared<histogram>(this->events, std::move(e_sets));
+        h->normalise(*this);
+        this->histograms[name] = std::move(h);
+        return *this;
+    }
+
+    lhe &lhe::make_histogram(const std::string &name, std::vector<event_bool_fn> comps)
+    {
+        auto h = std::make_shared<histogram>(this->events, std::move(comps));
+        h->normalise(*this);
+        this->histograms[name] = std::move(h);
+        return *this;
+    }
+
+    bool lhe::has_histogram(const std::string &name) const
+    {
+        return this->histograms.find(name) != this->histograms.end();
+    }
+
+    histogram &lhe::hist(const std::string &name)
+    {
+        auto it = this->histograms.find(name);
+        if (it == this->histograms.end())
+            throw std::out_of_range("lhe::hist: no histogram named '" + name + "'");
+        return *it->second;
+    }
+
+    const histogram &lhe::hist(const std::string &name) const
+    {
+        auto it = this->histograms.find(name);
+        if (it == this->histograms.end())
+            throw std::out_of_range("lhe::hist: no histogram named '" + name + "'");
+        return *it->second;
+    }
+
+    lhe &lhe::remove_histogram(const std::string &name)
+    {
+        this->histograms.erase(name);
+        return *this;
+    }
+
+    bool lhe::sync_histograms()
+    {
+        bool ok = true;
+        for (auto &kv : this->histograms)
         {
-            os << std::any_cast<std::string>(this->header);
+            if (kv.second && !kv.second->sync_wgts())
+                ok = false;
         }
-        else if (this->header.type() == typeid(std::shared_ptr<xmlNode>))
+        return ok;
+    }
+
+    bool lhe::sync_histogram(const std::string &name)
+    {
+        auto it = this->histograms.find(name);
+        if (it == this->histograms.end())
+            throw std::out_of_range("lhe::sync_histogram: no histogram named '" + name + "'");
+        return it->second ? it->second->sync_wgts() : true;
+    }
+
+    void print_lhe_header(const std::any &header, std::ostream &os)
+    {
+        if (header.type() == typeid(std::string))
         {
-            std::any_cast<std::shared_ptr<xmlNode>>(this->header)->write(os);
+            os << std::any_cast<std::string>(header);
         }
-        else if (this->header.type() == typeid(xmlNode))
+        else if (header.type() == typeid(std::shared_ptr<xmlNode>))
         {
-            std::any_cast<xmlNode>(this->header).write(os);
+            std::any_cast<std::shared_ptr<xmlNode>>(header)->write(os);
         }
-        else if (this->header.has_value())
+        else if (header.type() == typeid(xmlNode))
         {
-            warning("lhe::print_header() - Header is of unknown type, cannot print.");
+            std::any_cast<xmlNode>(header).write(os);
+        }
+        else if (header.has_value())
+        {
+            warning("print_lhe_header() - Header is of unknown type, cannot print.");
         }
     }
 
+    void lhe::pad_proc_arrays()
+    {
+        if (this->xSec_.size() < this->nProc_)
+            this->xSec_.resize(this->nProc_, 0.0);
+        if (this->xSecErr_.size() < this->nProc_)
+            this->xSecErr_.resize(this->nProc_, 0.0);
+        if (this->xMax_.size() < this->nProc_)
+            this->xMax_.resize(this->nProc_, 0.0);
+    }
+
+    void lhe::print_header(std::ostream &os) const
+    {
+        print_lhe_header(this->header, os);
+    }
+
     void lhe::print(std::ostream &os, bool include_ids)
     {
         this->append_weight_ids(include_ids);
@@ -5086,6 +6119,13 @@ namespace REX
     }
 
     std::shared_ptr<xmlNode> xmlNode::parse_element(const std::shared_ptr<const std::string> &doc, size_t &pos)
+    {
+        std::vector<std::string_view> ancestors;
+        return parse_element(doc, pos, ancestors);
+    }
+
+    std::shared_ptr<xmlNode> xmlNode::parse_element(const std::shared_ptr<const std::string> &doc, size_t &pos,
+                                                    std::vector<std::string_view> &ancestors)
     {
         const std::string &s = *doc;
         const size_t N = s.size();
@@ -5147,6 +6187,26 @@ namespace REX
             // Check for closing tag of this node
             if (lt + 1 < N && s[lt + 1] == '/')
             {
+                // A closing tag for one of the enclosing elements instead
+                // means this element is never closed: it is not an element
+                // at all but text that happens to look like a start tag --
+                // eg "<run>" or "<set>@<member>" in the comments of an MG7
+                // run card embedded unescaped in an LHE header. Give up on
+                // it so the caller keeps it as content; otherwise this
+                // closing tag would end this element instead, and every
+                // later sibling of the enclosing element (<initrwgt>, say)
+                // would end up nested inside it. A closing tag naming no
+                // open element at all still just ends this one, as before.
+                size_t close_beg = lt + 2;
+                while (close_beg < N && is_space(s[close_beg]))
+                    ++close_beg;
+                size_t close_end = close_beg;
+                while (close_end < N && !is_space(s[close_end]) && s[close_end] != '>')
+                    ++close_end;
+                std::string_view close_name(s.data() + close_beg, close_end - close_beg);
+                if (close_name != node->name_view_ &&
+                    std::find(ancestors.begin(), ancestors.end(), close_name) != ancestors.end())
+                    return nullptr;
                 node->content_end_ = lt;
                 // find end of closing tag
                 size_t gt = s.find('>', lt + 2);
@@ -5175,7 +6235,9 @@ namespace REX
 
             // Child node
             size_t child_pos = lt;
-            auto child = parse_element(doc, child_pos);
+            ancestors.push_back(node->name_view_);
+            auto child = parse_element(doc, child_pos, ancestors);
+            ancestors.pop_back();
             if (!child)
             {
                 // If parse failed, avoid infinite loop by moving cursor forward
@@ -5542,6 +6604,36 @@ namespace REX
         return false;
     }
 
+    bool xmlNode::remove_inserted_child(const xmlNode *child) noexcept
+    {
+        bool found = false;
+        for (auto it = inserts_.begin(); it != inserts_.end();)
+        {
+            if (it->node == child)
+            {
+                it = inserts_.erase(it);
+                found = true;
+            }
+            else
+            {
+                ++it;
+            }
+        }
+        for (auto it = children_.begin(); it != children_.end(); ++it)
+        {
+            if (it->get() == child)
+            {
+                children_.erase(it);
+                break;
+            }
+        }
+        if (found)
+        {
+            modified_ = true;
+        }
+        return found;
+    }
+
     const xmlNode *xmlNode::nth_in_doc_child(size_t ordinal) const noexcept
     {
         size_t seen = 0;
@@ -5847,7 +6939,8 @@ namespace REX
         std::string buf;
         while (buf.empty())
         {
-            std::getline(in, buf);
+            if (!std::getline(in, buf))
+                throw std::runtime_error("load_lhef: empty or unreadable stream");
         }
         if (buf.find("LesHouchesEvents") == npos)
         {
@@ -5876,7 +6969,8 @@ namespace REX
         }
         while (buf.find("init") == npos)
         {
-            std::getline(in, buf);
+            if (!std::getline(in, buf))
+                throw std::runtime_error("load_lhef: malformed LHEF file, no init block found");
         }
         if (buf.find("</init") != npos || buf.find("/>") != npos)
         {
@@ -5936,6 +7030,284 @@ namespace REX
         return load_lhef(stream);
     }
 
+    // -------------------- streaming eventSource / eventSink (LHE-XML) --------------------
+
+    namespace
+    {
+        // Resumable counterpart to load_lhef()'s scanning loop: the same
+        // line-by-line, one-<event>-block-at-a-time parse, but split across
+        // read_init()/pull(n) calls instead of running to completion in one
+        // go. `content` always holds either nothing, or the not-yet-flushed
+        // lines of the event currently being accumulated, so a pull() call
+        // can stop after n events and resume correctly on the next call
+        struct lheStreamState
+        {
+            std::shared_ptr<std::istream> in;
+            bool started = false;
+            bool exhausted = false;
+            std::string content;
+            initNode cached_init;
+            std::any cached_header;
+
+            void ensure_started()
+            {
+                if (started)
+                    return;
+                started = true;
+                std::string buf;
+                while (buf.empty())
+                {
+                    if (!std::getline(*in, buf))
+                        throw std::runtime_error("open_lhe_source: empty or unreadable stream");
+                }
+                if (buf.find("LesHouchesEvents") == npos)
+                    throw std::runtime_error("open_lhe_source: not a valid LHEF file");
+                std::getline(*in, buf);
+                if (buf.find("header") != npos)
+                {
+                    std::string hdr_content;
+                    if (buf.find("</header") != npos || buf.find("/>") != npos)
+                    {
+                        cached_header = xmlNode::parse(buf);
+                    }
+                    else
+                    {
+                        hdr_content += buf + "\n";
+                        while (std::getline(*in, buf))
+                        {
+                            hdr_content += buf + "\n";
+                            if (buf.find("/header") != npos)
+                                break;
+                        }
+                        cached_header = xmlNode::parse(hdr_content);
+                    }
+                }
+                while (buf.find("init") == npos)
+                {
+                    if (!std::getline(*in, buf))
+                        throw std::runtime_error("open_lhe_source: malformed LHEF file, no init block found");
+                }
+                if (buf.find("</init") != npos || buf.find("/>") != npos)
+                    throw std::runtime_error("open_lhe_source: malformed init block");
+                std::string init_content = buf + "\n";
+                while (std::getline(*in, buf))
+                {
+                    init_content += buf + "\n";
+                    if (buf.find("/init") != npos)
+                        break;
+                }
+                cached_init = xml_to_init(xmlNode::parse(init_content));
+            }
+
+            initNode read_init()
+            {
+                ensure_started();
+                return cached_init;
+            }
+
+            std::any read_header()
+            {
+                ensure_started();
+                return cached_header;
+            }
+
+            std::vector<std::shared_ptr<event>> pull(size_t n)
+            {
+                ensure_started();
+                std::vector<std::shared_ptr<event>> events;
+                if (exhausted || n == 0)
+                    return events;
+                std::string buf;
+                while (std::getline(*in, buf))
+                {
+                    if (buf.find("<event") != npos)
+                    {
+                        if (!content.empty())
+                        {
+                            auto curr_event = xml_to_event(xmlNode::parse(content));
+                            if (curr_event)
+                                events.push_back(curr_event);
+                        }
+                        content.clear();
+                        content += buf + "\n";
+                        if (events.size() >= n)
+                            return events;
+                    }
+                    else if (buf.find("</LesHouchesEvents") != npos)
+                    {
+                        if (!content.empty())
+                        {
+                            auto curr_event = xml_to_event(xmlNode::parse(content));
+                            if (curr_event)
+                                events.push_back(curr_event);
+                        }
+                        content.clear();
+                        exhausted = true;
+                        return events;
+                    }
+                    else
+                    {
+                        content += buf + "\n";
+                    }
+                }
+                // Stream ended without a closing tag: mirrors load_lhef()'s
+                // behaviour on a truncated file (whatever is left pending in
+                // `content` is not flushed).
+                exhausted = true;
+                return events;
+            }
+        };
+
+        eventSource make_lhe_source(std::shared_ptr<std::istream> in)
+        {
+            auto state = std::make_shared<lheStreamState>();
+            state->in = std::move(in);
+            eventSource src;
+            src.read_init = [state]()
+            { return state->read_init(); };
+            src.read_header = [state]()
+            { return state->read_header(); };
+            src.pull = [state](size_t n)
+            { return state->pull(n); };
+            // Forward-only: rewind is intentionally left null. Re-open the
+            // file via open_lhe_source() again for a second pass.
+            return src;
+        }
+    } // namespace
+
+    eventSource open_lhe_source(const std::string &filename)
+    {
+        auto stream = std::make_shared<std::ifstream>(filename);
+        if (!*stream)
+            throw std::ios_base::failure("open_lhe_source: could not open file for reading");
+        return make_lhe_source(stream);
+    }
+
+    eventSource open_lhe_source(std::istream &in)
+    {
+        // Non-owning: caller-supplied stream must outlive the source.
+        std::shared_ptr<std::istream> shared_in(&in, [](std::istream *) {});
+        return make_lhe_source(shared_in);
+    }
+
+    eventSink open_lhe_sink(const std::string &filename, bool include_ids)
+    {
+        auto out = std::make_shared<std::ofstream>(filename);
+        if (!*out)
+            throw std::ios_base::failure("open_lhe_sink: could not open file for writing");
+        eventSink sink;
+        sink.write_open = [out](const initNode &init, const std::any &header)
+        {
+            *out << "<LesHouchesEvents version=\"3.0\">\n";
+            print_lhe_header(header, *out);
+            init.print_init(*out);
+        };
+        sink.write_event = [out, include_ids](std::shared_ptr<event> ev)
+        {
+            if (ev)
+                ev->print(*out, include_ids);
+        };
+        sink.write_close = [out]()
+        {
+            *out << "\n</LesHouchesEvents>";
+            // Explicit close rather than relying on the shared_ptr<ofstream>
+            // eventually going out of scope: a caller that reads the file
+            // back right after write_close() returns (the natural thing to
+            // do -- and the only viable timing from a GC'd binding, where
+            // "this eventSink is unreachable" isn't scope-bound) must see a
+            // complete file, not whatever ofstream's own internal buffer
+            // happened to have flushed to the OS on its own.
+            out->close();
+        };
+        return sink;
+    }
+
+    std::vector<size_t> split_source(
+        eventSource &src,
+        size_t n_shards,
+        const std::function<eventSink(size_t)> &make_sink,
+        size_t batch_size
+    )
+    {
+        if (n_shards == 0)
+            throw std::runtime_error("split_source: n_shards must be > 0");
+        if (!src.read_init)
+            throw std::runtime_error("split_source: eventSource has no read_init set");
+        if (!src.pull)
+            throw std::runtime_error("split_source: eventSource has no pull set");
+        if (!make_sink)
+            throw std::runtime_error("split_source: make_sink is not set");
+        if (batch_size == 0 || batch_size == npos)
+            batch_size = 1;
+
+        initNode init = src.read_init();
+        std::any header = src.read_header ? src.read_header() : std::any{};
+
+        std::vector<eventSink> sinks;
+        sinks.reserve(n_shards);
+        for (size_t i = 0; i < n_shards; ++i)
+        {
+            eventSink sink = make_sink(i);
+            if (!sink.write_open || !sink.write_event || !sink.write_close)
+                throw std::runtime_error(
+                    "split_source: make_sink(" + std::to_string(i) +
+                    ") returned an incomplete eventSink (missing write_open/write_event/write_close)");
+            sink.write_open(init, header);
+            sinks.push_back(std::move(sink));
+        }
+
+        std::vector<size_t> counts(n_shards, 0);
+        size_t next_shard = 0;
+        std::vector<std::shared_ptr<event>> batch;
+        while (!(batch = src.pull(batch_size)).empty())
+        {
+            for (auto &ev : batch)
+            {
+                sinks[next_shard].write_event(ev);
+                ++counts[next_shard];
+                next_shard = (next_shard + 1) % n_shards;
+            }
+        }
+
+        for (auto &sink : sinks)
+            sink.write_close();
+
+        return counts;
+    }
+
+    lhe lhe::open_streaming(std::shared_ptr<eventSource> src, std::size_t batch_size)
+    {
+        if (!src || !src->read_init)
+            throw std::runtime_error("lhe::open_streaming: eventSource has no read_init set");
+        if (!src->pull)
+            throw std::runtime_error("lhe::open_streaming: eventSource has no pull set");
+        initNode init = src->read_init();
+        lhe doc(init);
+        if (src->read_header)
+        {
+            auto header = src->read_header();
+            if (header.has_value())
+                doc.set_header(header);
+        }
+        doc.source = src;
+        doc.batch_size = (batch_size == 0 || batch_size == npos) ? 1 : batch_size;
+        return doc;
+    }
+
+    std::size_t lhe::pull_batch()
+    {
+        if (!this->streaming())
+            throw std::runtime_error("lhe::pull_batch: lhe is not in streaming mode (no source set); use open_streaming() first");
+        this->events.clear();
+        this->sorted_events.clear();
+        this->processes.clear();
+        this->process_group.clear();
+        this->process_order.clear();
+        this->events = this->source->pull(this->batch_size);
+        this->events_pulled_total += this->events.size();
+        return this->events.size();
+    }
+
     void write_lhef(lhe &doc, std::ostream &out, bool include_weight_ids)
     {
         doc.print(out, include_weight_ids);
@@ -6164,6 +7536,17 @@ namespace REX
         return get(block, {i1}, fallback);
     }
 
+    double slha::get(const std::string &block,
+                     const std::vector<int> &indices,
+                     double fallback) const
+    {
+        auto itb = blocks_.find(upper(block));
+        if (itb == blocks_.end())
+            return fallback;
+        auto itv = itb->second.entries.find(indices);
+        return (itv == itb->second.entries.end()) ? fallback : itv->second;
+    }
+
     void slha::set(const std::string &block,
                    std::initializer_list<int> indices,
                    double value)
@@ -6180,6 +7563,13 @@ namespace REX
         set(block, {i1}, value);
     }
 
+    void slha::set(const std::string &block,
+                   const std::vector<int> &indices,
+                   double value)
+    {
+        blocks_[upper(block)].entries[indices] = value;
+    }
+
     double slha::get_decay(int pid, double fallback) const
     {
         auto it = decays_.find(pid);
@@ -6225,7 +7615,7 @@ namespace REX
         return to_slha(content);
     }
 
-    slha to_slha(lhe &doc)
+    slha to_slha(const lhe &doc)
     {
         if (doc.header.type() == typeid(slha))
             return std::any_cast<slha>(doc.header);
diff --git a/madtrex/Rex/src/python/rex.cpp b/madtrex/Rex/src/python/rex.cpp
new file mode 100644
index 0000000000..1be1271892
--- /dev/null
+++ b/madtrex/Rex/src/python/rex.cpp
@@ -0,0 +1,985 @@
+#include <pybind11/functional.h>
+#include <pybind11/pybind11.h>
+#include <pybind11/stl.h>
+
+#include "Rex.h"
+#include "tearex.hpp"
+
+#include <sstream>
+
+namespace py = pybind11;
+using namespace REX;
+
+namespace {
+
+// ---- arrN<T,N> / vecArrN<T,N> <-> plain Python lists ----
+//
+// Rex's fixed-size arrN<T,N> (a single momentum/mother/icol tuple) and
+// vecArrN<T,N> (a flat, strided vector of those) aren't STL types, so
+// pybind11/stl.h doesn't know how to convert them. Rather than writing a
+// generic type_caster, bind them as ordinary list-valued properties: read
+// copies out to a list, write rebuilds the arrN/vecArrN from a list
+
+template <typename T, std::size_t N>
+std::vector<T> arrN_to_vec(const arrN<T, N> &a) {
+    std::vector<T> v(N);
+    for (std::size_t i = 0; i < N; ++i) v[i] = a[i];
+    return v;
+}
+
+template <typename T, std::size_t N>
+arrN<T, N> vec_to_arrN(const std::vector<T> &v) {
+    if (v.size() != N) {
+        throw std::invalid_argument(
+            "expected a sequence of length " + std::to_string(N) + ", got "
+            + std::to_string(v.size())
+        );
+    }
+    arrN<T, N> a;
+    for (std::size_t i = 0; i < N; ++i) a[i] = v[i];
+    return a;
+}
+
+template <typename T, std::size_t N>
+std::vector<std::vector<T>> vecArrN_to_nested(const vecArrN<T, N> &v) {
+    std::vector<std::vector<T>> out(v.size());
+    for (std::size_t i = 0; i < v.size(); ++i) {
+        out[i].resize(N);
+        for (std::size_t j = 0; j < N; ++j) out[i][j] = v[i][j];
+    }
+    return out;
+}
+
+template <typename T, std::size_t N>
+vecArrN<T, N> nested_to_vecArrN(const std::vector<std::vector<T>> &in) {
+    vecArrN<T, N> out(in.size());
+    for (std::size_t i = 0; i < in.size(); ++i) {
+        if (in[i].size() != N) {
+            throw std::invalid_argument(
+                "expected inner sequences of length " + std::to_string(N)
+            );
+        }
+        for (std::size_t j = 0; j < N; ++j) out[i][j] = in[i][j];
+    }
+    return out;
+}
+
+template <typename T, std::size_t N, typename C>
+void bind_arrN_property(py::classh<C> &cls, const char *name, arrN<T, N> C::*field) {
+    cls.def_property(
+        name,
+        [field](const C &self) { return arrN_to_vec(self.*field); },
+        [field](C &self, const std::vector<T> &v) { self.*field = vec_to_arrN<T, N>(v); }
+    );
+}
+
+template <typename T, std::size_t N, typename C>
+void bind_vecArrN_property(py::classh<C> &cls, const char *name, vecArrN<T, N> C::*field) {
+    cls.def_property(
+        name,
+        [field](const C &self) { return vecArrN_to_nested(self.*field); },
+        [field](C &self, const std::vector<std::vector<T>> &v) {
+            self.*field = nested_to_vecArrN<T, N>(v);
+        }
+    );
+}
+
+// ---- arrNRef<T,N> <-> plain Python lists ----
+//
+// event::particle/const_particle expose their momentum/mother/icol fields
+// as arrNRef<T,N> -- a raw-pointer proxy into the *owning event's* own
+// storage vectors (not an owning array like arrN<T,N> above), so writing
+// through it writes straight back into that event. Converts the same way:
+// read copies out to a list, write copies a list's values back through the
+// proxy (never rebinding what it points at)
+template <typename T, std::size_t N>
+std::vector<std::remove_const_t<T>> arrNRef_to_vec(const arrNRef<T, N> &a) {
+    std::vector<std::remove_const_t<T>> v(N);
+    for (std::size_t i = 0; i < N; ++i) v[i] = a[i];
+    return v;
+}
+
+template <typename T, std::size_t N>
+void vec_to_arrNRef(arrNRef<T, N> &a, const std::vector<std::remove_const_t<T>> &v) {
+    if (v.size() != N) {
+        throw std::invalid_argument(
+            "expected a sequence of length " + std::to_string(N) + ", got "
+            + std::to_string(v.size())
+        );
+    }
+    for (std::size_t i = 0; i < N; ++i) a[i] = v[i];
+}
+
+// ---- index-based iterator helpers for Event/EventView ----
+//
+// event::particle_iterator/event::event_view::iterator only implement
+// operator!=, not operator==, but pybind11::make_iterator (this version)
+// needs operator== to detect the end. Wrapping plain indices here (rather
+// than the library's own iterator types) sidesteps that without changing
+// Rex.h's public iterator API
+struct event_particle_iter {
+    event *evt;
+    std::size_t i;
+    event::particle operator*() const { return evt->get_particle(i); }
+    event_particle_iter &operator++() {
+        ++i;
+        return *this;
+    }
+    bool operator==(const event_particle_iter &o) const { return i == o.i; }
+    bool operator!=(const event_particle_iter &o) const { return i != o.i; }
+};
+
+struct event_view_particle_iter {
+    event::event_view *view;
+    std::size_t i;
+    event::particle operator*() const { return (*view)[i]; }
+    event_view_particle_iter &operator++() {
+        ++i;
+        return *this;
+    }
+    bool operator==(const event_view_particle_iter &o) const { return i == o.i; }
+    bool operator!=(const event_view_particle_iter &o) const { return i != o.i; }
+};
+
+} // namespace
+
+PYBIND11_MODULE(_rex_py, m) {
+    m.doc() = "Python bindings for the Rex/teaRex LHE parsing and reweighting libraries";
+    // Generic "not found"/"unset" sentinel used eg by Lhe.max_batch (unset)
+    // EventSorter.position()/sort() (no match), and Lhe.process_group (catch-all)
+    m.attr("npos") = npos;
+
+    // ---- parton ----
+    auto parton_cls =
+        py::classh<parton>(m, "Parton")
+            .def(py::init<>())
+            .def_readwrite("mass", &parton::mass_)
+            .def_readwrite("vtim", &parton::vtim_)
+            .def_readwrite("spin", &parton::spin_)
+            .def_readwrite("pdg", &parton::pdg_)
+            .def_readwrite("status", &parton::status_)
+            .def_property_readonly("pT", &parton::pT)
+            .def_property_readonly("pT2", &parton::pT2)
+            .def_property_readonly("pL", &parton::pL)
+            .def_property_readonly("pL2", &parton::pL2)
+            .def_property_readonly("eT", &parton::eT)
+            .def_property_readonly("eT2", &parton::eT2)
+            .def_property_readonly("phi", &parton::phi)
+            .def_property_readonly("theta", &parton::theta)
+            .def_property_readonly("eta", &parton::eta)
+            .def_property_readonly("rap", &parton::rap)
+            .def_property_readonly("mT", &parton::mT)
+            .def_property_readonly("mT2", &parton::mT2)
+            .def_property_readonly("m2", &parton::m2);
+    bind_arrN_property(parton_cls, "momentum", &parton::momenta_);
+    bind_arrN_property(parton_cls, "mother", &parton::mother_);
+    bind_arrN_property(parton_cls, "icol", &parton::icol_);
+
+    // ---- event::particle / event::const_particle ----
+    // Live views into the *owning event's* own storage vectors (arrNRef
+    // members are raw-pointer proxies, not copies) -- so writing to a
+    // Particle's fields writes straight back into the Event it came from.
+    // Bound before Event itself since Event's get_particle/__getitem__/
+    // __iter__/view() return these. keep_alive<0, 1>() on every factory
+    // that returns one (see below) ties the returned view's lifetime to
+    // its parent Event, so the proxied vectors can't be freed out from
+    // under it
+    py::classh<event::particle>(m, "Particle")
+        .def_property(
+            "momentum",
+            [](const event::particle &p) { return arrNRef_to_vec(p.momentum_); },
+            [](event::particle &p, const std::vector<double> &v) {
+                vec_to_arrNRef(p.momentum_, v);
+            }
+        )
+        .def_property(
+            "mass",
+            [](const event::particle &p) { return p.mass_; },
+            [](event::particle &p, double v) { p.mass_ = v; }
+        )
+        .def_property(
+            "vtim",
+            [](const event::particle &p) { return p.vtim_; },
+            [](event::particle &p, double v) { p.vtim_ = v; }
+        )
+        .def_property(
+            "spin",
+            [](const event::particle &p) { return p.spin_; },
+            [](event::particle &p, double v) { p.spin_ = v; }
+        )
+        .def_property(
+            "pdg",
+            [](const event::particle &p) { return p.pdg_; },
+            [](event::particle &p, long int v) { p.pdg_ = v; }
+        )
+        .def_property(
+            "status",
+            [](const event::particle &p) { return p.status_; },
+            [](event::particle &p, short int v) { p.status_ = v; }
+        )
+        .def_property(
+            "mother",
+            [](const event::particle &p) { return arrNRef_to_vec(p.mother_); },
+            [](event::particle &p, const std::vector<short int> &v) {
+                vec_to_arrNRef(p.mother_, v);
+            }
+        )
+        .def_property(
+            "icol",
+            [](const event::particle &p) { return arrNRef_to_vec(p.icol_); },
+            [](event::particle &p, const std::vector<short int> &v) {
+                vec_to_arrNRef(p.icol_, v);
+            }
+        )
+        .def_property_readonly("pT", &event::particle::pT)
+        .def_property_readonly("pT2", &event::particle::pT2)
+        .def_property_readonly("pL", &event::particle::pL)
+        .def_property_readonly("pL2", &event::particle::pL2)
+        .def_property_readonly("eT", &event::particle::eT)
+        .def_property_readonly("eT2", &event::particle::eT2)
+        .def_property_readonly("phi", &event::particle::phi)
+        .def_property_readonly("theta", &event::particle::theta)
+        .def_property_readonly("eta", &event::particle::eta)
+        .def_property_readonly("rap", &event::particle::rap)
+        .def_property_readonly("mT", &event::particle::mT)
+        .def_property_readonly("mT2", &event::particle::mT2)
+        .def_property_readonly("m2", &event::particle::m2)
+        .def("__str__", [](const event::particle &p) {
+            std::ostringstream ss;
+            p.print(ss);
+            return ss.str();
+        });
+
+    py::classh<event::const_particle>(m, "ConstParticle")
+        .def_property_readonly(
+            "momentum", [](const event::const_particle &p) { return arrNRef_to_vec(p.momentum_); }
+        )
+        .def_property_readonly("mass", [](const event::const_particle &p) { return p.mass_; })
+        .def_property_readonly("vtim", [](const event::const_particle &p) { return p.vtim_; })
+        .def_property_readonly("spin", [](const event::const_particle &p) { return p.spin_; })
+        .def_property_readonly("pdg", [](const event::const_particle &p) { return p.pdg_; })
+        .def_property_readonly("status", [](const event::const_particle &p) { return p.status_; })
+        .def_property_readonly(
+            "mother", [](const event::const_particle &p) { return arrNRef_to_vec(p.mother_); }
+        )
+        .def_property_readonly(
+            "icol", [](const event::const_particle &p) { return arrNRef_to_vec(p.icol_); }
+        )
+        .def_property_readonly("pT", &event::const_particle::pT)
+        .def_property_readonly("pT2", &event::const_particle::pT2)
+        .def_property_readonly("pL", &event::const_particle::pL)
+        .def_property_readonly("pL2", &event::const_particle::pL2)
+        .def_property_readonly("eT", &event::const_particle::eT)
+        .def_property_readonly("eT2", &event::const_particle::eT2)
+        .def_property_readonly("phi", &event::const_particle::phi)
+        .def_property_readonly("theta", &event::const_particle::theta)
+        .def_property_readonly("eta", &event::const_particle::eta)
+        .def_property_readonly("rap", &event::const_particle::rap)
+        .def_property_readonly("mT", &event::const_particle::mT)
+        .def_property_readonly("mT2", &event::const_particle::mT2)
+        .def_property_readonly("m2", &event::const_particle::m2)
+        .def("__str__", [](const event::const_particle &p) {
+            std::ostringstream ss;
+            p.print(ss);
+            return ss.str();
+        });
+
+    // ---- event::event_view ----
+    // A reordered, read/write view over an event's particles according to
+    // its `indices` (see event.set_indices()/view() below), without
+    // touching the underlying storage vectors
+    py::classh<event::event_view>(m, "EventView")
+        .def("__len__", &event::event_view::size)
+        .def(
+            "__getitem__",
+            [](event::event_view &self, std::size_t i) { return self[i]; },
+            py::keep_alive<0, 1>()
+        )
+        .def(
+            "__iter__",
+            [](event::event_view &self) {
+                return py::make_iterator(
+                    event_view_particle_iter{&self, 0},
+                    event_view_particle_iter{&self, self.size()}
+                );
+            },
+            py::keep_alive<0, 1>()
+        );
+
+    // ---- event ----
+    auto event_cls =
+        py::classh<event>(m, "Event")
+            .def(py::init<>())
+            .def(py::init<std::size_t>(), py::arg("n_particles"))
+            .def_readwrite("n", &event::n_)
+            .def_readwrite("proc_id", &event::proc_id_)
+            .def_readwrite("weight", &event::weight_)
+            .def_readwrite("scale", &event::scale_)
+            .def_readwrite("muF", &event::muF_)
+            .def_readwrite("muR", &event::muR_)
+            .def_readwrite("muPS", &event::muPS_)
+            .def_readwrite("alphaEW", &event::alphaEW_)
+            .def_readwrite("alphaS", &event::alphaS_)
+            .def_readwrite("mass", &event::mass_)
+            .def_readwrite("vtim", &event::vtim_)
+            .def_readwrite("spin", &event::spin_)
+            .def_readwrite("pdg", &event::pdg_)
+            .def_readwrite("status", &event::status_)
+            .def_readwrite("wgts", &event::wgts_)
+            .def_readwrite("helicity", &event::helicity_)
+            .def_readwrite("flavor", &event::flavor_)
+            .def_readwrite("indices", &event::indices)
+            .def(
+                "add_particle",
+                py::overload_cast<const parton &>(&event::add_particle)
+            )
+            .def("size", &event::size)
+            .def("__len__", &event::size)
+            .def(
+                "__getitem__",
+                [](event &self, std::size_t i) { return self.get_particle(i); },
+                py::arg("index"),
+                py::keep_alive<0, 1>()
+            )
+            .def(
+                "__iter__",
+                [](event &self) {
+                    return py::make_iterator(
+                        event_particle_iter{&self, 0}, event_particle_iter{&self, self.size()}
+                    );
+                },
+                py::keep_alive<0, 1>()
+            )
+            .def(
+                "get_particle",
+                [](event &self, std::size_t i) { return self.get_particle(i); },
+                py::arg("index"),
+                py::keep_alive<0, 1>()
+            )
+            .def(
+                "view", [](event &self) { return self.view(); }, py::keep_alive<0, 1>()
+            )
+            .def("set_indices", py::overload_cast<>(&event::set_indices))
+            .def(
+                "set_indices",
+                py::overload_cast<const event &, bool>(&event::set_indices),
+                py::arg("other"),
+                py::arg("fail_on_mismatch") = false
+            )
+            .def(
+                "set_indices",
+                py::overload_cast<const std::vector<std::size_t> &>(&event::set_indices),
+                py::arg("indices")
+            )
+            .def("get_muF", &event::get_muF)
+            .def("get_muR", &event::get_muR)
+            .def("get_muPS", &event::get_muPS)
+            .def("__eq__", [](const event &a, const event &b) { return a == b; })
+            .def("__ne__", [](const event &a, const event &b) { return a != b; })
+            .def("validate", &event::validate);
+    bind_vecArrN_property(event_cls, "momenta", &event::momenta_);
+    bind_vecArrN_property(event_cls, "mother", &event::mother_);
+    bind_vecArrN_property(event_cls, "icol", &event::icol_);
+    m.def("default_event_equal", &default_event_equal, py::arg("lhs"), py::arg("rhs"));
+    // set_event_comparator's only overload takes cevent_equal_fn -- unlike
+    // eg eventBelongs/eventSorter's constructor pairs, there's no sibling
+    // event_equal_fn overload of the same name to collide with, so this is
+    // bindable directly despite the const-vs-mutable comparator ambiguity 
+    m.def("set_event_comparator", &set_event_comparator, py::arg("comparator"));
+    m.def("reset_event_comparator", &reset_event_comparator);
+
+    // ---- eventComparatorConfig ----
+    py::classh<eventComparatorConfig>(m, "EventComparatorConfig")
+        .def(py::init<>())
+        .def_readwrite("status_filter", &eventComparatorConfig::status_filter)
+        .def_readwrite("compare_momentum", &eventComparatorConfig::compare_momentum)
+        .def_readwrite("compare_momentum_x", &eventComparatorConfig::compare_momentum_x)
+        .def_readwrite("compare_momentum_y", &eventComparatorConfig::compare_momentum_y)
+        .def_readwrite("compare_momentum_z", &eventComparatorConfig::compare_momentum_z)
+        .def_readwrite("compare_momentum_E", &eventComparatorConfig::compare_momentum_E)
+        .def_readwrite("compare_mass", &eventComparatorConfig::compare_mass)
+        .def_readwrite("compare_vtim", &eventComparatorConfig::compare_vtim)
+        .def_readwrite("compare_spin", &eventComparatorConfig::compare_spin)
+        .def_readwrite("compare_pdg", &eventComparatorConfig::compare_pdg)
+        .def_readwrite("compare_status", &eventComparatorConfig::compare_status)
+        .def_readwrite("compare_mother", &eventComparatorConfig::compare_mother)
+        .def_readwrite("compare_icol", &eventComparatorConfig::compare_icol)
+        .def_readwrite("compare_n", &eventComparatorConfig::compare_n)
+        .def_readwrite("compare_proc_id", &eventComparatorConfig::compare_proc_id)
+        .def_readwrite("compare_weight", &eventComparatorConfig::compare_weight)
+        .def_readwrite("compare_scale", &eventComparatorConfig::compare_scale)
+        .def_readwrite("compare_alphaEW", &eventComparatorConfig::compare_alphaEW)
+        .def_readwrite("compare_alphaS", &eventComparatorConfig::compare_alphaS)
+        .def_readwrite("mass_tol", &eventComparatorConfig::mass_tol)
+        .def_readwrite("vtim_tol", &eventComparatorConfig::vtim_tol)
+        .def_readwrite("spin_tol", &eventComparatorConfig::spin_tol)
+        .def_readwrite("momentum_tol", &eventComparatorConfig::momentum_tol)
+        .def_readwrite("weight_tol", &eventComparatorConfig::weight_tol)
+        .def_readwrite("scale_tol", &eventComparatorConfig::scale_tol)
+        .def_readwrite("alphaEW_tol", &eventComparatorConfig::alphaEW_tol)
+        .def_readwrite("alphaS_tol", &eventComparatorConfig::alphaS_tol);
+    m.def("compare_legs_only", &compare_legs_only);
+    m.def("compare_final_state_only", &compare_final_state_only);
+    m.def("compare_physics_fields", &compare_physics_fields);
+
+    // ---- eventBelongs ----
+    // Only the mutable-event (event_equal_fn) comparator overload is bound,
+    // not its cevent_equal_fn twin: pybind11 can't distinguish a Python
+    // callable meant for one std::function signature from another of the
+    // same arity, so binding both would leave the second unreachable. The
+    // mutable path is also what belongs(event&)/__call__ actually use
+    py::classh<eventBelongs>(m, "EventBelongs")
+        .def(py::init<>())
+        .def(py::init<const event &>(), py::arg("event"))
+        .def(py::init<std::vector<event>>(), py::arg("events"))
+        .def_readwrite("events", &eventBelongs::events)
+        .def(
+            "add_event", py::overload_cast<const event &>(&eventBelongs::add_event)
+        )
+        .def(
+            "add_event",
+            py::overload_cast<const std::vector<event> &>(&eventBelongs::add_event)
+        )
+        .def(
+            "set_events", py::overload_cast<const event &>(&eventBelongs::set_events)
+        )
+        .def(
+            "set_events",
+            py::overload_cast<const std::vector<event> &>(&eventBelongs::set_events)
+        )
+        .def(
+            "set_comparator",
+            py::overload_cast<event_equal_fn>(&eventBelongs::set_comparator)
+        )
+        .def(
+            "set_comparator",
+            py::overload_cast<const eventComparatorConfig &>(&eventBelongs::set_comparator)
+        )
+        .def("belongs", py::overload_cast<event &>(&eventBelongs::belongs))
+        .def("__call__", py::overload_cast<event &>(&eventBelongs::operator()));
+    m.def("all_events_belong", &all_events_belong);
+    m.def("external_legs_comparator", &external_legs_comparator);
+    m.def("always_true", &always_true);
+
+    // ---- eventSorter ----
+    // Same mutable-vs-const ambiguity as eventBelongs applies to its
+    // event_bool_fn/cevent_bool_fn constructor twins; only event_bool_fn is
+    // bound. Unlike that documented omission, the single-vs-vector forms
+    // below are ALL bound in pairs (constructor, add_bool, set_event_sets,
+    // set_bools), matching add_event_set's existing single+vector pair
+    py::classh<eventSorter>(m, "EventSorter")
+        .def(py::init<>())
+        .def(py::init<const eventBelongs &>(), py::arg("event_set"))
+        .def(py::init<std::vector<eventBelongs>>(), py::arg("event_sets"))
+        .def(py::init<event_bool_fn>(), py::arg("comparator"))
+        .def(py::init<std::vector<event_bool_fn>>(), py::arg("comparators"))
+        .def(
+            "add_event_set",
+            py::overload_cast<const eventBelongs &>(&eventSorter::add_event_set)
+        )
+        .def(
+            "add_event_set",
+            py::overload_cast<const std::vector<eventBelongs> &>(&eventSorter::add_event_set)
+        )
+        .def("add_bool", py::overload_cast<event_bool_fn>(&eventSorter::add_bool))
+        .def(
+            "add_bool",
+            py::overload_cast<std::vector<event_bool_fn>>(&eventSorter::add_bool)
+        )
+        .def(
+            "set_event_sets",
+            py::overload_cast<const eventBelongs &>(&eventSorter::set_event_sets)
+        )
+        .def(
+            "set_event_sets",
+            py::overload_cast<const std::vector<eventBelongs> &>(&eventSorter::set_event_sets)
+        )
+        .def(
+            "set_bools", py::overload_cast<const event_bool_fn>(&eventSorter::set_bools)
+        )
+        .def(
+            "set_bools",
+            py::overload_cast<const std::vector<event_bool_fn>>(&eventSorter::set_bools)
+        )
+        .def("size", &eventSorter::size)
+        .def("position", py::overload_cast<event &>(&eventSorter::position))
+        .def(
+            "position",
+            py::overload_cast<std::vector<event> &>(&eventSorter::position)
+        )
+        .def("sort", py::overload_cast<std::vector<event> &>(&eventSorter::sort));
+    m.def(
+        "make_sample_sorter",
+        py::overload_cast<const std::vector<event> &, event_equal_fn>(&make_sample_sorter),
+        py::arg("sample"),
+        py::arg("comparator") = event_equal_fn(external_legs_comparator)
+    );
+    m.def(
+        "make_sample_sorter",
+        py::overload_cast<std::vector<std::shared_ptr<event>>, event_equal_fn>(
+            &make_sample_sorter
+        ),
+        py::arg("sample"),
+        py::arg("comparator") = event_equal_fn(external_legs_comparator)
+    );
+
+    // ---- process ----
+    auto process_cls =
+        py::classh<process>(m, "Process")
+            .def(
+                py::init<std::vector<std::shared_ptr<event>>, bool, bool>(),
+                py::arg("events"),
+                py::arg("filter_partons") = false,
+                py::arg("column_major") = false
+            )
+            .def_readwrite("n", &process::n_)
+            .def_readonly("n_summed", &process::n_summed)
+            .def_readwrite("proc_id", &process::proc_id_)
+            .def_readwrite("weight", &process::weight_)
+            .def_readwrite("scale", &process::scale_)
+            .def_readwrite("muF", &process::muF_)
+            .def_readwrite("muR", &process::muR_)
+            .def_readwrite("muPS", &process::muPS_)
+            .def_readwrite("alphaEW", &process::alphaEW_)
+            .def_readwrite("alphaS", &process::alphaS_)
+            .def_readwrite("mass", &process::mass_)
+            .def_readwrite("vtim", &process::vtim_)
+            .def_readwrite("spin", &process::spin_)
+            .def_readwrite("pdg", &process::pdg_)
+            .def_readwrite("status", &process::status_)
+            .def_readwrite("wgts", &process::wgts_)
+            .def_readwrite("helicity", &process::helicity_)
+            .def_readwrite("flavor", &process::flavor_)
+            .def_readwrite("column_major", &process::column_major)
+            .def_readwrite("filter", &process::filter)
+            .def_readonly("umami_momenta", &process::umami_momenta_)
+            .def_readonly("umami_mass", &process::umami_mass_)
+            .def_readonly("umami_vtim", &process::umami_vtim_)
+            .def_readonly("umami_spin", &process::umami_spin_)
+            .def_readonly("umami_pdg", &process::umami_pdg_)
+            .def_readonly("umami_status", &process::umami_status_)
+            .def_readonly("umami_mother", &process::umami_mother_)
+            .def_readonly("umami_icol", &process::umami_icol_)
+            .def_readwrite("events", &process::events)
+            .def("add_event", py::overload_cast<const event &>(&process::add_event))
+            .def(
+                "add_event",
+                py::overload_cast<const std::vector<event> &>(&process::add_event)
+            )
+            .def(
+                "add_event_raw", py::overload_cast<const event &>(&process::add_event_raw)
+            )
+            .def(
+                "add_event_filtered",
+                py::overload_cast<const event &>(&process::add_event_filtered)
+            )
+            .def(
+                "add_event_umami_filtered",
+                py::overload_cast<const std::vector<event> &>(
+                    &process::add_event_umami_filtered
+                )
+            )
+            .def("to_umami", &process::to_umami)
+            .def("from_umami", &process::from_umami)
+            .def("validate", &process::validate)
+            .def("transpose", &process::transpose)
+            // Unlike the other 25+ transpose_*() partials transpose_wgts()
+            // isn't just a redundant narrower transpose(): transpose()
+            // rebuilds process::events as brand-new Event objects, while
+            // transpose_wgts() writes wgts_ into the *existing* Event
+            // objects in place -- the ones a Reweightor's own .events (and
+            // an xSecAccumulator reading through it) still points to.
+            // Needed for callers driving Reweightor.run_streaming()'s loop
+            // by hand
+            .def("transpose_wgts", &process::transpose_wgts)
+            .def("size", &process::size)
+            .def("__len__", &process::size)
+            .def("gS", &process::gS)
+            .def("set_gS", &process::set_gS)
+            .def("get_muF", &process::get_muF)
+            .def("get_muR", &process::get_muR)
+            .def("get_muPS", &process::get_muPS)
+            .def("append_wgts", &process::append_wgts, py::arg("wgts"))
+            // Specific momentum components
+            .def("E", &process::E)
+            .def("x", &process::x)
+            .def("y", &process::y)
+            .def("z", &process::z)
+            .def("set_E", &process::set_E, py::arg("E"))
+            .def("set_x", &process::set_x, py::arg("x"))
+            .def("set_y", &process::set_y, py::arg("y"))
+            .def("set_z", &process::set_z, py::arg("z"));
+    bind_vecArrN_property(process_cls, "momenta", &process::momenta_);
+    bind_vecArrN_property(process_cls, "mother", &process::mother_);
+    bind_vecArrN_property(process_cls, "icol", &process::icol_);
+
+    // ---- initNode ----
+    auto initnode_cls =
+        py::classh<initNode>(m, "InitNode")
+            .def(py::init<>())
+            .def_readwrite("idWgt", &initNode::idWgt_)
+            .def_readwrite("nProc", &initNode::nProc_)
+            .def_readwrite("xSec", &initNode::xSec_)
+            .def_readwrite("xSecErr", &initNode::xSecErr_)
+            .def_readwrite("xMax", &initNode::xMax_)
+            .def_readwrite("lProc", &initNode::lProc_)
+            .def("validate_init", &initNode::validate_init);
+    bind_arrN_property(initnode_cls, "idBm", &initNode::idBm_);
+    bind_arrN_property(initnode_cls, "eBm", &initNode::eBm_);
+    bind_arrN_property(initnode_cls, "pdfG", &initNode::pdfG_);
+    bind_arrN_property(initnode_cls, "pdfS", &initNode::pdfS_);
+
+    // weight_norm's `verbose` parameter is left unbound (defaults to
+    // nullptr, ie no warning printed)
+    m.def(
+        "weight_norm",
+        [](const initNode &init, const std::vector<std::shared_ptr<event>> &evts) {
+            return weight_norm(init, evts);
+        },
+        py::arg("init"),
+        py::arg("events")
+    );
+
+    // ---- histogram::entry ----
+    py::classh<histogram::entry>(m, "HistogramEntry")
+        .def(py::init<>())
+        .def_readwrite("evt", &histogram::entry::evt)
+        .def_readwrite("index", &histogram::entry::index)
+        .def_readwrite("base_weight", &histogram::entry::base_weight);
+
+    // ---- histogram ----
+    // As with EventSorter above, only the explicit-indices/EventSorter/
+    // event_hash_fn overloads are bound for the constructors and
+    // add_event(s) -- the EventBelongs/event_bool_fn/cevent_bool_fn/vector
+    // forms in Rex.h are pure convenience wrappers that build an
+    // EventSorter internally, so nothing is lost by asking Python callers
+    // to do that EventSorter(...) wrapping themselves; skipping them also
+    // sidesteps the same event_bool_fn/cevent_bool_fn ambiguity noted for
+    // EventSorter's own bindings
+    //
+    // nominal/mult/n_wgts_synced/entries are derived/internal state --
+    // rebuilt by rebuild()/sync_wgts(), not meant to be hand-edited --
+    // so, like ProcReweightor.normalisation/proc above, they're exposed
+    // read-only. `norm` is likewise read-only: mutating it has the side
+    // effect of forcing a full re-sync (see the C++ comment on
+    // histogram::set_norm()), which a plain attribute write can't trigger;
+    // use set_norm()/normalise() instead. n_bins stays read-write, same as
+    // in C++, since setting it has no side effect until the next
+    // rebuild()/sync_wgts() call
+    py::classh<histogram>(m, "Histogram")
+        .def(py::init<>())
+        .def(
+            py::init<const std::vector<std::shared_ptr<event>> &, const std::vector<std::size_t> &>(),
+            py::arg("events"),
+            py::arg("indices")
+        )
+        .def(
+            py::init<const std::vector<std::shared_ptr<event>> &, eventSorter>(),
+            py::arg("events"),
+            py::arg("sorter")
+        )
+        .def(
+            py::init<const std::vector<std::shared_ptr<event>> &, event_hash_fn>(),
+            py::arg("events"),
+            py::arg("hash")
+        )
+        .def_readonly("entries", &histogram::entries)
+        .def_readonly("nominal", &histogram::nominal)
+        .def_readonly("mult", &histogram::mult)
+        .def_readonly("n_wgts_synced", &histogram::n_wgts_synced)
+        .def_readonly("norm", &histogram::norm)
+        .def_readwrite("n_bins", &histogram::n_bins)
+        .def(
+            "add_event",
+            py::overload_cast<std::shared_ptr<event>, std::size_t>(&histogram::add_event),
+            py::arg("event"),
+            py::arg("index")
+        )
+        .def(
+            "add_event",
+            py::overload_cast<std::shared_ptr<event>, eventSorter>(&histogram::add_event),
+            py::arg("event"),
+            py::arg("sorter")
+        )
+        .def(
+            "add_event",
+            py::overload_cast<std::shared_ptr<event>, event_hash_fn>(&histogram::add_event),
+            py::arg("event"),
+            py::arg("hash")
+        )
+        .def(
+            "add_events",
+            py::overload_cast<const std::vector<std::shared_ptr<event>> &, const std::vector<std::size_t> &>(
+                &histogram::add_events
+            ),
+            py::arg("events"),
+            py::arg("indices")
+        )
+        .def(
+            "add_events",
+            py::overload_cast<const std::vector<std::shared_ptr<event>> &, eventSorter>(&histogram::add_events),
+            py::arg("events"),
+            py::arg("sorter")
+        )
+        .def(
+            "add_events",
+            py::overload_cast<const std::vector<std::shared_ptr<event>> &, event_hash_fn>(&histogram::add_events),
+            py::arg("events"),
+            py::arg("hash")
+        )
+        .def("rebuild", &histogram::rebuild)
+        .def("normalise", &histogram::normalise, py::arg("init"))
+        .def("set_norm", &histogram::set_norm, py::arg("norm"))
+        .def("sync_wgts", &histogram::sync_wgts)
+        .def("size", &histogram::size)
+        .def("__len__", &histogram::size)
+        .def("bins", &histogram::bins)
+        .def("rows", &histogram::rows)
+        .def("counts", &histogram::counts)
+        .def(
+            "heights",
+            py::overload_cast<std::size_t>(&histogram::heights, py::const_),
+            py::arg("row")
+        )
+        .def("heights", py::overload_cast<>(&histogram::heights, py::const_));
+
+    // ---- lhe ----
+    // Real inheritance from InitNode (not a flattened mirror), so REX::tea's
+    // reweightor/param_rwgt (bound in tearex.cpp, and themselves derived from
+    // lhe in C++) can be registered as proper Python subclasses too.
+    // As with eventBelongs/eventSorter above, the vector<event>-by-value
+    // constructor is skipped in favour of the vector<shared_ptr<event>>
+    // form: a Python list of Event objects converts to either one, so
+    // binding both would leave one of them permanently shadowed
+    py::classh<lhe, initNode>(m, "Lhe")
+        .def(py::init<>())
+        .def(py::init<const initNode &>(), py::arg("init"))
+        .def(
+            py::init<std::vector<std::shared_ptr<event>>>(), py::arg("events")
+        )
+        .def(
+            py::init<const initNode &, std::vector<std::shared_ptr<event>>>(),
+            py::arg("init"),
+            py::arg("events")
+        )
+        .def_readwrite("filter_processes", &lhe::filter_processes)
+        .def_readwrite("column_major", &lhe::column_major)
+        .def_readwrite("max_batch", &lhe::max_batch)
+        .def_readwrite("events", &lhe::events)
+        .def_readwrite("processes", &lhe::processes)
+        .def_readonly("sorted_events", &lhe::sorted_events)
+        .def_readonly("process_order", &lhe::process_order)
+        .def_readonly("process_group", &lhe::process_group)
+        .def_property(
+            "weight_ids",
+            [](const lhe &self) {
+                return self.weight_ids ? *self.weight_ids : std::vector<std::string>{};
+            },
+            [](lhe &self, const std::vector<std::string> &ids) {
+                self.weight_ids = std::make_shared<std::vector<std::string>>(ids);
+            }
+        )
+        .def_readwrite("weight_context", &lhe::weight_context)
+        // ---- streaming (see open_lhe_source()/open_lhe_sink() below) ----
+        // `source` is read-only: it's only ever meant to be set via
+        // open_streaming(), which also eagerly reads init/header off of it --
+        // assigning one directly would leave those unset
+        .def_readonly("source", &lhe::source)
+        .def_readwrite("batch_size", &lhe::batch_size)
+        .def_readonly("events_pulled_total", &lhe::events_pulled_total)
+        .def_property_readonly("streaming", &lhe::streaming)
+        .def_static(
+            "open_streaming",
+            &lhe::open_streaming,
+            py::arg("source"),
+            py::arg("batch_size") = 10000
+        )
+        .def("pull_batch", &lhe::pull_batch)
+        .def("add_event", py::overload_cast<std::shared_ptr<event>>(&lhe::add_event))
+        .def("add_event", py::overload_cast<const event &>(&lhe::add_event))
+        .def("set_sorter", py::overload_cast<>(&lhe::set_sorter))
+        .def("set_sorter", py::overload_cast<const eventSorter &>(&lhe::set_sorter))
+        .def("set_sorter", py::overload_cast<event_equal_fn>(&lhe::set_sorter))
+        .def("sort_events", &lhe::sort_events)
+        .def("unsort_events", &lhe::unsort_events)
+        .def("events_to_processes", &lhe::events_to_processes)
+        .def("processes_to_events", &lhe::processes_to_events)
+        .def("transpose", py::overload_cast<>(&lhe::transpose))
+        .def(
+            "transpose",
+            py::overload_cast<std::string>(&lhe::transpose),
+            py::arg("direction")
+        )
+        .def("set_filter", &lhe::set_filter)
+        .def("set_max_batch", &lhe::set_max_batch)
+        .def(
+            "add_weight_id",
+            py::overload_cast<const std::string &>(&lhe::add_weight_id),
+            py::arg("id")
+        )
+        .def("extract_weight_ids", &lhe::extract_weight_ids)
+        .def("sync_weight_ids", &lhe::sync_weight_ids)
+        .def(
+            "append_weight_ids", &lhe::append_weight_ids, py::arg("include") = false
+        )
+        // ---- histogram management ----
+        // Owned histograms, built against self.events. As with add_event
+        // above, add_histogram is only bound in its shared_ptr<Histogram>
+        // form -- a Python Histogram converts to that transparently via
+        // classh, so the Histogram-by-value overload would just be
+        // permanently shadowed. make_histogram mirrors the same
+        // explicit-indices/EventSorter/event_hash_fn subset Histogram
+        // itself binds above, for the same reason
+        .def_readonly("histograms", &lhe::histograms)
+        .def(
+            "add_histogram",
+            py::overload_cast<const std::string &, std::shared_ptr<histogram>>(&lhe::add_histogram),
+            py::arg("name"),
+            py::arg("histogram")
+        )
+        .def(
+            "make_histogram",
+            py::overload_cast<const std::string &, const std::vector<std::size_t> &>(&lhe::make_histogram),
+            py::arg("name"),
+            py::arg("indices")
+        )
+        .def(
+            "make_histogram",
+            py::overload_cast<const std::string &, eventSorter>(&lhe::make_histogram),
+            py::arg("name"),
+            py::arg("sorter")
+        )
+        .def(
+            "make_histogram",
+            py::overload_cast<const std::string &, event_hash_fn>(&lhe::make_histogram),
+            py::arg("name"),
+            py::arg("hash")
+        )
+        .def("has_histogram", &lhe::has_histogram, py::arg("name"))
+        // Bound as a lambda returning the map's own shared_ptr<histogram>
+        // rather than directly to lhe::hist() (which returns a plain
+        // histogram&): a bare C++ reference return only resolves back to
+        // the same live Python object if that particular instance has
+        // already been seen through a shared_ptr/holder pathway (eg via
+        // .histograms[name] below) -- the very first access otherwise
+        // silently hands back a disconnected copy, so mutating it (or
+        // syncing it later via sync_histogram()) wouldn't touch the
+        // histogram actually stored on this Lhe. Returning the shared_ptr
+        // directly, exactly as .histograms[name] does, sidesteps that
+        // regardless of access order
+        .def(
+            "hist",
+            [](lhe &self, const std::string &name) -> std::shared_ptr<histogram> {
+                auto it = self.histograms.find(name);
+                if (it == self.histograms.end())
+                    throw std::out_of_range("Lhe.hist: no histogram named '" + name + "'");
+                return it->second;
+            },
+            py::arg("name")
+        )
+        .def("remove_histogram", &lhe::remove_histogram, py::arg("name"))
+        .def("sync_histograms", &lhe::sync_histograms)
+        .def("sync_histogram", &lhe::sync_histogram, py::arg("name"))
+        // In-memory string round-trip, mirroring Slha.write() below: only
+        // write_lhef(lhe&, ostream&, bool)/write_lhef(lhe&, filename, bool)
+        // exist in C++, so file-path I/O was the only Python-reachable
+        // form; wrapping the ostream overload in an ostringstream gives a
+        // string result without needing to touch disk
+        .def(
+            "to_string",
+            [](lhe &self, bool include_ids) {
+                std::ostringstream ss;
+                write_lhef(self, ss, include_ids);
+                return ss.str();
+            },
+            py::arg("include_ids") = false
+        );
+    m.def(
+        "load_lhef",
+        py::overload_cast<const std::string &>(&load_lhef),
+        py::arg("filename")
+    );
+    m.def(
+        "write_lhef",
+        py::overload_cast<lhe &, const std::string &, bool>(&write_lhef),
+        py::arg("doc"),
+        py::arg("filename"),
+        py::arg("include_ids") = false
+    );
+    // to_lhe(const string&) parses an XML string directly into an Lhe --
+    // the in-memory counterpart to load_lhef(filename); the
+    // shared_ptr<xmlNode>-taking overload stays unbound (xmlNode itself
+    // isn't exposed to Python)
+    m.def("to_lhe", py::overload_cast<const std::string &>(&to_lhe), py::arg("xml"));
+
+    // ---- eventSource / eventSink (streaming) ----
+    // Both are otherwise-opaque handles from Python: their std::function
+    // members exist to be invoked internally by Lhe.pull_batch()/
+    // Reweightor.run_streaming(), not called directly. Concrete instances
+    // come from open_lhe_source()/open_lhe_sink() below -- the only
+    // constructors bound, matching how xmlNode/xmlDoc stay unexposed and
+    // load_lhef()/write_lhef() are the practical entry points for the
+    // non-streaming path. Only the filename-taking open_lhe_source()
+    // overload is bound; the std::istream& overload has no Python-side
+    // counterpart to pass in
+    py::classh<eventSource>(m, "EventSource")
+        .def_property_readonly("seekable", &eventSource::seekable);
+    py::classh<eventSink>(m, "EventSink");
+    m.def(
+        "open_lhe_source",
+        py::overload_cast<const std::string &>(&open_lhe_source),
+        py::arg("filename")
+    );
+    m.def(
+        "open_lhe_sink", &open_lhe_sink, py::arg("filename"), py::arg("include_ids") = false
+    );
+
+    // ---- slha ----
+    // Only the single-index get/set overloads are bound: the multi-index
+    // overloads take std::initializer_list<int>, which pybind11 has no
+    // caster for (it can't be built from a runtime-sized Python list), and
+    // has_entry only exists in that form, so it has no bindable overload at
+    // all
+    py::classh<slha>(m, "Slha")
+        .def(py::init<>())
+        .def(
+            "get",
+            py::overload_cast<const std::string &, int, double>(&slha::get, py::const_),
+            py::arg("block"),
+            py::arg("index"),
+            py::arg("fallback") = 0.0
+        )
+        .def(
+            "set",
+            py::overload_cast<const std::string &, int, double>(&slha::set),
+            py::arg("block"),
+            py::arg("index"),
+            py::arg("value")
+        )
+        .def(
+            "get_decay", &slha::get_decay, py::arg("pid"), py::arg("fallback") = 0.0
+        )
+        .def("set_decay", &slha::set_decay, py::arg("pid"), py::arg("width"))
+        .def("has_block", &slha::has_block, py::arg("block"))
+        .def(
+            "write",
+            [](const slha &self, int value_precision, bool scientific, const std::string &indent) {
+                std::ostringstream ss;
+                self.write(ss, value_precision, scientific, indent);
+                return ss.str();
+            },
+            py::arg("value_precision") = 6,
+            py::arg("scientific") = true,
+            py::arg("indent") = "      "
+        );
+    m.def(
+        "load_slha", py::overload_cast<const std::string &>(&load_slha), py::arg("filename")
+    );
+    m.def("to_slha", py::overload_cast<const std::string &>(&to_slha), py::arg("text"));
+    // Extracts an already-loaded Lhe doc's SLHA param card (eg from its
+    // <slha> header block), distinct from to_slha(text) above which parses
+    // a raw SLHA-format string
+    m.def("to_slha", py::overload_cast<const lhe &>(&to_slha), py::arg("doc"));
+
+    bind_tearex(m.def_submodule("tea"));
+}
diff --git a/madtrex/Rex/src/python/tearex.cpp b/madtrex/Rex/src/python/tearex.cpp
new file mode 100644
index 0000000000..9a410455a9
--- /dev/null
+++ b/madtrex/Rex/src/python/tearex.cpp
@@ -0,0 +1,461 @@
+#include "tearex.hpp"
+
+#include <pybind11/functional.h>
+#include <pybind11/pybind11.h>
+#include <pybind11/stl.h>
+
+#include "teaRex.h"
+
+namespace py = pybind11;
+using namespace REX;
+using namespace REX::tea;
+
+namespace {
+
+// weightor is std::function<shared_ptr<vector<double>>(process&)>. pybind11's
+// functional.h caster can't handle that return type directly: its
+// shared_ptr<T> caster assumes T is a pybind11-registered class (for which it
+// has a holder-based caster), not an STL container like vector<double> (which
+// is instead handled by stl.h's *non*-holder vector caster) -- so
+// shared_ptr<vector<double>> falls through both and fails to compile. Route
+// Python callables through this plain-vector-returning signature instead,
+// which pybind11 can cast natively, and adapt to the real weightor here in
+// hand-written glue rather than via automatic function casting
+using py_weightor = std::function<std::vector<double>(process &)>;
+
+weightor to_weightor(py_weightor f) {
+    return [f = std::move(f)](process &p) -> std::shared_ptr<std::vector<double>> {
+        return std::make_shared<std::vector<double>>(f(p));
+    };
+}
+
+std::vector<weightor> to_weightors(std::vector<py_weightor> fs) {
+    std::vector<weightor> out;
+    out.reserve(fs.size());
+    for (auto &f : fs) out.push_back(to_weightor(std::move(f)));
+    return out;
+}
+
+// Same rationale as py_weightor above, one level further removed: a
+// deferred_weightor is std::function<weight_puller(process&)>, and
+// weight_puller is itself std::function<shared_ptr<vector<double>>()> --
+// two holder-vs-non-holder shared_ptr<vector<double>> casters that can't be
+// bound directly. Python callables instead produce a plain vector<double>
+// puller (py_weight_puller), and the outer Python callable returns *that*
+// rather than a real weight_puller; both layers get adapted to the real
+// deferred_weightor/weight_puller here
+using py_weight_puller = std::function<std::vector<double>()>;
+using py_deferred_weightor = std::function<py_weight_puller(process &)>;
+
+deferred_weightor to_deferred_weightor(py_deferred_weightor f) {
+    return [f = std::move(f)](process &p) -> weight_puller {
+        py_weight_puller puller = f(p);
+        return [puller = std::move(puller)]() -> std::shared_ptr<std::vector<double>> {
+            return std::make_shared<std::vector<double>>(puller());
+        };
+    };
+}
+
+std::vector<deferred_weightor> to_deferred_weightors(std::vector<py_deferred_weightor> fs) {
+    std::vector<deferred_weightor> out;
+    out.reserve(fs.size());
+    for (auto &f : fs) out.push_back(to_deferred_weightor(std::move(f)));
+    return out;
+}
+
+} // namespace
+
+void bind_tearex(py::module_ m) {
+    m.doc() = "REX::tea (teaRex) generic event reweighting bindings";
+
+    m.def("true_function", &true_function);
+    m.def("helicity_selector", &helicity_selector, py::arg("helicity"));
+    m.def("flavor_selector", &flavor_selector, py::arg("flavor"));
+    m.def(
+        "helicity_flavor_selector",
+        &helicity_flavor_selector,
+        py::arg("helicity"),
+        py::arg("flavor")
+    );
+
+    // ---- procReweightor ----
+    // Constructors taking an event selector are bound only in their
+    // shared_ptr<eventBelongs> form, not the sibling eventBelongs-by-value
+    // overloads: a Python EventBelongs object converts to shared_ptr<...>
+    // transparently via classh either way, and keeping only one form avoids
+    // pybind11 having two simultaneously-viable overloads for the same call
+    //
+    // The deferred_weightor-taking constructors are NOT bound as further
+    // py::init(...) overloads alongside these: pybind11's std::function
+    // caster accepts any Python callable regardless of what it returns, so
+    // a single-callable or list-of-callables argument is equally "valid"
+    // for both a py_weightor- and a py_deferred_weightor-shaped overload --
+    // pybind11 can't tell them apart the way it can eg a callable from a
+    // list, and the first-registered (eager) overload would silently
+    // shadow the deferred one for every call. They're bound as separate
+    // ProcReweightor.from_deferred(...) static factories below instead,
+    // which sidesteps the collision entirely since they're never in the
+    // same overload-resolution pool as these
+    py::classh<procReweightor>(m, "ProcReweightor")
+        .def(py::init<>())
+        .def(
+            py::init([](py_weightor f) { return procReweightor(to_weightor(std::move(f))); }),
+            py::arg("reweight_function")
+        )
+        .def(
+            py::init([](py_weightor f, std::shared_ptr<eventBelongs> sel) {
+                return procReweightor(to_weightor(std::move(f)), sel);
+            }),
+            py::arg("reweight_function"),
+            py::arg("selector")
+        )
+        .def(
+            py::init([](std::vector<py_weightor> fs) {
+                return procReweightor(to_weightors(std::move(fs)));
+            }),
+            py::arg("reweight_functions")
+        )
+        .def(
+            py::init([](std::vector<py_weightor> fs, std::shared_ptr<eventBelongs> sel) {
+                return procReweightor(to_weightors(std::move(fs)), sel);
+            }),
+            py::arg("reweight_functions"),
+            py::arg("selector")
+        )
+        .def(
+            py::init([](std::vector<py_weightor> fs,
+                        std::shared_ptr<eventBelongs> sel,
+                        py_weightor norm) {
+                return procReweightor(
+                    to_weightors(std::move(fs)), sel, to_weightor(std::move(norm))
+                );
+            }),
+            py::arg("reweight_functions"),
+            py::arg("selector"),
+            py::arg("normaliser")
+        )
+        // ---- deferred_weightor construction (see comment above) ----
+        .def_static(
+            "from_deferred",
+            [](py_deferred_weightor f) {
+                return procReweightor(to_deferred_weightor(std::move(f)));
+            },
+            py::arg("reweight_function")
+        )
+        .def_static(
+            "from_deferred",
+            [](py_deferred_weightor f, std::shared_ptr<eventBelongs> sel) {
+                return procReweightor(to_deferred_weightor(std::move(f)), sel);
+            },
+            py::arg("reweight_function"),
+            py::arg("selector")
+        )
+        .def_static(
+            "from_deferred",
+            [](std::vector<py_deferred_weightor> fs) {
+                return procReweightor(to_deferred_weightors(std::move(fs)));
+            },
+            py::arg("reweight_functions")
+        )
+        .def_static(
+            "from_deferred",
+            [](std::vector<py_deferred_weightor> fs, std::shared_ptr<eventBelongs> sel) {
+                return procReweightor(to_deferred_weightors(std::move(fs)), sel);
+            },
+            py::arg("reweight_functions"),
+            py::arg("selector")
+        )
+        .def_static(
+            "from_deferred",
+            [](std::vector<py_deferred_weightor> fs,
+               std::shared_ptr<eventBelongs> sel,
+               py_weightor norm) {
+                return procReweightor(
+                    to_deferred_weightors(std::move(fs)), sel, to_weightor(std::move(norm))
+                );
+            },
+            py::arg("reweight_functions"),
+            py::arg("selector"),
+            py::arg("normaliser")
+        )
+        .def_readonly("normalisation", &procReweightor::normalisation)
+        .def_readonly("proc", &procReweightor::proc)
+        // procReweightor::verbose is a REX::verbosity, not a plain bool --
+        // exposed as one anyway since that's all Python callers need; a
+        // procReweightor owned by a Reweightor has its verbosity linked to
+        // its owner's (see Reweightor.verbose below), so muting the latter
+        // mutes this one too, but it can still be read/overridden directly
+        // here (eg for a procReweightor used standalone)
+        .def_property(
+            "verbose",
+            [](const procReweightor &self) { return bool(self.verbose); },
+            [](procReweightor &self, bool v) { self.verbose = v; }
+        )
+        // procReweightor::backlog is now std::vector<backlog_entry>, a
+        // struct holding a raw weight_puller (a C++ closure, not itself
+        // pybind11-registered) plus the immediate flag -- no longer the
+        // plain std::vector<std::vector<double>> stl.h could convert
+        // automatically. Exposing its length is enough to let callers
+        // check whether anything is pending without reaching into C++
+        // closures from Python
+        .def_property_readonly(
+            "backlog_size", [](const procReweightor &self) { return self.backlog.size(); }
+        )
+        .def(
+            "set_event_checker",
+            py::overload_cast<eventBelongs>(&procReweightor::set_event_checker)
+        )
+        .def(
+            "set_event_checker",
+            py::overload_cast<REX::event_bool_fn>(&procReweightor::set_event_checker)
+        )
+        .def(
+            "set_normaliser",
+            [](procReweightor &self, py_weightor f) -> procReweightor & {
+                return self.set_normaliser(to_weightor(std::move(f)));
+            },
+            py::arg("normaliser")
+        )
+        // ---- normalisation access besides evaluating a normaliser ----
+        // calc_normalisation derives the normalisation array from a vector
+        // of target weights (relative to proc's original weight_);
+        // set_initial_weights does the same but also marks the
+        // normalisation as explicitly set, so initialise() won't overwrite
+        // it later; set_normalisation instead takes the normalisation array
+        // itself, verbatim
+        .def(
+            "calc_normalisation", &procReweightor::calc_normalisation, py::arg("wgts")
+        )
+        .def(
+            "set_initial_weights", &procReweightor::set_initial_weights, py::arg("weights")
+        )
+        .def(
+            "set_normalisation", &procReweightor::set_normalisation, py::arg("normalisation")
+        )
+        .def(
+            "set_reweight_functions",
+            [](procReweightor &self, py_weightor f) -> procReweightor & {
+                return self.set_reweight_functions(to_weightor(std::move(f)));
+            }
+        )
+        .def(
+            "set_reweight_functions",
+            [](procReweightor &self, std::vector<py_weightor> fs) -> procReweightor & {
+                return self.set_reweight_functions(to_weightors(std::move(fs)));
+            }
+        )
+        .def(
+            "add_reweight_function",
+            [](procReweightor &self, py_weightor f) -> procReweightor & {
+                return self.add_reweight_function(to_weightor(std::move(f)));
+            },
+            py::arg("rwgt")
+        )
+        .def(
+            "set_deferred_reweight_functions",
+            [](procReweightor &self, py_deferred_weightor f) -> procReweightor & {
+                return self.set_deferred_reweight_functions(to_deferred_weightor(std::move(f)));
+            }
+        )
+        .def(
+            "set_deferred_reweight_functions",
+            [](procReweightor &self, std::vector<py_deferred_weightor> fs) -> procReweightor & {
+                return self.set_deferred_reweight_functions(to_deferred_weightors(std::move(fs))
+                );
+            }
+        )
+        .def(
+            "add_deferred_reweight_function",
+            [](procReweightor &self, py_deferred_weightor f) -> procReweightor & {
+                return self.add_deferred_reweight_function(to_deferred_weightor(std::move(f)));
+            },
+            py::arg("rwgt")
+        )
+        .def("set_process", &procReweightor::set_process, py::arg("process"))
+        .def("initialise", py::overload_cast<>(&procReweightor::initialise))
+        .def(
+            "initialise",
+            py::overload_cast<std::shared_ptr<process>>(&procReweightor::initialise),
+            py::arg("process")
+        )
+        .def("evaluate", py::overload_cast<>(&procReweightor::evaluate))
+        .def(
+            "evaluate", py::overload_cast<std::size_t>(&procReweightor::evaluate), py::arg("amp")
+        )
+        .def("append_zero_weights", &procReweightor::append_zero_weights)
+        .def("append_backlog", &procReweightor::append_backlog)
+        .def("pull_deferred_backlog", &procReweightor::pull_deferred_backlog);
+
+    // ---- xSecAccumulator ----
+    // Running-reduction accumulator standing in for
+    // calc_xSec_from_weights()/calc_norm()/calc_xSecs()/calc_xErrs() when
+    // driving Reweightor.run_streaming()'s per-batch loop by hand instead of
+    // calling run_streaming() itself -- absorb() folds one finished batch's
+    // events into a handful of running scalars, finalise_into() reproduces
+    // the same xSec/xSecErr/rwgt_xSec/rwgt_xErr a full run_streaming() call
+    // would set on the given Reweightor. See run_streaming's own reference
+    // sequence in teaRex.cc: pull_batch() -> extract_sorter_stream() ->
+    // initialise_reweightors() -> run_all_iterations() ->
+    // pull_deferred_backlog() per procReweightor -> absorb(events) -> write
+    // each event -> repeat per batch -> finalise_into(reweightor) once, only
+    // after every batch has been absorbed
+    py::classh<xSecAccumulator>(m, "XSecAccumulator")
+        .def(py::init<>())
+        .def_readonly("n_events", &xSecAccumulator::n_events)
+        .def_readonly("weight_sum", &xSecAccumulator::weight_sum)
+        .def_readonly("weight_sumSq", &xSecAccumulator::weight_sumSq)
+        .def_readonly("wgt_sum", &xSecAccumulator::wgt_sum)
+        .def_readonly("ratio_sum", &xSecAccumulator::ratio_sum)
+        .def_readonly("ratio_sumSq", &xSecAccumulator::ratio_sumSq)
+        .def("absorb", &xSecAccumulator::absorb, py::arg("events"))
+        .def("finalise_into", &xSecAccumulator::finalise_into, py::arg("reweightor"));
+
+    // ---- reweightor ----
+    // Bound as a real subclass of Lhe (registered in rex.cpp, which runs
+    // before this function is called). Constructor overloads taking
+    // vector<procReweightor> by value are skipped in favour of the
+    // shared_ptr<procReweightor> forms, same rationale as above; the
+    // lhe&&-taking (move) overloads are skipped since Python has no rvalue
+    // distinction -- the const lhe& forms cover every practical call
+    py::classh<reweightor, lhe>(m, "Reweightor")
+        .def(py::init<>())
+        .def(py::init<const lhe &>(), py::arg("mother"))
+        .def(
+            py::init<
+                const lhe &,
+                std::vector<std::shared_ptr<procReweightor>>,
+                std::vector<iterator>>(),
+            py::arg("mother"),
+            py::arg("reweightors"),
+            // Must match reweightor::iterators's own default member
+            // initializer ({true_function}), not an empty vector: passing an
+            // explicit empty vector here would override that default and
+            // leave run_all_iterations()'s `curr_iter < iterators.size()`
+            // loop with nothing to iterate over
+            py::arg("iterators") = std::vector<iterator>{true_function}
+        )
+        .def_readwrite("reweightors", &reweightor::reweightors)
+        .def_readwrite("launch_names", &reweightor::launch_names)
+        .def_readwrite("pool_threads", &reweightor::pool_threads)
+        // Controls whether non-forced diagnostic output is emitted (eg the
+        // per-iteration "." progress marker); forced warnings still get
+        // through regardless
+        .def_property(
+            "verbose",
+            [](const reweightor &self) { return bool(self.verbose); },
+            [](reweightor &self, bool v) { self.verbose = v; }
+        )
+        .def_readonly("norm_factor", &reweightor::norm_factor)
+        .def_readonly("rwgt_xSec", &reweightor::rwgt_xSec)
+        .def_readonly("rwgt_xErr", &reweightor::rwgt_xErr)
+        // curr_iter/n_amps are read-write: setup()/run_streaming() maintain
+        // them automatically for run()/run_streaming(), but a caller driving
+        // run_iteration()/run_all_iterations() by hand per streamed batch
+        // (see the manual-loop half of streaming_demo() in showcase.py)
+        // needs to reset curr_iter to 0 and resync n_amps (via
+        // sync_n_amps() below) itself once per batch
+        .def_readwrite("curr_iter", &reweightor::curr_iter)
+        .def_readwrite("n_amps", &reweightor::n_amps)
+        .def(
+            "set_reweightors",
+            py::overload_cast<std::vector<std::shared_ptr<procReweightor>>>(
+                &reweightor::set_reweightors
+            )
+        )
+        .def(
+            "add_reweightor",
+            py::overload_cast<std::shared_ptr<procReweightor>>(&reweightor::add_reweightor)
+        )
+        .def("set_initialise", &reweightor::set_initialise, py::arg("init"))
+        .def("set_finalise", &reweightor::set_finalise, py::arg("fin"))
+        .def("set_iterators", &reweightor::set_iterators, py::arg("iterators"))
+        .def(
+            "add_iterator",
+            [](reweightor &self, iterator it) -> reweightor & {
+                return self.add_iterator(it);
+            },
+            py::arg("iterator")
+        )
+        .def("set_launch_names", &reweightor::set_launch_names, py::arg("names"))
+        .def("add_launch_name", &reweightor::add_launch_name, py::arg("name"))
+        .def("extract_sorter", &reweightor::extract_sorter)
+        // Per-batch counterpart to extract_sorter(), used by run_streaming()
+        // (and available here for callers driving the streaming loop by
+        // hand): `base` is the procReweightor list as registered before
+        // streaming began (eg self.reweightors, saved off beforehand) --
+        // unlike extract_sorter(), this never rebinds into `base` itself, so
+        // the same list stays a valid template call after call, once per
+        // batch
+        .def(
+            "extract_sorter_stream",
+            &reweightor::extract_sorter_stream,
+            py::arg("base")
+        )
+        .def("initialise_reweightors", &reweightor::initialise_reweightors)
+        .def("finalise_reweighting", &reweightor::finalise_reweighting)
+        .def("sync_n_amps", &reweightor::sync_n_amps)
+        .def("setup", &reweightor::setup)
+        // run_iteration()/run_all_iterations()/run() dispatch procReweightor
+        // evaluation onto REX::tea::threadPool's worker threads and then
+        // block the calling thread in pool->wait_batch() until they finish.
+        // Since Python-defined weightor/event_bool_fn callables need the GIL
+        // to run (pybind11's functional.h caster acquires it automatically
+        // before calling in), leaving the GIL held on the blocked calling
+        // thread would deadlock every worker against it. Release it for the
+        // duration of the call so the workers can actually acquire it
+        .def("run_iteration", &reweightor::run_iteration, py::call_guard<py::gil_scoped_release>())
+        .def(
+            "run_all_iterations",
+            &reweightor::run_all_iterations,
+            py::call_guard<py::gil_scoped_release>()
+        )
+        .def("run", &reweightor::run, py::call_guard<py::gil_scoped_release>())
+        // Streaming counterpart to run(): repeatedly pulls a batch from
+        // self's source (see Lhe.open_streaming()), reweights it, writes it
+        // to sink, and frees it before pulling the next -- see run_streaming's
+        // C++ doc comment in teaRex.h for the full per-batch sequence. Same
+        // GIL-release rationale as run_iteration()/run_all_iterations()/run()
+        .def(
+            "run_streaming",
+            &reweightor::run_streaming,
+            py::arg("sink"),
+            py::arg("include_ids") = false,
+            py::call_guard<py::gil_scoped_release>()
+        )
+        .def("calc_norm", &reweightor::calc_norm)
+        .def("set_norm", &reweightor::set_norm, py::arg("norm"))
+        .def("calc_xSecs", &reweightor::calc_xSecs)
+        .def("calc_xErrs", &reweightor::calc_xErrs);
+
+    // ---- param_rwgt ----
+    // rwgt_slha/rwgt_card/rwgt_block stay unbound for now (mirrors slha's own
+    // scope trim) -- read_slha_rwgt(slha_file, rwgt_file) is the practical
+    // entry point for the SLHA-reweight-card workflow
+    py::classh<param_rwgt, reweightor>(m, "ParamRwgt")
+        .def(py::init<>())
+        .def(py::init<const lhe &>(), py::arg("mother"))
+        .def(
+            py::init<const lhe &, std::vector<std::shared_ptr<procReweightor>>>(),
+            py::arg("mother"),
+            py::arg("reweightors")
+        )
+        .def(
+            py::init<
+                const lhe &,
+                std::vector<std::shared_ptr<procReweightor>>,
+                const std::string &,
+                const std::string &>(),
+            py::arg("mother"),
+            py::arg("reweightors"),
+            py::arg("slha_path"),
+            py::arg("rwgt_path")
+        )
+        .def(
+            "read_slha_rwgt",
+            py::overload_cast<const std::string &, const std::string &>(
+                &param_rwgt::read_slha_rwgt
+            ),
+            py::arg("slha_file"),
+            py::arg("rwgt_file")
+        );
+}
diff --git a/madtrex/Rex/src/python/tearex.hpp b/madtrex/Rex/src/python/tearex.hpp
new file mode 100644
index 0000000000..bb74ee41fc
--- /dev/null
+++ b/madtrex/Rex/src/python/tearex.hpp
@@ -0,0 +1,7 @@
+#pragma once
+
+#include <pybind11/pybind11.h>
+
+// Binds the REX::tea (teaRex) types into the given module. Split into its own
+// translation unit purely to mirror the Rex/teaRex source split
+void bind_tearex(pybind11::module_ m);
diff --git a/madtrex/Rex/src/teaRex.cc b/madtrex/Rex/src/teaRex.cc
new file mode 100644
index 0000000000..fc530ad75d
--- /dev/null
+++ b/madtrex/Rex/src/teaRex.cc
@@ -0,0 +1,1536 @@
+/***
+ *     _            ______
+ *    | |           | ___ \
+ *    | |_ ___  __ _| |_/ /_____  __
+ *    | __/ _ \/ _` |    // _ \ \/ /
+ *    | ||  __/ (_| | |\ \  __/>  <
+ *     \__\___|\__,_\_| \_\___/_/\_\
+ *
+ ***/
+//
+// *t*ensorial *e*vent *a*daption with *R*e*x* Version 1.1.0
+// teaRex is an extension to the Rex library for the generic reweighting of parton-level events.
+// It provides a flexible framework for applying weight modifications to events based on user-defined criteria,
+// using the underlying Rex formats to sort, extract, and rewrite event-level information,
+// and extending it to allow for generic reweighting using any information stored in an LHE file as input for a
+// user-provided reweighting function acting on REX::process objects, which are SoA (Structure of Arrays)
+// objects for storing event information. Users can either provide the REX::process objects themselves,
+// or use the flexible Rex sorting architecture to extract the necessary information from an LHE file.
+//
+// Copyright © 2023-2025 CERN, CERN Author Zenny Wettersten.
+// Copyright © 2025-2026 Zenny Wettersten.
+// Licensed under the GNU Lesser General Public License (version 3 or later).
+// All rights not expressly granted are reserved.
+//
+
+#ifndef _TEAREX_CPP_
+#define _TEAREX_CPP_
+
+#include "teaRex.h"
+
+namespace REX::tea
+{
+
+    bool true_function()
+    {
+        return true;
+    }
+
+    event_bool_fn helicity_selector(int helicity)
+    {
+        return [helicity](REX::event &e) -> bool
+        { return helicity < 0 || e.helicity_ == helicity; };
+    }
+
+    event_bool_fn flavor_selector(int flavor)
+    {
+        return [flavor](REX::event &e) -> bool
+        { return flavor < 0 || e.flavor_ == flavor; };
+    }
+
+    event_bool_fn helicity_flavor_selector(int helicity, int flavor)
+    {
+        return [helicity, flavor](REX::event &e) -> bool
+        { return (helicity < 0 || e.helicity_ == helicity) && (flavor < 0 || e.flavor_ == flavor); };
+    }
+
+    void rwgt_slha::rwgt_card::add_param(const std::string &block_name, std::pair<int, double> param)
+    {
+        if (this->blocks.find(block_name) == this->blocks.end())
+            this->blocks[block_name] = rwgt_block{block_name, {}};
+        this->blocks[block_name].params.push_back(param);
+    }
+
+    void rwgt_slha::rwgt_card::add_param(const std::string &block_name, int param_id, double param_value)
+    {
+        this->add_param(block_name, std::make_pair(param_id, param_value));
+    }
+
+    void rwgt_slha::rwgt_card::add_param(const std::string &block_name, std::vector<int> param_ids, double param_value)
+    {
+        if (param_ids.size() == 1)
+            return this->add_param(block_name, param_ids[0], param_value);
+        if (this->blocks.find(block_name) == this->blocks.end())
+            this->blocks[block_name] = rwgt_block{block_name, {}};
+        this->blocks[block_name].multi_params.emplace_back(std::move(param_ids), param_value);
+    }
+
+    void rwgt_slha::parse_rwgt_card(std::istream &is)
+    {
+        this->cards.clear();
+        size_t curr_launch = 0;
+        this->cards.emplace_back();
+        std::string line;
+        bool first_launch = false;
+        while (std::getline(is, line))
+        {
+            if (line.empty() || line[0] == '#')
+                continue; // Skip empty lines and comments
+            std::string iss = line.c_str();
+            if (iss.find("launch") != std::string::npos)
+            {
+                if (!first_launch)
+                {
+                    first_launch = true;
+                }
+                else
+                {
+                    this->cards.emplace_back();
+                    curr_launch++;
+                }
+                size_t name_pos = iss.find("rwgt_name");
+                if (name_pos != std::string::npos)
+                {
+                    std::string curr_name = iss.substr(name_pos + 10);
+                    auto name_lines = REX::blank_splitter(curr_name);
+                    this->cards.back().launch_name = std::string(name_lines[0]);
+                }
+                continue;
+            }
+            if (iss.find("set") != std::string::npos)
+            {
+                auto rwgt_line = REX::blank_splitter(iss);
+                if (rwgt_line[0] != "set")
+                    continue;
+                if (rwgt_line.size() < 3)
+                    continue;
+                if (rwgt_line.size() == 3)
+                    throw std::runtime_error("rwgt_slha::parse_rwgt_card: \"set\" line appears to use parameter names. This is not supported by teaRex.");
+                if (!first_launch)
+                {
+                    REX::warning("rwgt_slha::parse_rwgt_card: \"set\" line appears before first launch command. Assuming launch command missed. May end up appending meaningless reweighting iterations.");
+                    first_launch = true;
+                }
+                size_t curr_ind_of_line = 1;
+                if (rwgt_line[curr_ind_of_line] == "param_card")
+                    ++curr_ind_of_line;
+                std::string curr_block = std::string(rwgt_line[curr_ind_of_line]);
+                ++curr_ind_of_line;
+                // one index, or several for a multi-index block ("set NMIX 1 2
+                // 0.5"): every word between the block and the value
+                if (curr_ind_of_line + 1 >= rwgt_line.size())
+                    throw std::runtime_error("rwgt_slha::parse_rwgt_card: \"set\" line without a parameter index: " + line);
+                std::vector<int> curr_params;
+                for (; curr_ind_of_line + 1 < rwgt_line.size(); ++curr_ind_of_line)
+                    curr_params.push_back(REX::ctoi(rwgt_line[curr_ind_of_line]));
+                double curr_val = REX::ctod(rwgt_line[curr_ind_of_line]);
+                this->cards[curr_launch].add_param(curr_block, std::move(curr_params), curr_val);
+                this->cards[curr_launch].rwgt_com += line + "\n";
+                continue;
+            }
+        }
+    }
+
+    rwgt_slha &rwgt_slha::set_card_path(const std::string &path)
+    {
+        this->card_path = path;
+        return *this;
+    }
+
+    rwgt_slha &rwgt_slha::set_mod_card_path(const std::string &path)
+    {
+        this->mod_card_path = path;
+        return *this;
+    }
+
+    bool rwgt_slha::move_param_card(const std::string &new_path)
+    {
+        if (this->card_path.empty())
+            throw std::runtime_error("rwgt_slha::move_param_card: card_path not set");
+        if (new_path.empty() && this->mod_card_path.empty())
+        {
+            this->mod_card_path = this->card_path + ".mod";
+        }
+        else if (!new_path.empty())
+        {
+            this->mod_card_path = new_path;
+        }
+        try
+        {
+            std::filesystem::rename(this->card_path, this->mod_card_path);
+        }
+        catch (const std::filesystem::filesystem_error &e)
+        {
+            throw std::runtime_error("rwgt_slha::move_param_card: failed to rename original card to .mod backup: " + std::string(e.what()));
+        }
+        return true;
+    }
+
+    bool rwgt_slha::remove_mod_card()
+    {
+        if (this->card_path.empty())
+            throw std::runtime_error("rwgt_slha::remove_mod_card: card_path not set");
+        if (this->mod_card_path.empty())
+            throw std::runtime_error("rwgt_slha::remove_mod_card: mod_card_path not set");
+        try
+        {
+            std::filesystem::remove(this->card_path);
+        }
+        catch (const std::filesystem::filesystem_error &e)
+        {
+            throw std::runtime_error("rwgt_slha::remove_mod_card: failed to remove mod card: " + std::string(e.what()));
+        }
+        try
+        {
+            std::filesystem::rename(this->mod_card_path, this->card_path);
+        }
+        catch (const std::filesystem::filesystem_error &e)
+        {
+            throw std::runtime_error("rwgt_slha::remove_mod_card: failed to rename .mod backup to original card: " + std::string(e.what()));
+        }
+        return true;
+    }
+
+    bool rwgt_slha::write_rwgt_card(size_t idx)
+    {
+        if (idx >= this->cards.size())
+            throw std::out_of_range("rwgt_slha::write_rwgt_card: index out of range");
+        if (this->card_path.empty())
+            throw std::runtime_error("rwgt_slha::write_rwgt_card: card_path not set");
+        if (this->mod_card_path.empty())
+            this->mod_card_path = this->card_path + ".mod";
+        if (!std::filesystem::exists(this->mod_card_path))
+        {
+            this->move_param_card();
+        }
+        if (!this->orig_params.blocks.empty())
+        {
+            for (const auto &[key, val] : this->orig_params.blocks)
+            {
+                if (REX::to_upper(key) == "DECAY")
+                {
+                    for (const auto &[pid, width] : val.params)
+                    {
+                        this->REX::slha::set_decay(pid, width);
+                    }
+                }
+                else
+                {
+                    for (auto [param_id, param_value] : val.params)
+                    {
+                        this->REX::slha::set(key, param_id, param_value);
+                    }
+                    for (const auto &[param_ids, param_value] : val.multi_params)
+                    {
+                        this->REX::slha::set(key, param_ids, param_value);
+                    }
+                }
+            }
+            this->orig_params.blocks.clear();
+        }
+        for (const auto &[key, val] : this->cards[idx].blocks)
+        {
+            if (REX::to_upper(key) == "DECAY")
+            {
+                for (const auto &[pid, width] : val.params)
+                {
+                    auto curr_decay = this->REX::slha::get_decay(pid);
+                    this->orig_params.add_param("DECAY", pid, curr_decay);
+                    this->REX::slha::set_decay(pid, width);
+                }
+            }
+            else
+            {
+                for (auto [param_id, param_value] : val.params)
+                {
+                    auto curr_param = this->REX::slha::get(key, param_id);
+                    this->orig_params.add_param(key, param_id, curr_param);
+                    this->REX::slha::set(key, param_id, param_value);
+                }
+                for (const auto &[param_ids, param_value] : val.multi_params)
+                {
+                    auto curr_param = this->REX::slha::get(key, param_ids);
+                    this->orig_params.add_param(key, param_ids, curr_param);
+                    this->REX::slha::set(key, param_ids, param_value);
+                }
+            }
+        }
+        std::ofstream ofs(this->card_path);
+        if (!ofs)
+            throw std::runtime_error("rwgt_slha::write_rwgt_card: failed to open file");
+        this->REX::slha::write(ofs);
+        return true;
+    }
+
+    std::vector<std::function<bool()>> rwgt_slha::get_card_writers()
+    {
+        std::vector<std::function<bool()>> writers;
+        for (size_t i = 0; i < this->cards.size(); ++i)
+        {
+            writers.push_back([this, i]()
+                              { return this->write_rwgt_card(i); });
+        }
+        return writers;
+    }
+
+    std::vector<std::string> rwgt_slha::get_launch_names()
+    {
+        std::vector<std::string> names;
+        for (const auto &card : this->cards)
+        {
+            names.push_back(card.launch_name);
+        }
+        return names;
+    }
+
+    std::vector<std::string> rwgt_slha::get_rwgt_commands()
+    {
+        std::vector<std::string> commands;
+        for (const auto &card : this->cards)
+        {
+            commands.push_back(card.rwgt_com);
+        }
+        return commands;
+    }
+
+    threadPool::threadPool(unsigned nthreads)
+        : stop_(false), active_(0)
+    {
+        workers_.reserve(nthreads);
+        for (unsigned i = 0; i < nthreads; ++i)
+        {
+            workers_.emplace_back([this]
+                                  {
+                for (;;) {
+                    Task task;
+                    {
+                        std::unique_lock<std::mutex> lk(m_);
+                        cv_.wait(lk, [this]{ return stop_ || !q_.empty(); });
+                        if (stop_ && q_.empty()) return;
+                        task = std::move(q_.front());
+                        q_.pop();
+                        ++active_;
+                    }
+                    try {
+                        task();
+                    } catch (...) {
+                        // Record first exception and signal cancellation
+                        {
+                            std::lock_guard<std::mutex> g(err_m_);
+                            if (!first_error_) first_error_ = std::current_exception();
+                        }
+                        cancel_.store(true, std::memory_order_relaxed);
+                    }
+                    {
+                        std::lock_guard<std::mutex> lk(m_);
+                        --active_;
+                        if (q_.empty() && active_ == 0) drained_.notify_all();
+                    }
+                } });
+        }
+    }
+
+    threadPool::~threadPool()
+    {
+        {
+            std::lock_guard<std::mutex> lk(m_);
+            stop_ = true;
+        }
+        cv_.notify_all();
+        for (auto &t : workers_)
+            t.join();
+    }
+
+    void threadPool::enqueue(Task t)
+    {
+        {
+            std::lock_guard<std::mutex> lk(m_);
+            q_.push(std::move(t));
+        }
+        cv_.notify_one();
+    }
+
+    void threadPool::begin_batch()
+    {
+        cancel_.store(false, std::memory_order_relaxed);
+        std::lock_guard<std::mutex> g(err_m_);
+        first_error_ = nullptr;
+    }
+
+    void threadPool::wait_batch()
+    {
+        std::unique_lock<std::mutex> lk(m_);
+        drained_.wait(lk, [this]
+                      { return (q_.empty() && active_ == 0) || first_error_; });
+        lk.unlock();
+        if (first_error_)
+            std::rethrow_exception(first_error_);
+    }
+
+    bool threadPool::cancel_requested() const noexcept
+    {
+        return cancel_.load(std::memory_order_relaxed);
+    }
+
+    procReweightor::rwgt_slot procReweightor::wrap(weightor w)
+    {
+        if (!w)
+            return {};
+        return {[w](process &p) -> weight_puller
+                {
+                    auto result = w(p);
+                    return [result]()
+                    { return result; };
+                },
+                true};
+    }
+
+    procReweightor::rwgt_slot procReweightor::wrap(deferred_weightor w)
+    {
+        if (!w)
+            return {};
+        return {w, false};
+    }
+
+    procReweightor::procReweightor(weightor reweight_function)
+    {
+        this->reweight_functions.push_back(wrap(reweight_function));
+    }
+
+    procReweightor::procReweightor(weightor reweight_function, eventBelongs selector)
+    {
+        this->reweight_functions.push_back(wrap(reweight_function));
+        this->event_checker = std::make_shared<eventBelongs>(std::move(selector));
+        this->event_checker_fn = this->event_checker->get_event_bool();
+    }
+
+    procReweightor::procReweightor(weightor reweight_function, std::shared_ptr<eventBelongs> selector)
+    {
+        this->reweight_functions.push_back(wrap(reweight_function));
+        this->event_checker = selector;
+        this->event_checker_fn = selector->get_event_bool();
+    }
+
+    procReweightor::procReweightor(std::vector<weightor> rwgts)
+    {
+        for (auto &w : rwgts)
+            this->reweight_functions.push_back(wrap(w));
+    }
+
+    procReweightor::procReweightor(std::vector<weightor> rwgts, std::shared_ptr<eventBelongs> selector)
+    {
+        for (auto &w : rwgts)
+            this->reweight_functions.push_back(wrap(w));
+        this->event_checker = selector;
+        this->event_checker_fn = selector->get_event_bool();
+    }
+
+    procReweightor::procReweightor(std::vector<weightor> rwgts, std::shared_ptr<eventBelongs> selector, weightor normaliser)
+    {
+        for (auto &w : rwgts)
+            this->reweight_functions.push_back(wrap(w));
+        this->event_checker = selector;
+        this->event_checker_fn = selector->get_event_bool();
+        this->normaliser = normaliser;
+    }
+
+    procReweightor::procReweightor(std::vector<weightor> rwgts, eventBelongs selector)
+    {
+        for (auto &w : rwgts)
+            this->reweight_functions.push_back(wrap(w));
+        this->event_checker = std::make_shared<eventBelongs>(std::move(selector));
+        this->event_checker_fn = this->event_checker->get_event_bool();
+    }
+
+    procReweightor::procReweightor(std::vector<weightor> rwgts, eventBelongs selector, weightor normaliser)
+    {
+        for (auto &w : rwgts)
+            this->reweight_functions.push_back(wrap(w));
+        this->event_checker = std::make_shared<eventBelongs>(std::move(selector));
+        this->event_checker_fn = this->event_checker->get_event_bool();
+        this->normaliser = normaliser;
+    }
+
+    procReweightor::procReweightor(deferred_weightor reweight_function)
+    {
+        this->reweight_functions.push_back(wrap(reweight_function));
+    }
+
+    procReweightor::procReweightor(deferred_weightor reweight_function, eventBelongs selector)
+    {
+        this->reweight_functions.push_back(wrap(reweight_function));
+        this->event_checker = std::make_shared<eventBelongs>(std::move(selector));
+        this->event_checker_fn = this->event_checker->get_event_bool();
+    }
+
+    procReweightor::procReweightor(deferred_weightor reweight_function, std::shared_ptr<eventBelongs> selector)
+    {
+        this->reweight_functions.push_back(wrap(reweight_function));
+        this->event_checker = selector;
+        this->event_checker_fn = selector->get_event_bool();
+    }
+
+    procReweightor::procReweightor(std::vector<deferred_weightor> rwgts)
+    {
+        for (auto &w : rwgts)
+            this->reweight_functions.push_back(wrap(w));
+    }
+
+    procReweightor::procReweightor(std::vector<deferred_weightor> rwgts, std::shared_ptr<eventBelongs> selector)
+    {
+        for (auto &w : rwgts)
+            this->reweight_functions.push_back(wrap(w));
+        this->event_checker = selector;
+        this->event_checker_fn = selector->get_event_bool();
+    }
+
+    procReweightor::procReweightor(std::vector<deferred_weightor> rwgts, std::shared_ptr<eventBelongs> selector, weightor normaliser)
+    {
+        for (auto &w : rwgts)
+            this->reweight_functions.push_back(wrap(w));
+        this->event_checker = selector;
+        this->event_checker_fn = selector->get_event_bool();
+        this->normaliser = normaliser;
+    }
+
+    procReweightor::procReweightor(std::vector<deferred_weightor> rwgts, eventBelongs selector)
+    {
+        for (auto &w : rwgts)
+            this->reweight_functions.push_back(wrap(w));
+        this->event_checker = std::make_shared<eventBelongs>(std::move(selector));
+        this->event_checker_fn = this->event_checker->get_event_bool();
+    }
+
+    procReweightor::procReweightor(std::vector<deferred_weightor> rwgts, eventBelongs selector, weightor normaliser)
+    {
+        for (auto &w : rwgts)
+            this->reweight_functions.push_back(wrap(w));
+        this->event_checker = std::make_shared<eventBelongs>(std::move(selector));
+        this->event_checker_fn = this->event_checker->get_event_bool();
+        this->normaliser = normaliser;
+    }
+
+    procReweightor &procReweightor::calc_normalisation(std::vector<double> wgts)
+    {
+        if (!this->proc)
+            throw std::runtime_error("procReweightor::calc_normalisation: process not set before calculating normalisation");
+        if (wgts.size() != this->proc->weight_.size())
+            throw std::runtime_error("procReweightor::calc_normalisation: weights vector size does not match number of original weights in process");
+        this->normalisation = wgts;
+        std::transform(this->normalisation.begin(), this->normalisation.end(), this->normalisation.begin(),
+                       [](double val)
+                       { return (val == 0.0) ? 0.0 : 1.0 / val; });
+        this->normalisation = *REX::vec_elem_mult<double>(this->normalisation, this->proc->weight_);
+        return *this;
+    }
+
+    procReweightor &procReweightor::set_event_checker(eventBelongs checker)
+    {
+        this->event_checker = std::make_shared<eventBelongs>(std::move(checker));
+        this->event_checker_fn = this->event_checker->get_event_bool();
+        return *this;
+    }
+
+    procReweightor &procReweightor::set_event_checker(REX::event_bool_fn checker)
+    {
+        this->event_checker = nullptr;
+        this->event_checker_fn = checker;
+        return *this;
+    }
+
+    procReweightor &procReweightor::set_normaliser(weightor normaliser)
+    {
+        this->normaliser = normaliser;
+        return *this;
+    }
+
+    procReweightor &procReweightor::set_normalisation(std::vector<double> normalisation)
+    {
+        if (!this->proc)
+            throw std::runtime_error("procReweightor::set_normalisation: process not set before setting normalisation");
+        if(normalisation.size() != this->proc->weight_.size() && this->proc->weight_.size() > 0)
+            throw std::runtime_error("procReweightor::set_normalisation: normalisation vector size does not match number of original weights in process");
+        this->normalisation = normalisation;
+        this->explicitly_set_norm = true;
+        return *this;
+    }
+
+    procReweightor &procReweightor::set_reweight_functions(weightor rwgt)
+    {
+        this->reweight_functions = {wrap(rwgt)};
+        if (!this->normaliser)
+            this->normaliser = rwgt;
+        return *this;
+    }
+
+    procReweightor &procReweightor::set_reweight_functions(std::vector<weightor> rwgts)
+    {
+        this->reweight_functions.clear();
+        this->reweight_functions.reserve(rwgts.size());
+        for (auto &w : rwgts)
+            this->reweight_functions.push_back(wrap(w));
+        return *this;
+    }
+
+    procReweightor &procReweightor::add_reweight_function(weightor rwgt)
+    {
+        this->reweight_functions.push_back(wrap(rwgt));
+        return *this;
+    }
+
+    procReweightor &procReweightor::set_deferred_reweight_functions(deferred_weightor rwgt)
+    {
+        this->reweight_functions = {wrap(rwgt)};
+        return *this;
+    }
+
+    procReweightor &procReweightor::set_deferred_reweight_functions(std::vector<deferred_weightor> rwgts)
+    {
+        this->reweight_functions.clear();
+        this->reweight_functions.reserve(rwgts.size());
+        for (auto &w : rwgts)
+            this->reweight_functions.push_back(wrap(w));
+        return *this;
+    }
+
+    procReweightor &procReweightor::add_deferred_reweight_function(deferred_weightor rwgt)
+    {
+        this->reweight_functions.push_back(wrap(rwgt));
+        return *this;
+    }
+
+    procReweightor &procReweightor::set_process(std::shared_ptr<process> p)
+    {
+        this->proc = p;
+        return *this;
+    }
+
+    procReweightor &procReweightor::set_initial_weights(std::vector<double> weights)
+    {
+        this->explicitly_set_norm = true;
+        return this->calc_normalisation(weights);
+    }
+
+    // Member functions for handling reweighting
+    void procReweightor::initialise()
+    {
+        if (!this->proc)
+            throw std::runtime_error("procReweightor::initialise: process not set before initialisation");
+        if (this->explicitly_set_norm) return; // If normalisation is explicitly set, skip the rest of the initialisation
+        if (!this->normaliser)
+        {
+            // reweight_functions may hold a mix of immediate and genuinely
+            // deferred slots -- scan for the first immediate one, since
+            // only its result is guaranteed to already exist right now.
+            size_t first_immediate = this->reweight_functions.size();
+            size_t immediate_count = 0;
+            bool any_slot = false;
+            for (size_t i = 0; i < this->reweight_functions.size(); ++i)
+            {
+                if (!this->reweight_functions[i])
+                    continue;
+                any_slot = true;
+                if (!this->reweight_functions[i].immediate)
+                    continue;
+                if (first_immediate == this->reweight_functions.size())
+                    first_immediate = i;
+                ++immediate_count;
+            }
+            if (first_immediate == this->reweight_functions.size())
+            {
+                if (!any_slot)
+                {
+                    this->verbose.warn("procReweightor::initialise: no reweight functions set, process will only yield zero weights.");
+                    this->normalisation = std::vector<double>(this->proc->weight_.size(), 0.0);
+                    return;
+                }
+                throw std::runtime_error(
+                    "procReweightor::initialise: only deferred reweight functions "
+                    "are registered and no normaliser was set; call "
+                    "set_normaliser() explicitly");
+            }
+            if (immediate_count != 1)
+                this->verbose.warn("procReweightor::initialise: multiple reweight functions set, assuming first is default evaluator and using it for normalisation.");
+            auto slot = this->reweight_functions[first_immediate];
+            this->normaliser = [slot](process &p)
+            { return slot.fn(p)(); };
+        }
+        auto normalised = this->normaliser(*this->proc);
+        this->calc_normalisation(*normalised);
+    }
+
+    void procReweightor::initialise(std::shared_ptr<process> p)
+    {
+        this->proc = p;
+        initialise();
+    }
+
+    void procReweightor::evaluate()
+    {
+        return this->evaluate(0);
+    }
+
+    void procReweightor::evaluate(size_t amp)
+    {
+        if (!this->proc)
+            throw std::runtime_error("procReweightor::evaluate: process not set before evaluation");
+        if (amp >= this->reweight_functions.size() || !this->reweight_functions[amp])
+            return this->append_zero_weights();
+        if (this->normalisation.empty())
+            this->initialise();
+        auto &slot = this->reweight_functions[amp];
+        auto puller = slot.fn(*this->proc);
+        if (!puller)
+            throw std::runtime_error("procReweightor::evaluate: reweight function returned an empty puller");
+        this->backlog.push_back({std::move(puller), slot.immediate});
+    }
+
+    void procReweightor::append_zero_weights()
+    {
+        if (!this->proc)
+            throw std::runtime_error("procReweightor::append_zero_weights: process not set before appending zero weights");
+        auto zeros = std::make_shared<std::vector<double>>(this->proc->weight_.size(), 0.0);
+        this->backlog.push_back({[zeros]()
+                                  { return zeros; },
+                                  true});
+    }
+
+    void procReweightor::append_backlog()
+    {
+        if (this->normalisation.empty())
+            throw std::runtime_error("procReweightor::append_backlog: normalisation is empty; call initialise() first");
+
+        std::vector<backlog_entry> remaining;
+        remaining.reserve(this->backlog.size());
+        for (auto &entry : this->backlog)
+        {
+            if (!entry.immediate)
+            {
+                remaining.push_back(std::move(entry));
+                continue;
+            }
+            auto weights = entry.pull();
+            if (!weights)
+                throw std::runtime_error("procReweightor::append_backlog: reweight function returned null pointer");
+            if (weights->size() != this->normalisation.size())
+                throw std::runtime_error("procReweightor::append_backlog: size mismatch between weights and normalisation");
+
+            this->proc->append_wgts(*REX::vec_elem_mult<double>(*weights, this->normalisation));
+        }
+        this->backlog = std::move(remaining);
+
+        // One outer reweighting iteration just finished (this is called
+        // exactly once per iteration, from reweightor::run_all_iterations()/
+        // run_streaming()); once max_deferred_iterations of them have gone
+        // by without a flush, pull and clear whatever deferred results are
+        // still pending -- see max_deferred_iterations' own doc comment
+        ++this->deferred_iterations_since_flush;
+        if (this->max_deferred_iterations != npos &&
+            this->deferred_iterations_since_flush >= this->max_deferred_iterations)
+        {
+            this->pull_deferred_backlog();
+        }
+    }
+
+    void procReweightor::pull_deferred_backlog()
+    {
+        this->deferred_iterations_since_flush = 0;
+        if (this->backlog.empty())
+            return;
+        if (this->normalisation.empty())
+            throw std::runtime_error("procReweightor::pull_deferred_backlog: normalisation is empty; call initialise() first");
+
+        // append_backlog() flushes every immediate entry each iteration, so
+        // whatever remains here, whether this runs mid-run (because
+        // max_deferred_iterations fired) or once at the very end, must be a
+        // genuinely deferred entry
+        for (auto &entry : this->backlog)
+        {
+            auto weights = entry.pull();
+            if (!weights)
+                throw std::runtime_error("procReweightor::pull_deferred_backlog: weight puller returned null pointer");
+            if (weights->size() != this->normalisation.size())
+                throw std::runtime_error("procReweightor::pull_deferred_backlog: size mismatch between pulled weights and normalisation");
+
+            this->proc->append_wgts(*REX::vec_elem_mult<double>(*weights, this->normalisation));
+        }
+        this->backlog.clear();
+    }
+
+    void xSecAccumulator::absorb(const std::vector<std::shared_ptr<REX::event>> &events)
+    {
+        for (auto &ev : events)
+        {
+            if (!ev)
+                continue;
+            ++this->n_events;
+            this->weight_sum += ev->weight_;
+            this->weight_sumSq += ev->weight_ * ev->weight_;
+            if (this->wgt_sum.size() < ev->wgts_.size())
+            {
+                this->wgt_sum.resize(ev->wgts_.size(), 0.0);
+                this->ratio_sum.resize(ev->wgts_.size(), 0.0);
+                this->ratio_sumSq.resize(ev->wgts_.size(), 0.0);
+            }
+            for (size_t k = 0; k < ev->wgts_.size(); ++k)
+            {
+                this->wgt_sum[k] += ev->wgts_[k];
+                double ratio = (ev->weight_ != 0.0) ? ev->wgts_[k] / ev->weight_ : 0.0;
+                this->ratio_sum[k] += ratio;
+                this->ratio_sumSq[k] += ratio * ratio;
+            }
+        }
+    }
+
+    void xSecAccumulator::finalise_into(reweightor &r) const
+    {
+        if (this->n_events == 0)
+            throw std::runtime_error("xSecAccumulator::finalise_into: no events were streamed, cannot calculate cross sections");
+
+        // Mirrors reweightor::calc_xSec_from_weights(): only estimate
+        // xSec_/xSecErr_ from event weights if no header cross section (eg a
+        // real <init> block) is already available
+        if (std::accumulate(r.xSec_.begin(), r.xSec_.end(), 0.0) == 0.0)
+        {
+            double mean = this->weight_sum / double(this->n_events);
+            double variance = std::max(this->weight_sumSq / double(this->n_events) - std::pow(mean, 2), 0.0);
+            r.xSec_ = {mean};
+            r.xSecErr_ = {std::sqrt(variance / double(this->n_events))};
+        }
+
+        // Mirrors reweightor::calc_norm()
+        r.norm_factor = 1.0;
+        if (std::abs(r.idWgt_) == 3)
+        {
+            r.norm_factor = std::accumulate(r.xSec_.begin(), r.xSec_.end(), 0.0) / double(this->n_events);
+        }
+        else if (std::abs(r.idWgt_) == 4)
+        {
+            r.norm_factor = 1.0 / double(this->n_events);
+        }
+        else
+        {
+            if (std::abs(r.idWgt_) > 2 || r.idWgt_ == 0)
+                r.verbose.warn("reweightor::run_streaming: idWgt is not set to a value defined in the LHE standard. Assuming weighted events.");
+            double total_xSec = std::accumulate(r.xSec_.begin(), r.xSec_.end(), 0.0);
+            if (this->weight_sum == 0.0)
+                throw std::runtime_error("xSecAccumulator::finalise_into: total weight is zero, cannot calculate norm");
+            r.norm_factor = total_xSec / this->weight_sum;
+        }
+
+        // Mirrors reweightor::calc_xSecs()
+        r.rwgt_xSec = this->wgt_sum;
+        for (auto &x : r.rwgt_xSec)
+            x *= r.norm_factor;
+
+        // Mirrors reweightor::calc_xErrs()
+        double loc_xSec = std::accumulate(r.xSec_.begin(), r.xSec_.end(), 0.0);
+        double loc_xErr = std::sqrt(std::accumulate(r.xSecErr_.begin(), r.xSecErr_.end(), 0.0, [](double a, double b)
+                                                     { return a + b * b; }));
+        double invNoEvs = 1.0 / double(this->n_events);
+        double sqrtInvNoEvs = std::sqrt(invNoEvs);
+        r.rwgt_xErr.assign(r.rwgt_xSec.size(), 0.0);
+        for (size_t k = 0; k < r.rwgt_xSec.size(); ++k)
+        {
+            double variance = (this->ratio_sumSq[k] - std::pow(this->ratio_sum[k], 2) * invNoEvs) * invNoEvs;
+            variance = std::max(variance, 0.0);
+            r.rwgt_xErr[k] = loc_xSec * std::sqrt(variance) * sqrtInvNoEvs + loc_xErr * this->ratio_sum[k] * invNoEvs;
+            if (std::isnan(r.rwgt_xErr[k]) || std::isinf(r.rwgt_xErr[k]) || r.rwgt_xErr[k] <= 0.0)
+            {
+                if (loc_xSec <= 0.0 || r.rwgt_xSec[k] == 0.0)
+                {
+                    r.verbose.warn("reweightor::run_streaming: Error propagation failed for weight " + std::to_string(k) + " and no baseline cross section is available to approximate it from; setting the error to 0.", true);
+                    r.rwgt_xErr[k] = 0.0;
+                }
+                else
+                {
+                    r.verbose.warn("reweightor::run_streaming: Error propagation failed for weight " + std::to_string(k) + ". Approximating the error at the level of the cross section.");
+                    r.rwgt_xErr[k] = loc_xErr * std::max(loc_xSec / r.rwgt_xSec[k], r.rwgt_xSec[k] / loc_xSec);
+                }
+            }
+        }
+    }
+
+    reweightor::reweightor(lhe &&mother) : lhe(std::move(mother)) {}
+
+    reweightor::reweightor(const lhe &mother) : lhe(mother) {}
+
+    reweightor::reweightor(lhe &&mother, std::vector<std::shared_ptr<procReweightor>> rws) : lhe(std::move(mother)), reweightors(rws) {}
+
+    reweightor::reweightor(const lhe &mother, std::vector<std::shared_ptr<procReweightor>> rws) : lhe(mother), reweightors(rws) {}
+
+    reweightor::reweightor(lhe &&mother, std::vector<std::shared_ptr<procReweightor>> rws, std::vector<iterator> iters) : lhe(std::move(mother)), reweightors(rws), iterators(std::move(iters)) {}
+
+    reweightor::reweightor(const lhe &mother, std::vector<std::shared_ptr<procReweightor>> rws, std::vector<iterator> iters) : lhe(mother), reweightors(rws), iterators(std::move(iters)) {}
+
+    reweightor::reweightor(lhe &&mother, std::vector<procReweightor> rws) : lhe(std::move(mother))
+    {
+        this->set_reweightors(rws);
+    }
+
+    reweightor::reweightor(const lhe &mother, std::vector<procReweightor> rws) : lhe(mother)
+    {
+        this->set_reweightors(rws);
+    }
+
+    reweightor::reweightor(lhe &&mother, std::vector<procReweightor> rws, std::vector<iterator> iters) : lhe(std::move(mother)), iterators(std::move(iters))
+    {
+        this->set_reweightors(rws);
+    }
+
+    reweightor::reweightor(const lhe &mother, std::vector<procReweightor> rws, std::vector<iterator> iters) : lhe(mother), iterators(std::move(iters))
+    {
+        this->set_reweightors(rws);
+    }
+
+    reweightor &reweightor::set_reweightors(std::vector<std::shared_ptr<procReweightor>> rws)
+    {
+        this->reweightors = rws;
+        return *this;
+    }
+
+    reweightor &reweightor::set_reweightors(std::vector<procReweightor> rws)
+    {
+        this->reweightors.clear();
+        for (auto &rw : rws)
+        {
+            this->reweightors.push_back(std::make_shared<procReweightor>(std::move(rw)));
+        }
+        return *this;
+    }
+
+    reweightor &reweightor::add_reweightor(procReweightor &rw)
+    {
+        this->reweightors.push_back(std::make_shared<procReweightor>(rw));
+        return *this;
+    }
+
+    reweightor &reweightor::add_reweightor(procReweightor &&rw)
+    {
+        this->reweightors.push_back(std::make_shared<procReweightor>(std::move(rw)));
+        return *this;
+    }
+
+    reweightor &reweightor::add_reweightor(std::shared_ptr<procReweightor> rw)
+    {
+        this->reweightors.push_back(rw);
+        return *this;
+    }
+
+    reweightor &reweightor::set_initialise(iterator init)
+    {
+        this->initialise = init;
+        return *this;
+    }
+
+    reweightor &reweightor::set_finalise(iterator fin)
+    {
+        this->finalise = fin;
+        return *this;
+    }
+
+    reweightor &reweightor::set_iterators(const std::vector<iterator> &iters)
+    {
+        this->iterators = iters;
+        return *this;
+    }
+
+    reweightor &reweightor::add_iterator(const iterator &iter)
+    {
+        this->iterators.push_back(iter);
+        return *this;
+    }
+
+    reweightor &reweightor::add_iterator(iterator &&iter)
+    {
+        this->iterators.push_back(std::move(iter));
+        return *this;
+    }
+
+    reweightor &reweightor::set_launch_names(const std::vector<std::string> &names)
+    {
+        this->launch_names = names;
+        return *this;
+    }
+
+    reweightor &reweightor::add_launch_name(const std::string &name)
+    {
+        this->launch_names.push_back(name);
+        return *this;
+    }
+
+    void reweightor::calc_xSec_from_weights()
+    {
+        if (std::accumulate(this->xSec_.begin(), this->xSec_.end(), 0.0) != 0.0)
+            return; // a header cross section (e.g. from a real <init> block) is already available
+        if (this->events.empty())
+            this->transpose();
+        if (this->events.empty())
+            throw std::runtime_error("reweightor::calc_xSec_from_weights: no events found, cannot estimate a cross section");
+        double sum = 0.0, sumSq = 0.0;
+        for (auto ev : this->events)
+        {
+            sum += ev->weight_;
+            sumSq += std::pow(ev->weight_, 2);
+        }
+        double nEvs = double(this->events.size());
+        double mean = sum / nEvs;
+        double variance = std::max(sumSq / nEvs - std::pow(mean, 2), 0.0);
+        this->xSec_ = {mean};
+        this->xSecErr_ = {std::sqrt(variance / nEvs)};
+    }
+
+    void reweightor::calc_norm()
+    {
+        if (this->events.empty())
+            throw std::runtime_error("reweightor::calc_norm: no events loaded, cannot calculate norm");
+        this->calc_xSec_from_weights();
+        this->norm_factor = 1.0;
+        if (std::abs(this->idWgt_) == 3)
+        {
+            this->norm_factor = std::accumulate(this->xSec_.begin(), this->xSec_.end(), 0.0);
+            this->norm_factor /= this->events.size();
+        }
+        else if (std::abs(this->idWgt_) == 4)
+        {
+            this->norm_factor = 1. / this->events.size();
+        }
+        else
+        {
+            if (std::abs(this->idWgt_) > 2 || this->idWgt_ == 0)
+                this->verbose.warn("reweightor::calc_norm: idWgt is not set to a value defined in the LHE standard. Assuming weighted events.");
+            this->norm_factor = std::accumulate(this->xSec_.begin(), this->xSec_.end(), 0.0);
+            double accumulated_wgts = 0.0;
+            for (const auto &proc : this->processes)
+            {
+                accumulated_wgts += std::accumulate(proc->weight_.begin(), proc->weight_.end(), 0.0);
+            }
+            if (accumulated_wgts == 0.0)
+            {
+                for (auto ev : this->events)
+                {
+                    accumulated_wgts += ev->weight_;
+                }
+            }
+            if (accumulated_wgts == 0.0)
+                throw std::runtime_error("reweightor::calc_norm: total weight is zero, cannot calculate norm");
+            this->norm_factor /= accumulated_wgts;
+        }
+    }
+
+    void reweightor::set_norm(double norm)
+    {
+        this->norm_factor = norm;
+    }
+
+    void reweightor::setup_pool()
+    {
+        if (!pool)
+        {
+            unsigned hc = std::thread::hardware_concurrency();
+            if (hc == 0)
+                hc = 1;
+            unsigned want = this->pool_threads ? this->pool_threads : hc;
+            unsigned n = static_cast<unsigned>(
+                std::max<size_t>(1, std::min<size_t>(reweightors.size(), want)));
+            this->pool = std::make_unique<threadPool>(n);
+        }
+    }
+
+    void reweightor::extract_sorter()
+    {
+        if (this->reweightors.empty())
+            throw std::runtime_error("reweightor::extract_sorter: no procReweightors set in reweightor");
+
+        std::vector<event_bool_fn> preds;
+        preds.reserve(this->reweightors.size());
+        for (const auto &rw : this->reweightors)
+        {
+            if (rw->event_checker_fn)
+            {
+                preds.push_back(rw->event_checker_fn);
+            }
+            else
+            {
+                preds.push_back(rw->event_checker->get_event_bool());
+            }
+        }
+
+        this->set_sorter(eventSorter(std::move(preds)));
+        this->sorted_events.clear();
+        this->processes.clear();
+        // classify the events on as many threads as the reweighting uses
+        unsigned hc = std::thread::hardware_concurrency();
+        this->sort_threads = this->pool_threads ? this->pool_threads : std::max(hc, 1u);
+        this->events_to_processes(); // sorts the events; also (re)populates process_group, 1:1 with processes
+
+        const size_t R = this->reweightors.size();
+        if (this->processes.size() != this->process_group.size())
+            throw std::runtime_error("reweightor::extract_sorter: processes/process_group size mismatch.");
+
+        auto processes_full = std::move(this->processes);
+        auto groups_full = std::move(this->process_group);
+        auto base_reweightors = std::move(this->reweightors);
+
+        // Tracks, per registered procReweightor (ie per sorter class), whether its
+        // first batch has already been assigned. That first batch reuses the
+        // original procReweightor object; any further max_batch overflow batches
+        // of the same class each get a fresh clone of it -- identical reweight
+        // functions, normaliser and event checker, bound (via set_process below)
+        // to their own process/event subsample -- mirroring lhe::process_group's
+        // "additional index per overflow batch" on the reweightor side
+        std::vector<bool> group_claimed(R, false);
+
+        std::vector<std::shared_ptr<process>> procs;
+        std::vector<std::shared_ptr<procReweightor>> rwgs;
+        std::vector<size_t> groups_kept;
+        procs.reserve(processes_full.size());
+        rwgs.reserve(processes_full.size());
+        groups_kept.reserve(processes_full.size());
+
+        for (size_t i = 0; i < processes_full.size(); ++i)
+        {
+            if (processes_full[i]->events.empty())
+                continue;
+
+            size_t group = groups_full[i];
+            std::shared_ptr<procReweightor> rw;
+            if (group == npos)
+            {
+                // Catch-all "matched no registered procReweightor" bucket: every
+                // batch of it gets its own independent, unconfigured
+                // procReweightor (which yields zero weights, same as before)
+                rw = std::make_shared<procReweightor>();
+            }
+            else
+            {
+                if (group >= R)
+                    throw std::runtime_error("reweightor::extract_sorter: process_group index out of range.");
+                if (!group_claimed[group])
+                {
+                    rw = base_reweightors[group];
+                    group_claimed[group] = true;
+                }
+                else
+                {
+                    rw = std::make_shared<procReweightor>(*base_reweightors[group]);
+                    // Clone only the config; runtime state (proc/normalisation/
+                    // backlog) belongs to the batch it will be bound to below,
+                    // not whatever the base procReweightor may have accumulated
+                    rw->proc = nullptr;
+                    rw->normalisation.clear();
+                    rw->backlog.clear();
+                }
+            }
+
+            procs.push_back(processes_full[i]);
+            rwgs.push_back(std::move(rw));
+            groups_kept.push_back(group);
+        }
+
+        this->processes = std::move(procs);
+        this->reweightors = std::move(rwgs);
+        this->process_group = std::move(groups_kept);
+
+        if (this->processes.size() != this->reweightors.size())
+            throw std::runtime_error("reweightor::extract_sorter: number of processes does not match number of reweightors.");
+
+        for (size_t i = 0; i < this->reweightors.size(); ++i)
+        {
+            auto &p = this->processes[i];
+            p->validate();
+            this->reweightors[i]->set_process(p);
+            // Mute this batch's procReweightor whenever this reweightor is
+            // muted, without touching its own default (a procReweightor
+            // used standalone, outside of any reweightor, stays verbose)
+            this->reweightors[i]->verbose.link(this->verbose);
+        }
+    }
+
+    void reweightor::extract_sorter_stream(const std::vector<std::shared_ptr<procReweightor>> &base)
+    {
+        if (base.empty())
+            throw std::runtime_error("reweightor::extract_sorter_stream: no procReweightors registered");
+
+        std::vector<event_bool_fn> preds;
+        preds.reserve(base.size());
+        for (const auto &rw : base)
+        {
+            if (rw->event_checker_fn)
+                preds.push_back(rw->event_checker_fn);
+            else
+                preds.push_back(rw->event_checker->get_event_bool());
+        }
+
+        this->set_sorter(eventSorter(std::move(preds)));
+        this->sorted_events.clear();
+        this->processes.clear();
+        this->events_to_processes(); // sorts the events; also (re)populates process_group, 1:1 with processes
+
+        const size_t R = base.size();
+        if (this->processes.size() != this->process_group.size())
+            throw std::runtime_error("reweightor::extract_sorter_stream: processes/process_group size mismatch.");
+
+        std::vector<std::shared_ptr<process>> procs;
+        std::vector<std::shared_ptr<procReweightor>> rwgs;
+        std::vector<size_t> groups_kept;
+        procs.reserve(this->processes.size());
+        rwgs.reserve(this->processes.size());
+        groups_kept.reserve(this->processes.size());
+
+        for (size_t i = 0; i < this->processes.size(); ++i)
+        {
+            if (this->processes[i]->events.empty())
+                continue;
+
+            size_t group = this->process_group[i];
+            std::shared_ptr<procReweightor> rw;
+            if (group == npos)
+            {
+                // Catch-all "matched no registered procReweightor" bucket
+                rw = std::make_shared<procReweightor>();
+            }
+            else
+            {
+                if (group >= R)
+                    throw std::runtime_error("reweightor::extract_sorter_stream: process_group index out of range.");
+                // Unlike extract_sorter(), always clone: `base` must stay an
+                // untouched, reusable template across every batch, so no
+                // batch's runtime state (proc/normalisation/backlog) is ever
+                // written into it
+                rw = std::make_shared<procReweightor>(*base[group]);
+                rw->proc = nullptr;
+                rw->normalisation.clear();
+                rw->backlog.clear();
+            }
+
+            procs.push_back(this->processes[i]);
+            rwgs.push_back(std::move(rw));
+            groups_kept.push_back(group);
+        }
+
+        this->processes = std::move(procs);
+        this->reweightors = std::move(rwgs);
+        this->process_group = std::move(groups_kept);
+
+        if (this->processes.size() != this->reweightors.size())
+            throw std::runtime_error("reweightor::extract_sorter_stream: number of processes does not match number of reweightors.");
+
+        for (size_t i = 0; i < this->reweightors.size(); ++i)
+        {
+            auto &p = this->processes[i];
+            p->validate();
+            this->reweightors[i]->set_process(p);
+            this->reweightors[i]->verbose.link(this->verbose);
+        }
+    }
+
+    void reweightor::initialise_reweightors()
+    {
+        if (this->reweightors.size() != this->processes.size())
+            throw std::runtime_error("initialise_reweightors: reweightors/processes size mismatch");
+
+        // The normalisations, one task per batch on the worker pool, as the
+        // reweighting iterations (run_iteration())
+        setup_pool();
+        pool->begin_batch();
+        for (size_t i = 0; i < this->reweightors.size(); ++i)
+        {
+            pool->enqueue([this, i]
+                          {
+                    if (pool->cancel_requested()) return;
+                    this->reweightors[i]->initialise(this->processes[i]); });
+        }
+        pool->wait_batch();
+    }
+
+    void reweightor::finalise_reweighting()
+    {
+        // Materializes every genuinely-deferred result accumulated across
+        // the whole iteration loop, one pull per procReweightor rather than
+        // one per iteration
+        for (auto &rwgt : this->reweightors)
+        {
+            rwgt->pull_deferred_backlog();
+        }
+        for (auto proc : this->processes)
+        {
+            proc->transpose_wgts();
+            proc->validate();
+        }
+        if (!this->finalise())
+            this->verbose.warn("reweightor::finalise_reweighting: finalise iterator returned false, something might have gone wrong. Validate output manually.");
+        if (this->launch_names.size() > 0)
+        {
+            this->extract_weight_ids();
+            size_t nWgts = this->weight_ids->size();
+            for (size_t i = 0; i < this->launch_names.size(); ++i)
+            {
+                std::string curr_name = (this->launch_names[i].empty()) ? "rwgt_" + std::to_string(i + nWgts + 1) : this->launch_names[i];
+                this->weight_ids->push_back(curr_name);
+            }
+        }
+        this->calc_xSecs();
+        this->calc_xErrs();
+    }
+
+    void reweightor::sync_n_amps()
+    {
+        this->n_amps = 0;
+        for (auto &rwgt : this->reweightors)
+            this->n_amps = std::max(this->n_amps, rwgt->reweight_functions.size());
+        if (this->n_amps == 0)
+            throw std::runtime_error("reweightor::sync_n_amps: no reweight functions found, something went wrong.");
+    }
+
+    void reweightor::setup()
+    {
+        if (!this->initialise())
+            throw std::runtime_error("reweightor::setup: initialise iterator returned false, something went wrong.");
+        this->extract_sorter();
+        this->initialise_reweightors();
+        this->sync_n_amps();
+        this->setup_pool();
+    }
+
+    void reweightor::run_iteration()
+    {
+        if (!this->iterators[this->curr_iter]())
+            throw std::runtime_error("reweightor::run_iteration: iterator returned false, something went wrong in iteration " + std::to_string(this->curr_iter) + ".");
+        this->curr_iter++;
+
+        // Nothing to do?
+        const size_t N = this->reweightors.size();
+        if (N == 0 || this->n_amps == 0)
+            return;
+
+        // Ensure pool exists (persisted across iterations)
+        setup_pool();
+
+        // Parallel "reweightor" phase
+        pool->begin_batch();
+        for (size_t i = 0; i < N; ++i)
+        {
+            pool->enqueue([this, i]
+                          {
+                    // Early cancel check (best-effort)
+                    if (pool->cancel_requested()) return;
+        
+                    auto &rwgt = this->reweightors[i];
+                    for (size_t amp = 0; amp < this->n_amps; ++amp) {
+                        // Safe because each task owns a distinct rwgt
+                        rwgt->evaluate(amp);
+                        if (pool->cancel_requested()) return; // responsive cancellation
+                    } });
+        }
+        // Wait for completion (or rethrow first error)
+        pool->wait_batch();
+    }
+
+    void reweightor::run_all_iterations()
+    {
+        while (this->curr_iter < this->iterators.size())
+        {
+            this->run_iteration();
+            for (auto &rwgt : this->reweightors)
+            {
+                rwgt->append_backlog();
+            }
+#pragma optimize("", off)
+            this->verbose.print(".");
+#pragma optimize("", on)
+            std::cout.flush();
+        }
+    }
+
+    void reweightor::run()
+    {
+        this->setup();
+        this->run_all_iterations();
+        this->finalise_reweighting();
+    }
+
+    void reweightor::run_streaming(eventSink &sink, bool include_ids)
+    {
+        if (!this->streaming())
+            throw std::runtime_error("reweightor::run_streaming: lhe is not in streaming mode (no source set); use run() instead");
+        if (!this->initialise())
+            throw std::runtime_error("reweightor::run_streaming: initialise iterator returned false, something went wrong.");
+        if (this->reweightors.empty())
+            throw std::runtime_error("reweightor::run_streaming: no procReweightors set in reweightor");
+
+        // The procReweightors as originally registered, before any batch is
+        // pulled: extract_sorter_stream() only ever reads from this (cloning
+        // fresh instances per batch), never writes into it, so it stays a
+        // valid template for every batch of the run
+        auto base_reweightors = this->reweightors;
+
+        // Weight IDs (pre-existing <initrwgt> ids plus this run's
+        // launch_names) are entirely header/config-derived -- see
+        // lhe::extract_weight_ids() and finalise_reweighting() -- so they
+        // can be resolved once, up front, instead of once per batch. That
+        // way every batch's events already carry the final weight_ids as
+        // they're written
+        this->extract_weight_ids();
+        if (this->launch_names.size() > 0)
+        {
+            size_t nWgts = this->weight_ids->size();
+            for (size_t i = 0; i < this->launch_names.size(); ++i)
+            {
+                std::string curr_name = (this->launch_names[i].empty()) ? "rwgt_" + std::to_string(i + nWgts + 1) : this->launch_names[i];
+                this->weight_ids->push_back(curr_name);
+            }
+        }
+        this->append_weight_ids(include_ids);
+        sink.write_open(static_cast<const initNode &>(*this), this->header);
+
+        this->setup_pool();
+
+        xSecAccumulator acc;
+        while (this->pull_batch() > 0)
+        {
+            this->extract_sorter_stream(base_reweightors);
+            this->initialise_reweightors();
+            this->sync_n_amps();
+
+            this->curr_iter = 0;
+            this->run_all_iterations();
+
+            // Materialize every genuinely-deferred result for this batch
+            // now (rather than once at the very end, as finalise_reweighting()
+            // does for the non-streaming path): this batch's procReweightor
+            // instances are clones, about to be discarded, and the next
+            // batch's clones start from a fresh copy of `base_reweightors`
+            // regardless -- but draining now is also what makes it safe for
+            // extract_sorter_stream() to matter-of-factly clone `base` again
+            // next iteration without inheriting stale backlog state
+            for (auto &rwgt : this->reweightors)
+                rwgt->pull_deferred_backlog();
+            for (auto &proc : this->processes)
+            {
+                proc->transpose_wgts();
+                proc->validate();
+            }
+
+            acc.absorb(this->events);
+            for (auto &ev : this->events)
+                sink.write_event(ev);
+
+            this->verbose.print(".");
+            std::cout.flush();
+        }
+
+        if (!this->finalise())
+            this->verbose.warn("reweightor::run_streaming: finalise iterator returned false, something might have gone wrong. Validate output manually.");
+
+        acc.finalise_into(*this);
+        sink.write_close();
+    }
+
+    void reweightor::calc_xSecs()
+    {
+        if (this->norm_factor == 0.0)
+            this->calc_norm();
+        this->rwgt_xSec = std::vector<double>(this->events[0]->wgts_.size(), 0.0);
+        for (auto ev : this->events)
+        {
+            for (size_t i = 0; i < ev->wgts_.size(); ++i)
+            {
+                this->rwgt_xSec[i] += ev->wgts_[i];
+            }
+        }
+        for (auto &x : this->rwgt_xSec)
+            x *= this->norm_factor;
+    }
+
+    void reweightor::calc_xErrs()
+    {
+        if (this->rwgt_xSec.size() == 0)
+            this->calc_xSecs();
+        this->calc_xSec_from_weights();
+        double loc_xSec = std::accumulate(this->xSec_.begin(), this->xSec_.end(), 0.0);
+        double loc_xErr = std::sqrt(std::accumulate(this->xSecErr_.begin(), this->xSecErr_.end(), 0.0, [](double a, double b)
+                                                    { return a + b * b; }));
+        size_t nEvs = this->events.size();
+        if (nEvs == 0)
+        {
+            this->transpose();
+            nEvs = this->events.size();
+            if (nEvs == 0)
+                throw std::runtime_error("reweightor::calc_xErrs: no events found, cannot calculate errors");
+        }
+        this->rwgt_xErr = std::vector<double>(this->rwgt_xSec.size(), 0.0);
+        auto omg = std::vector<double>(this->rwgt_xSec.size(), 0.0);
+        auto omgSq = std::vector<double>(this->rwgt_xSec.size(), 0.0);
+        for (auto ev : this->events)
+        {
+            for (size_t k = 0; k < ev->wgts_.size(); ++k)
+            {
+                double ratio = ev->wgts_[k] / ev->weight_;
+                omg[k] += ratio;
+                omgSq[k] += std::pow(ratio, 2);
+            }
+        }
+        double invNoEvs = 1.0 / double(nEvs);
+        double sqrtInvNoEvs = std::sqrt(invNoEvs);
+        for (size_t k = 0; k < this->rwgt_xSec.size(); ++k)
+        {
+            double variance = (omgSq[k] - std::pow(omg[k], 2) * invNoEvs) * invNoEvs;
+            variance = std::max(variance, 0.0);
+            this->rwgt_xErr[k] = loc_xSec * std::sqrt(variance) * sqrtInvNoEvs + loc_xErr * omg[k] * invNoEvs;
+            if (std::isnan(this->rwgt_xErr[k]) || std::isinf(this->rwgt_xErr[k]) || this->rwgt_xErr[k] <= 0.0)
+            {
+                if (loc_xSec <= 0.0 || this->rwgt_xSec[k] == 0.0)
+                {
+                    this->verbose.warn("reweightor::calc_xErrs: Error propagation failed for weight " + std::to_string(k) + " and no baseline cross section is available to approximate it from; setting the error to 0.", true);
+                    this->rwgt_xErr[k] = 0.0;
+                }
+                else
+                {
+                    this->verbose.warn("reweightor::calc_xErrs: Error propagation failed for weight " + std::to_string(k) + ". Approximating the error at the level of the cross section.");
+                    this->rwgt_xErr[k] = loc_xErr * std::max(loc_xSec / this->rwgt_xSec[k], this->rwgt_xSec[k] / loc_xSec);
+                }
+            }
+        }
+    }
+
+    param_rwgt::param_rwgt(const lhe &mother, std::vector<std::shared_ptr<procReweightor>> rws, const std::string &slha_path, const std::string &rwgt_path)
+        : reweightor(mother, rws)
+    {
+        this->read_slha_rwgt(slha_path, rwgt_path);
+    }
+
+    void param_rwgt::read_slha_rwgt(std::istream &slha_in, std::istream &rwgt_in)
+    {
+        this->card_iter = rwgt_slha::create(slha_in, rwgt_in);
+        this->initialise = [&]()
+        { return this->card_iter.move_param_card(); };
+        this->finalise = [&]()
+        { return this->card_iter.remove_mod_card(); };
+        this->iterators = this->card_iter.get_card_writers();
+        this->launch_names = this->card_iter.get_launch_names();
+        this->weight_context = this->card_iter.get_rwgt_commands();
+    }
+
+    void param_rwgt::read_slha_rwgt(const std::string &slha_file, const std::string &rwgt_file)
+    {
+        std::ifstream slha_in(slha_file);
+        std::ifstream rwgt_in(rwgt_file);
+        if (!slha_in || !rwgt_in)
+            throw std::runtime_error("param_rwgt::read_slha_rwgt: failed to open input files");
+        this->read_slha_rwgt(slha_in, rwgt_in);
+        this->card_iter.set_card_path(slha_file);
+    }
+
+} // namespace REX::tea
+#endif // _TEAREX_CPP_
\ No newline at end of file
diff --git a/madtrex/include/madtrex/binary_io.hpp b/madtrex/include/madtrex/binary_io.hpp
new file mode 100644
index 0000000000..1d4c2172e2
--- /dev/null
+++ b/madtrex/include/madtrex/binary_io.hpp
@@ -0,0 +1,141 @@
+// MadtRex binary I/O: builds/writes REX::lhe objects against MadGraph7's
+// internal binary event-file format (madspace::EventFile, io.hpp) instead of
+// LHEF XML, without teaching Rex itself anything about that format.
+//
+// Rex already factors "assemble/sort a REX::lhe from some raw format" into a
+// generic REX::lheReader<InitRaw, EventRaw, HeaderRaw>/REX::lheWriter<...>
+// pair driven by plain translator functions (see Rex.h); REX::xml_reader()/
+// xml_writer() are simply that template instantiated with shared_ptr<xmlNode>
+// as the raw type. This header instantiates the same machinery with
+// madspace::LHEMeta/LHEEvent (io.hpp's lhe_output.hpp raw structs, which are
+// already field-for-field identical to the binary format's
+// EventRecord::f_lhe_event/ParticleRecord::f_lhe_particle layout) as the raw
+// types, so all translation/sorting bookkeeping is still Rex's, and only the
+// binary<->LHEEvent/LHEMeta plumbing lives here.
+
+#pragma once
+
+#include "tupperware.hpp"
+
+#include "madspace/driver/io.hpp"
+#include "madspace/driver/lhe_output.hpp"
+
+#include <memory>
+#include <string>
+
+namespace madtrex {
+
+// Reads a MadGraph7 internal binary event file (an EventFile opened in
+// EventFile::load mode, expected to have been written with
+// EventRecord::f_lhe_event | ParticleRecord::f_lhe_particle layout, eg via
+// EventGenerator::combine_to_lhe_npy) into a REX::lhe.
+//
+// The binary format itself carries no init-block information (beam ids/
+// energies, PDF set, per-process cross sections): MadGraph7's own driver
+// constructs that separately (see madspace::LHEMeta) and does not persist it
+// alongside the binary event file. meta is therefore an optional, caller-
+// supplied source for that information; the resulting REX::lhe's initNode is
+// left at its default (all-zero) values for whichever fields meta doesn't
+// specify. This has no effect on reweighting itself: procReweightor::
+// evaluate() only ever reads per-event data (momenta, pdg, status, flavor_/
+// helicity_ indices, alphaS), never initNode fields.
+//
+// When the file has a "matrix_element" column (the generation-time squared
+// matrix elements madspace stores on request, EventRecord::f_matrix_element),
+// each event's value is kept as extra[stored_matrix_element_key], for
+// TupperWare::set_stored_normalisation(). open_lhe_source_binary() does the
+// same; open_lhe_sink_binary() does not write it back out.
+std::shared_ptr<REX::lhe> load_lhe_binary(
+    const std::string& path, const madspace::LHEMeta& meta = {}
+);
+// As above, given an already-open EventFile (must be in EventFile::load
+// mode, with a layout compatible with EventRecord::f_lhe_event |
+// ParticleRecord::f_lhe_particle).
+std::shared_ptr<REX::lhe> load_lhe_binary(
+    madspace::EventFile& file, const madspace::LHEMeta& meta = {}
+);
+
+// Streaming counterpart to load_lhe_binary(const std::string&, ...): opens
+// path as a load-mode EventFile and returns a REX::eventSource whose
+// pull(n) reads up to n further events at a time straight off disk (via
+// EventFile::read()) rather than loading the whole file up front --
+// otherwise translating each event exactly as load_lhe_binary() does.
+// read_init() returns exactly what load_lhe_binary(path, meta)'s own init
+// would have been (meta is consulted once, there, and nowhere else); the
+// binary format has no header, so no read_header is set. Forward-only, like
+// open_lhe_source(): no rewind is offered, open this again for a second
+// pass. The returned source keeps its EventFile open for as long as its
+// pull closure is kept alive.
+REX::eventSource open_lhe_source_binary(
+    const std::string& path, const madspace::LHEMeta& meta = {}
+);
+
+// Writes the per-event weights recorded in lhe.events[i]->wgts_ (ie the
+// output of running a REX::tea::reweightor over lhe) back out as one
+// madspace weight-only binary file per entry of lhe.weight_ids, using
+// madspace::weight_file_layout (EventRecord::f_weight, no particle data) --
+// the same index-aligned, kinematics-free weight-file convention already
+// used for per-channel weights in ChannelEventGenerator. Files are named
+// "<path_prefix>_<weight_id>.npy"; weight_id is sanitized to a
+// filesystem-safe token if needed. Throws if an event's wgts_ is shorter
+// than lhe.weight_ids (ie the reweighting loop didn't populate that weight
+// id for every event).
+void save_weights_binary(const REX::lhe& lhe, const std::string& path_prefix);
+
+// Streaming counterpart to save_weights_binary(): builds a REX::eventSink
+// that writes one weight-only binary EventFile per entry of weight_ids
+// (same "<path_prefix>_<weight_id>.npy" convention, same
+// madspace::weight_file_layout), one event's worth of rows at a time via
+// write_event(), rather than requiring the whole REX::lhe to already be
+// fully populated in memory the way save_weights_binary() does. Intended
+// for Driver::run_streaming()'s binary_output() path -- weight_ids must
+// already be final by the time this is called (write_open() only opens the
+// files this constructs from weight_ids up front; it can't discover more
+// weight ids later). write_event() throws if a given event's wgts_ is
+// shorter than weight_ids.
+REX::eventSink open_weights_sink_binary(
+    const std::vector<std::string>& weight_ids, const std::string& path_prefix
+);
+
+// Reads back what save_weights_binary(lhe, path_prefix) wrote for exactly
+// weight_ids (same order, same path_prefix), appending one entry to every
+// events[i]->wgts_ per weight id, in weight_ids order -- ie after this
+// call, events[i]->wgts_[k] is weight_ids[k]'s value for events[i], for
+// every k. events[i]->weight_ids is left untouched; call
+// REX::lhe::set_weight_ids()/sync_weight_ids() (or just assign it directly)
+// separately once the surrounding REX::lhe is assembled. Throws if a
+// "<path_prefix>_<weight_id>.npy" file is missing, or its row count doesn't
+// match events.size().
+void load_weights_binary(
+    std::vector<std::shared_ptr<REX::event>>& events,
+    const std::vector<std::string>& weight_ids,
+    const std::string& path_prefix
+);
+
+// Builds a REX::eventSink that writes full events (not just weights, unlike
+// save_weights_binary()) to a fresh madspace binary EventFile at path, laid
+// out the way load_lhe_binary() expects to read them back
+// (EventRecord::f_lhe_event | ParticleRecord::f_lhe_particle) -- so a shard
+// written by this sink is, on its own, a valid load_lhe_binary() source.
+// Intended for use with REX::split_source() to shard a large reweighting
+// job's input across independent per-shard jobs (see madtrex::split_events()),
+// always writing the binary format for the shards regardless of the
+// original source's own format.
+//
+// The file's particle-count is fixed once and for all the first time
+// write_event() is actually called (EventFile's own on-disk shape can't
+// grow afterward): if particle_count is left at 0, it is taken from that
+// first event's own particle count; passed explicitly, it instead fixes an
+// upper bound up front (useful to guarantee sibling shards end up with a
+// matching, or at least a known, capacity). Either way, every event this
+// sink is asked to write from then on must have at most that many
+// particles -- fewer are padded with zero-filled, status_code==0 particles
+// up to it (mirroring the padding load_lhe_binary() already strips back
+// out on read), more throws. write_open()'s initNode/header are accepted
+// (so this satisfies REX::eventSink) but otherwise ignored: the binary
+// format carries no init-block information, same as load_lhe_binary()'s
+// own caveat. If write_event() is never called (eg an empty shard), no
+// file is ever created on disk.
+REX::eventSink open_lhe_sink_binary(const std::string& path, std::size_t particle_count = 0);
+
+} // namespace madtrex
diff --git a/madtrex/include/madtrex/madtrex.hpp b/madtrex/include/madtrex/madtrex.hpp
new file mode 100644
index 0000000000..6a224d636e
--- /dev/null
+++ b/madtrex/include/madtrex/madtrex.hpp
@@ -0,0 +1,572 @@
+/* ################################################### */
+/* #                                                 # */
+/* #     ___  ___          _ _  ______               # */
+/* #     |  \/  |         | | | | ___ \              # */
+/* #     | .  . | __ _  __| | |_| |_/ /_____  __     # */
+/* #     | |\/| |/ _` |/ _` | __|    // _ \ \/ /     # */
+/* #     | |  | | (_| | (_| | |_| |\ \  __/>  <      # */
+/* #     \_|  |_/\__,_|\__,_|\__\_| \_\___/_/\_\     # */
+/* #                                                 # */
+/* #    ******************************************   # */
+/* #    * MadGraph7 teaRex Reweighting executors *   # */
+/* #    ******************************************   # */
+/* ################################################### */
+//
+// MadtRex is the primary MadGraph7 driver for data-parallel
+// event reweighting, using the Rex/teaRex libraries wrapper
+// with the TupperWare and WareHouse classes to simplify the
+// interface between MadGraph7 and Rex/teaRex.
+// MadtRex provides immediate access to SLHA parameter rewegighting
+// with UMAMI-compliant matrix element libraries through the 
+// MadtRex driver type, which can load data directly from
+// provided paths to relevant process information, i.e. LHE
+// files, reweighting cards, SLHA parameter cards, and the
+// utilised scattering amplitude libraries.
+// The driver is minimal, only wrapping the REX::tea::reweightor
+// and TupperWare types with necessary plumbing and bureaucracy
+// to utilise scattering amplitudes safely.
+
+#pragma once
+#include "binary_io.hpp"
+#include "param_handler.hpp"
+#include "tupperware.hpp"
+
+#include <nlohmann/json.hpp>
+
+#include <cstddef>
+#include <string>
+#include <vector>
+
+using json = nlohmann::json;
+
+namespace madtrex {
+
+using namespace madspace;
+
+    // Which on-disk representation a set of reweighting events came from, or
+    // should be written back out as: LHEF XML, or MadGraph7's default
+    // internal binary EventFile format (io.hpp).
+    enum class LheFormat { xml, binary };
+
+    // Peeks at path's first bytes to tell a madspace binary EventFile (NumPy
+    // magic "\x93NUMPY", see io.cpp's write_event_header) apart from LHEF
+    // XML, without relying on the file extension. Throws if path cannot be
+    // opened.
+    LheFormat detect_lhe_format(const std::string& path);
+
+    std::shared_ptr<REX::lhe> load_lhe_xml(const std::string& path);
+    std::shared_ptr<REX::lhe> load_lhe_xml(std::istream& stream);
+    // Dispatches to load_lhe_xml or madtrex::load_lhe_binary based on
+    // detect_lhe_format(path).
+    std::shared_ptr<REX::lhe> load_lhe(const std::string& path);
+
+    // ---------------------------------------------------------------------
+    // Splitting a job across independent shards, and recombining them
+    // ---------------------------------------------------------------------
+    //
+    // Large reweighting jobs can be split into n_shards independent pieces
+    // (eg to run one per cluster node) via split_events(), each piece
+    // reweighted completely independently (via the ordinary Driver API,
+    // pointed at that one shard -- no new API needed for that half), and
+    // the n_shards results stitched back together via combine_shards(),
+    // which is also where -- and the *only* place where -- the combined
+    // sample's cross section and error actually get computed.
+
+    // The result of split_events(): one binary EventFile path per shard
+    // (empty if that shard ended up with zero events, eg n_shards exceeded
+    // the source's event count -- no file is created for it), and the
+    // event count actually written to each, both in shard order.
+    struct SplitResult {
+        std::vector<std::string> paths;
+        std::vector<std::size_t> counts;
+    };
+
+    // Splits the event source at path (auto-detected XML or MadGraph7's
+    // internal binary EventFile via detect_lhe_format()) into n_shards
+    // separate binary EventFile shards on disk, named
+    // "<output_prefix>_shard<i>.npy" for i in [0, n_shards) -- always
+    // binary for the shards regardless of path's own format, via
+    // REX::split_source() + open_lhe_sink_binary(), so each one is
+    // independently a load_events()-ready input for its own, completely
+    // separate Driver/job (eg one per cluster node) reweighting only that
+    // shard. path is streamed directly either way (REX::open_lhe_source()
+    // for XML, open_lhe_source_binary() for binary; meta is only consulted
+    // for the latter, matching load_events()), so memory use stays
+    // O(batch_size) regardless of path's total size. particle_count bounds
+    // every shard's fixed per-event particle-slot count (see
+    // open_lhe_sink_binary()); left at 0, each shard determines its own
+    // from whichever event lands there first.
+    //
+    // This only ever splits *input* events, never weights: recombining
+    // per-shard reweighting results (and computing the combined sample's
+    // cross section/error exactly once, over the totality) is
+    // combine_shards()'s job.
+    SplitResult split_events(
+        const std::string& path,
+        const std::string& output_prefix,
+        std::size_t n_shards,
+        std::size_t particle_count = 0,
+        std::size_t batch_size = 10000,
+        const madspace::LHEMeta& meta = {}
+    );
+
+    // Recombines n_shards shards previously produced by split_events(),
+    // each independently reweighted (eg via a Driver::reweight()/
+    // write_weights() call per shard, with binary_output() left true --
+    // the natural, and only currently supported, output for a per-shard
+    // job in this workflow), back into a single result: for shard i, loads
+    // its original events from shard_event_paths[i] (via
+    // load_lhe_binary()) and its reweighted columns from
+    // "<shard_weight_prefixes[i]>_<weight_id>.npy" per entry of
+    // launch_names, in order (via load_weights_binary()), reattaches them
+    // as that shard's events' wgts_, then appends all n_shards shards'
+    // events, in order, into one combined REX::tea::reweightor.
+    //
+    // Cross sections and errors (reweightor::calc_xSecs()/calc_xErrs())
+    // are computed exactly once here, over the full recombined sample --
+    // this is the *only* place in the split/reweight/recombine workflow
+    // that does so. A per-shard job must never compute (or, worse, write
+    // out) its own cross-section/error estimate over just its own
+    // fraction of the events: unlike a per-event weight, cross section
+    // and (especially) its error do not additively decompose across an
+    // arbitrary partition of the sample, so combining separately-computed
+    // per-shard estimates afterward would silently be statistically
+    // wrong. Using save_weights_binary() as every shard's own output
+    // format (as this workflow requires) already sidesteps the temptation
+    // structurally -- it never writes xSec_/xSecErr_/rwgt_xSec/rwgt_xErr
+    // at all, only raw per-event weight columns.
+    //
+    // Writes the combined result to output_path as LHEF XML
+    // (REX::write_lhef) unless binary_output is true, in which case
+    // save_weights_binary() is used instead, the same convention
+    // Driver::write_weights() follows. meta supplies init-block
+    // information the same way it does for load_events()/load_lhe_binary()
+    // (the binary format itself carries none); consulted once, for
+    // shard_event_paths[0], not per shard. Returns the combined reweightor
+    // (already finalised: calc_xSecs()/calc_xErrs() already run) for
+    // callers that want to inspect rwgt_xSec/rwgt_xErr/norm_factor
+    // themselves afterward.
+    REX::tea::reweightor combine_shards(
+        const std::vector<std::string>& shard_event_paths,
+        const std::vector<std::string>& shard_weight_prefixes,
+        const std::vector<std::string>& launch_names,
+        const std::string& output_path,
+        bool binary_output = false,
+        const madspace::LHEMeta& meta = {}
+    );
+
+    // ---------------------------------------------------------------------
+    // MadMatrix/UMAMI subprocess JSON loading
+    // ---------------------------------------------------------------------
+    //
+    // MadMatrix compiles each physical subprocess of a MadGraph7 process (eg
+    // the "gg > ttx" piece of "p p > 2j") into its own UMAMI-compliant matrix
+    // element library, and describes the resulting set of subprocesses in a
+    // PROCESS_DIRECTORY/SubProcesses/subprocesses.json file: a JSON array with
+    // one entry per subprocess, giving its incoming/outgoing PDG codes, the
+    // (device-templated) path to its compiled library, the flavor channels it
+    // covers (eg one library often serves several external quark flavors, up
+    // to the color/coupling factors UMAMI computes internally), and the full
+    // list of helicity combinations it was generated for.
+    //
+    // The types and functions below parse that file and turn it into
+    // REX::tea-ready building blocks: SubProcessSpec is the parsed form of one
+    // entry, make_event_checker() builds the flavor/helicity-aware
+    // REX::eventBelongs for it, resolve_me_path() finds the concrete library
+    // file the device-templated me_path refers to, and MadtRex ties all of
+    // this together with a madspace::Context and a madtrex::WareHouse into a
+    // single driver for one process directory.
+
+    // One entry of a subprocess's JSON "flavors" array, expanded so that
+    // mirrored entries (incoming beams swapped, eg q qbar vs qbar q sharing one
+    // library) become a second, explicit FlavorChannel instead of needing
+    // special-casing at match time. `pdgs` is the exact, positional external-leg
+    // PDG tuple expected of an event's own particle slot order (incoming legs,
+    // in event order, followed by outgoing legs, in event order) for this
+    // specific variant; `mirrored` records whether this variant is the
+    // incoming-swapped copy of a JSON option, so per-event helicity lookups
+    // know to undo the swap before consulting SubProcessSpec::helicities, which
+    // is always recorded in the un-mirrored leg order.
+    struct FlavorChannel {
+        int index = -1;
+        std::vector<long int> pdgs;
+        bool mirrored = false;
+    };
+
+    // Parsed form of one entry of a subprocesses.json array: one physical
+    // MadMatrix subprocess, carrying enough information to load its matrix
+    // element library and build a flavor/helicity-aware event checker for it.
+    // Phase-space-only JSON fields (channels, color_flows, pdg_color_types) are
+    // intentionally not retained here: they describe multichannel phase-space
+    // generation, not the reweighting-facing contract between an event and a
+    // UMAMI matrix element call.
+    struct SubProcessSpec {
+        std::vector<long int> incoming;
+        std::vector<long int> outgoing;
+        std::string me_path_template; // eg "lib/libmadmatrix_P1_gg_ttx_{device}.so", relative to PROCESS_DIRECTORY
+        std::string source_path;      // JSON "path" field; informational only, may be stale if the process directory was moved since generation
+        std::size_t diagram_count = 0;
+        std::vector<FlavorChannel> flavors;       // expanded, incl. mirror variants
+        std::vector<std::vector<int>> helicities; // as given in JSON, un-mirrored leg order
+    };
+
+    // Parses a MadMatrix subprocesses.json file (eg
+    // PROCESS_DIRECTORY/SubProcesses/subprocesses.json) into one
+    // SubProcessSpec per top-level array entry.
+    std::vector<SubProcessSpec> parse_subprocesses_json(const std::string& subprocesses_json_path);
+
+    // Builds the flavor/helicity-aware REX::eventBelongs for a single
+    // SubProcessSpec: one template event per distinct (pre-mirror-expansion)
+    // flavor option, matched via the default external_legs_comparator (ie PDG
+    // multiset per status, which mirroring doesn't change), with
+    // flavor_index/helicity_index hash functions that positionally re-match
+    // the specific (possibly mirrored) variant to report the UMAMI flavor
+    // index and, if the event carries genuine per-particle helicities in its
+    // spin_ fields (LHEF-style +-1; anything else is treated as "no fixed
+    // helicity"), the position of that helicity combination in
+    // SubProcessSpec::helicities. Events without fixed helicities are left
+    // with helicity_ unset (REX::npos truncated into event::helicity_'s -1
+    // sentinel), meaning the matrix element should sum over helicities.
+    REX::eventBelongs make_event_checker(const SubProcessSpec& spec);
+
+    // Given PROCESS_DIRECTORY and a me_path template containing literal
+    // "{device}" and/or "{fptype}" placeholders (see me_lib_format in
+    // MadMatrix's own madmatrix/output.py, and MADMATRIX_LIB in
+    // madmatrix.mk), lists whichever concrete files actually exist on disk
+    // (eg "libmadmatrix_sm_P1_gg_ttx_cppnone_m.so") and resolves which one
+    // to use via device_priority/fptype_priority: device is resolved
+    // first -- the first device_priority entry with a matching file wins;
+    // if none match but exactly one device variant was found (for any
+    // fptype), that one is used regardless -- then, among files matching
+    // the chosen device, fptype is resolved the same way via
+    // fptype_priority. A template with only one of the two placeholders
+    // (or neither, eg an already-fully-resolved path) resolves that one
+    // placeholder alone, or is returned as-is if it exists. Throws (no
+    // match at any stage, or several found and no priority set) listing
+    // what was found, so the caller can call
+    // Driver::set_device_priority(...)/set_fptype_priority(...) and try
+    // again.
+    std::string resolve_me_path(
+        const std::string& process_directory,
+        const std::string& me_path_template,
+        const std::vector<std::string>& device_priority,
+        const std::vector<std::string>& fptype_priority = {}
+    );
+
+    // Driver is the top-level driver tying together JSON-described MadMatrix
+    // subprocesses, the madspace::Context that loads and owns their UMAMI
+    // matrix element libraries, and the madtrex::WareHouse of
+    // REX::tea::procReweightors built from them. One Driver instance
+    // corresponds to one MadGraph7 process directory (eg PROCMG7_sm_2):
+    // load_process_directory()/load_process() may be called more than once to
+    // register subprocesses from several subprocesses.json files into the same
+    // WareHouse before calling load_events() (or warehouse().build(lhe)
+    // directly, if the events are already in hand as a REX::lhe rather than
+    // on disk).
+    class Driver {
+    public:
+        Driver();
+        explicit Driver(madspace::ContextPtr context);
+
+        // Priority order used to disambiguate multiple {device}-suffixed
+        // library variants found for the same subprocess (eg prefer
+        // "cppavx2" over "cppnone" once vectorized kernels exist). Earlier
+        // entries win. Empty by default, meaning: auto-use a lone match,
+        // throw on ambiguity.
+        Driver& set_device_priority(std::vector<std::string> priority);
+        const std::vector<std::string>& device_priority() const;
+
+        // As device_priority(), but disambiguates multiple {fptype}-suffixed
+        // library variants found for the same subprocess (see MADMATRIX_LIB
+        // in madmatrix.mk, SUPPORTED_FPTYPES therein for the full set of
+        // legal values: "d", "f", "m"), among whichever ones exist for the
+        // already-resolved device (see resolve_me_path()). Defaults to
+        // {"m", "d", "f"}, matching madmatrix.mk's own FPTYPE=m default; an
+        // explicitly empty priority means: auto-use a lone match, throw on
+        // ambiguity.
+        Driver& set_fptype_priority(std::vector<std::string> priority);
+        const std::vector<std::string>& fptype_priority() const;
+
+        // Priority order used to resolve which {fptype}-suffixed library
+        // variant load_process()/load_process_directory() loads for each
+        // subprocess's *normaliser* (REX::tea::procReweightor's own
+        // normalisation weightor -- see TupperWare::set_normaliser()), kept
+        // deliberately independent of fptype_priority() (used for the
+        // reweight-*target* api instead). This is a fundamental assumption
+        // of how reweighting works here: every launch must be normalised
+        // against the sample's actual, original configuration -- never
+        // against whatever it's being reweighted TO -- exactly as ordinary
+        // physics-parameter reweighting already does (normaliser = the
+        // as-loaded param_card, untouched by any launch's "set BLOCK ID
+        // VALUE" override, which only ever gets applied to the reweight-
+        // target api via ParamHandler). A change of *fptype* additionally
+        // needs a differently-*compiled* library for that role, which is
+        // exactly what this priority controls.
+        //
+        // Left unset (the default) means: reuse fptype_priority() for the
+        // normaliser too, ie load_process() shares one api between both
+        // roles -- correct and harmless when a card never changes fptype
+        // (the sample's own fptype IS the target fptype in that case), and
+        // preserves this class's pre-existing behavior for every caller
+        // that never calls this setter. Call this explicitly with the
+        // sample's *original* fptype (madtrex.py's run_parameter_
+        // reweighting() reads this from the process directory's own
+        // run_card.toml [run] fptype) whenever fptype_priority() targets a
+        // *different* fptype than the sample was actually generated at --
+        // load_process() then loads a second, independent api for the
+        // normaliser role, deliberately never registered via add_api()
+        // (see its own implementation comment), so ParamHandler's per-
+        // launch parameter overrides can never touch it.
+        Driver& set_normaliser_fptype_priority(std::vector<std::string> priority);
+        const std::vector<std::string>& normaliser_fptype_priority() const;
+
+        // Helicity reweighting (true, the default) or helicity summation
+        // (false) for subprocesses loaded by load_process()/
+        // load_process_directory() from here on (it does not retroactively
+        // affect subprocesses already loaded). With helicity reweighting,
+        // each event is reweighted at its own helicity: the matrix elements
+        // are evaluated for the helicity combination of the event's LHE
+        // SPINUP column (UMAMI_IN_HELICITY_INDEX) rather than summed over
+        // helicities. A matrix element without per-event helicity support
+        // falls back to helicity summation with a one-time warning on stderr;
+        // a sample without helicity information (any particle with a SPINUP
+        // other than -1, 0, +1, eg 9) is an error -- reweight it with
+        // helicity_reweighting off. See default_input_keys() and
+        // input_keys_for().
+        Driver& set_helicity_reweighting(bool helicity_reweighting);
+        bool helicity_reweighting() const;
+        // The process directory (and param_card, default its Cards/
+        // param_card.dat) whose matrix elements subprocesses loaded from
+        // here on are normalised against, instead of their own: the
+        // original process of a "change model" reweighting, whose events
+        // were generated with it. The weights are then the new model's
+        // matrix element over the original one, MadGraph's reweighting
+        // convention. Every subprocess loaded must have an identical
+        // counterpart there (same legs, flavor channels and helicity
+        // table), or load_process() throws; its library is the one
+        // normaliser_fptype_priority() resolves, and it is evaluated the
+        // same way as the reweighted one (both at the event's helicity, or
+        // both helicity-summed). Empty (the default): the loaded process
+        // directory's own.
+        Driver& set_normaliser_process_directory(
+            const std::string& process_directory, const std::string& param_card = ""
+        );
+        const std::string& normaliser_process_directory() const;
+        // The param_card the normaliser is loaded with: the one the events
+        // were generated with, when the param_card the reweighted matrix
+        // elements start from (load_process()'s) is another one. Empty (the
+        // default): with a normaliser process directory, its Cards/
+        // param_card.dat, else load_process()'s param_card. Also set by
+        // set_normaliser_process_directory().
+        Driver& set_normaliser_param_card(const std::string& param_card);
+        const std::string& normaliser_param_card() const;
+
+        // Whether subprocesses loaded from here on normalise against the
+        // squared matrix elements the events were generated with, when the
+        // event file stores them (the "matrix_element" column of a madspace
+        // binary event file) and they are the quantity the normaliser
+        // computes, instead of evaluating the normaliser again (see
+        // TupperWare::set_stored_normalisation()). True by default; turn it
+        // off to always evaluate the normaliser.
+        Driver& set_use_stored_normalisation(bool use_stored_normalisation);
+        bool use_stored_normalisation() const;
+        // Former names of set_helicity_reweighting()/helicity_reweighting().
+        Driver& set_use_helicity(bool use_helicity);
+        bool use_helicity() const;
+
+        // The muted-by-default warning logger shared with the built
+        // WareHouse/reweightor (see WareHouse::verbose()). Call
+        // verbose().set_verbose(true) to unmute.
+        REX::verbosity& verbose();
+        const REX::verbosity& verbose() const;
+        // Plain-bool convenience for verbose().set_verbose(v)/is_verbose().
+        Driver& set_verbose(bool v);
+        bool is_verbose() const;
+
+        // Maximum number of events per REX::process (see
+        // WareHouse::set_max_batch()/REX::lhe::max_batch), applied by
+        // load_events()/warehouse().build(). Left at REX::npos (unset) by
+        // default. Setting this also makes every subprocess loaded via
+        // load_process()/load_process_directory() reload its own
+        // MatrixElementApi instance per batch instead of every max_batch
+        // overflow clone sharing (and serializing calls through) the one
+        // instance that was loaded for it -- see TupperWare::build() and
+        // ReloadSpec.
+        Driver& set_max_batch(std::size_t max_batch);
+        std::size_t max_batch() const;
+
+        // Worker threads and the minimum events per batch when splitting
+        // the sample across them: see WareHouse::set_threads()/
+        // set_min_batch()/batch_size_for(). Applied by load_events().
+        Driver& set_threads(std::size_t threads);
+        std::size_t threads() const;
+        Driver& set_min_batch(std::size_t min_batch);
+        std::size_t min_batch() const;
+
+        // Passed straight through to WareHouse::set_max_deferred_iterations()
+        // (see its own doc comment): caps how many outer reweighting
+        // iterations' worth of deferred (eg GPU) results are allowed to
+        // accumulate device-side, per process, before an automatic mid-run
+        // flush back to the host. REX::npos (the default) means no cap --
+        // deferred results are only ever copied back once, at the end of
+        // the whole run, same as before this existed.
+        Driver& set_max_deferred_iterations(std::size_t max_deferred_iterations);
+        std::size_t max_deferred_iterations() const;
+
+        // Enables Rex's streaming mode for the next load_events() call:
+        // rather than loading the whole event file into memory up front,
+        // load_events() opens a REX::eventSource for path and builds the
+        // WareHouse's reweightor via REX::lhe::open_streaming(), so events
+        // are pulled and processed batch_size at a time. Off by default.
+        // Only supported for LHEF XML event sources (REX::open_lhe_source);
+        // load_events() throws if streaming is enabled and path is detected
+        // as MadGraph7's internal binary EventFile format. Once enabled,
+        // run the reweighting via run_streaming() (or reweight()'s
+        // end-to-end convenience) instead of warehouse().get().run() +
+        // write_weights(): streamed output is written incrementally to a
+        // sink as each batch finishes, not held for a later
+        // write_weights() call.
+        Driver& set_streaming(bool enable, std::size_t batch_size = 10000);
+        bool streaming() const;
+        std::size_t stream_batch_size() const;
+
+        // Runs the streaming reweightor built by load_events() (see
+        // set_streaming()) to completion, writing events incrementally to
+        // output_path as an LHEF XML sink (REX::open_lhe_sink) unless
+        // binary_output() is true (set automatically by load_events() when
+        // it detected a binary source), in which case output_path is
+        // instead used as a path_prefix for one streamed weight-only
+        // binary file per launch name (open_weights_sink_binary()), the
+        // same "<output_path>_<weight_id>.npy" convention
+        // save_weights_binary()/write_weights() use -- this requires
+        // warehouse().launch_names() to be non-empty (eg via
+        // load_param_reweighting() or warehouse().set_launch_names()),
+        // since binary weight files are named per weight id; throws
+        // otherwise. Throws if streaming() is false, or the WareHouse
+        // hasn't been built yet.
+        Driver& run_streaming(const std::string& output_path, bool include_ids = false);
+
+        // Parses subprocesses_json_path, loads each entry's matrix element
+        // library (resolving {device} against what's actually on disk) via
+        // the owned Context, and registers one TupperWare (api +
+        // flavor/helicity-aware event checker) per entry into the owned
+        // WareHouse. param_card defaults to
+        // PROCESS_DIRECTORY/Cards/param_card.dat, where PROCESS_DIRECTORY is
+        // taken to be the grandparent directory of subprocesses_json_path (ie
+        // .../PROCESS_DIRECTORY/SubProcesses/subprocesses.json).
+        Driver& load_process(const std::string& subprocesses_json_path, const std::string& param_card = "");
+        // Convenience: process_directory/SubProcesses/subprocesses.json.
+        Driver& load_process_directory(const std::string& process_directory, const std::string& param_card = "");
+
+        // Configures UMAMI-direct parameter reweighting: builds (or
+        // replaces) the owned ParamHandler over every matrix element api
+        // registered so far (ie everything load_process()/
+        // load_process_directory() has loaded up to this call -- call this
+        // after those, not before), parses rwgt_path, and wires the
+        // resulting iterators/launch_names onto the WareHouse via
+        // warehouse().set_iterators()/set_launch_names(). Must be called
+        // before load_events() (or warehouse().build()), since that is what
+        // actually captures the WareHouse's current iterators/launch_names
+        // into the built reweightor. Requires every registered api to
+        // implement umami_set_parameter (see ParamHandler::add_api); throws
+        // otherwise -- for libraries that don't, build the iterators from
+        // REX::tea::param_rwgt's file-based approach instead and pass them to
+        // warehouse().set_iterators() directly.
+        Driver& load_param_reweighting(const std::string& rwgt_path);
+
+        ParamHandler& param_handler();
+        const ParamHandler& param_handler() const;
+
+        // Loads the event source the reweighting will run against: sniffs
+        // path via detect_lhe_format() to decide between MadGraph7's
+        // internal binary EventFile format and LHEF XML, loads it
+        // accordingly, and builds the WareHouse's REX::tea::reweightor
+        // against the result (ie warehouse().build(*load_lhe(path))).
+        // meta supplies init-block information (beam ids/energies, xsec,
+        // PDF set) for the binary path, which -- unlike LHEF -- stores none
+        // of that on disk; ignored when path is XML, whose own <init> block
+        // is used instead. Also sets binary_output() to match the detected
+        // input format, so write_weights() round-trips to the same
+        // representation by default; call set_binary_output() afterwards to
+        // override. May be called again to rebuild against a different
+        // event source, discarding any previously-built reweightor (see
+        // WareHouse::build()).
+        //
+        // If set_streaming(true) was called, this instead opens a streaming
+        // REX::eventSource for path (which must be LHEF XML; see
+        // set_streaming()) and builds against a REX::lhe::open_streaming()
+        // lhe, without pulling any events yet -- use run_streaming() rather
+        // than warehouse().get().run() + write_weights() to actually run
+        // the reweighting in that case.
+        Driver& load_events(const std::string& path, const madspace::LHEMeta& meta = {});
+
+        // Which format write_weights() targets: true for
+        // madtrex::save_weights_binary() (one weight-only binary EventFile
+        // per weight id), false for a full LHEF XML rewrite via
+        // REX::write_lhef() (events plus every recorded weight). Defaults to
+        // whatever load_events() last detected; false if load_events() was
+        // never called.
+        bool binary_output() const;
+        Driver& set_binary_output(bool binary);
+
+        // Writes out whatever weights are currently recorded on the built
+        // reweightor's events (ie warehouse().get(), typically after
+        // REX::tea::reweightor::run()) via save_weights_binary(path) or
+        // REX::write_lhef(path), according to binary_output(). Throws if
+        // the WareHouse hasn't been built yet (see load_events()).
+        Driver& write_weights(const std::string& path);
+
+        // End-to-end convenience running the entire pipeline in one call:
+        // load_process_directory(process_directory, param_card), then --
+        // if rwgt_path is non-empty -- load_param_reweighting(rwgt_path),
+        // then load_events(events_path, meta), then
+        // warehouse().get().run() (REX::tea::reweightor's full setup/
+        // run_all_iterations/finalise_reweighting loop), then
+        // write_weights(output_path). rwgt_path may be left empty to skip
+        // parameter reweighting (eg when every process's WareHouse entry
+        // already fully determines its own reweighting via
+        // add_api()/set_normaliser() and needs no per-iteration parameter
+        // changes at all). Returns *this so eg driver.reweight(...).warehouse()
+        // stays chainable for inspecting the result afterwards.
+        Driver& reweight(
+            const std::string& process_directory,
+            const std::string& events_path,
+            const std::string& output_path,
+            const std::string& rwgt_path = "",
+            const std::string& param_card = "",
+            const madspace::LHEMeta& meta = {}
+        );
+
+        madspace::ContextPtr context() const;
+        WareHouse& warehouse();
+        const WareHouse& warehouse() const;
+        const std::vector<SubProcessSpec>& specs() const;
+
+    private:
+        madspace::ContextPtr reload_context();
+
+        madspace::ContextPtr _context;
+        madspace::ContextPtr _reload_context;
+        std::vector<std::string> _device_priority = {"cpp512z", "cpp512y", "cppavx2", "cppsse4", "cppnone"};
+        std::vector<std::string> _fptype_priority = {"m", "d", "f"};
+        // Unset (see set_normaliser_fptype_priority()'s own doc comment) by
+        // default: _normaliser_fptype_priority_set distinguishes "never
+        // called set_normaliser_fptype_priority()" from "called it with an
+        // explicitly empty priority" (itself a legal, meaningful value --
+        // "auto-use a lone match, throw on ambiguity", same as
+        // fptype_priority()'s own empty-priority meaning).
+        std::vector<std::string> _normaliser_fptype_priority;
+        bool _normaliser_fptype_priority_set = false;
+        WareHouse _warehouse;
+        std::vector<SubProcessSpec> _specs;
+        ParamHandler _param_handler;
+        bool _binary_output = false;
+        bool _helicity_reweighting = true;
+        bool _use_stored_normalisation = true;
+        std::string _normaliser_process_directory;
+        std::string _normaliser_param_card;
+        bool _streaming = false;
+        std::size_t _stream_batch_size = 10000;
+    };
+
+}
\ No newline at end of file
diff --git a/madtrex/include/madtrex/param_handler.hpp b/madtrex/include/madtrex/param_handler.hpp
new file mode 100644
index 0000000000..3ec43c6fc4
--- /dev/null
+++ b/madtrex/include/madtrex/param_handler.hpp
@@ -0,0 +1,85 @@
+// ParamHandler: UMAMI-direct SLHA parameter reweighting.
+//
+// REX::tea::param_rwgt (teaRex.h) drives SLHA parameter reweighting by
+// rewriting param_card.dat between iterations and relying on the matrix
+// element library to reread it from scratch on its next call. Where the
+// UMAMI implementation exports umami_set_parameter/umami_get_parameter (see
+// umami.h), that reread is unnecessary: the already-loaded instance's
+// independent parameters can be updated in place. ParamHandler builds
+// REX::tea::iterator objects that do exactly that, reusing
+// REX::tea::rwgt_slha only for parsing the same "# launch / # set BLOCK_NAME
+// PARAM_ID PARAM_VALUE" reweight-card format param_rwgt itself reads.
+
+#pragma once
+
+#include "tupperware.hpp"
+
+#include <istream>
+#include <string>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+namespace madtrex {
+
+class ParamHandler {
+public:
+    ParamHandler() = default;
+
+    // Registers apis this handler's iterators should set parameters on.
+    // Typically every madspace::MatrixElementApi loaded for one physical
+    // MadGraph7 process (ie every TupperWare::apis() in one WareHouse/
+    // Driver), since they share one param_card.dat and so should be kept in
+    // sync with each other. Duplicate apis (by pointer) are ignored. Throws
+    // if api does not report supports_set_parameter() -- for such libraries,
+    // use REX::tea::param_rwgt's file-based approach instead.
+    ParamHandler& add_api(const madspace::MatrixElementApi& api);
+    ParamHandler& add_apis(const std::vector<const madspace::MatrixElementApi*>& apis);
+    // Convenience: add_apis(tw.apis()) for every TupperWare in house.
+    ParamHandler& add_apis(const WareHouse& house);
+    const std::vector<const madspace::MatrixElementApi*>& apis() const { return _apis; }
+
+    // Parses a reweight card in the same format REX::tea::rwgt_slha reads
+    // (see teaRex.h). Replaces any previously parsed card.
+    ParamHandler& read_rwgt_card(std::istream& rwgt_in);
+    ParamHandler& read_rwgt_card(const std::string& rwgt_path);
+
+    std::size_t launch_count() const;
+
+    // Applies launch idx's parameter changes to every registered api via
+    // MatrixElementApi::set_parameter(), first resetting every parameter any
+    // apply_launch() call on this handler has previously touched back to the
+    // value it had the first time this handler read it (its "nominal"
+    // value), so each launch is an independent perturbation from that
+    // baseline rather than compounding on top of the previous launch's
+    // changes -- mirrors REX::tea::rwgt_slha::write_rwgt_card's own
+    // orig_params restore-then-clear step. Returns true on success; matches
+    // REX::tea::iterator's signature so it can be wrapped directly (see
+    // iterators()). Throws if idx is out of range or no apis are registered.
+    bool apply_launch(std::size_t idx);
+
+    // One REX::tea::iterator per parsed launch, in card order, each calling
+    // apply_launch() for its own index; ready to hand to
+    // WareHouse::set_iterators() (or REX::tea::reweightor::set_iterators()
+    // directly). Each iterator captures this by pointer -- the ParamHandler
+    // must outlive them, exactly as REX::tea::rwgt_slha::get_card_writers()
+    // requires of its own rwgt_slha.
+    std::vector<REX::tea::iterator> iterators();
+    // Forwarded from the underlying REX::tea::rwgt_slha; see
+    // rwgt_slha::get_launch_names()/get_rwgt_commands().
+    std::vector<std::string> launch_names();
+    std::vector<std::string> weight_context();
+
+private:
+    std::vector<const madspace::MatrixElementApi*> _apis;
+    REX::tea::rwgt_slha _card_iter;
+    // (api, "BLOCK_NAME PARAM_ID") -> value found the first time this
+    // handler read that pair, ie the un-reweighted baseline every launch's
+    // changes are applied relative to.
+    std::unordered_map<
+        const madspace::MatrixElementApi*,
+        std::unordered_map<std::string, std::pair<double, double>>>
+        _nominal;
+};
+
+} // namespace madtrex
diff --git a/madtrex/include/madtrex/pdfware.hpp b/madtrex/include/madtrex/pdfware.hpp
new file mode 100644
index 0000000000..efea2b03ef
--- /dev/null
+++ b/madtrex/include/madtrex/pdfware.hpp
@@ -0,0 +1,353 @@
+/* ################################################### */
+/* #                                                 # */
+/* #     ___  ___          _ _  ______               # */
+/* #     |  \/  |         | | | | ___ \              # */
+/* #     | .  . | __ _  __| | |_| |_/ /_____  __     # */
+/* #     | |\/| |/ _` |/ _` | __|    // _ \ \/ /     # */
+/* #     | |  | | (_| | (_| | |_| |\ \  __/>  <      # */
+/* #     \_|  |_/\__,_|\__,_|\__\_| \_\___/_/\_\     # */
+/* #                                                 # */
+/* #    ******************************************   # */
+/* #    *    MadGraph7 internal PDF reweighting   *   # */
+/* #    ******************************************   # */
+/* ################################################### */
+//
+// pdfware wraps madspace's internal PDF interpolators (PdfGrid/PartonDensity
+// and AlphaSGrid/RunningCoupling, madspace/phasespace/pdf.hpp) into REX::tea
+// weightors, so that PDF-only reweighting can run through the same
+// REX::tea::reweightor machinery madtrex::Driver/TupperWare already use
+//
+
+#pragma once
+
+#include "madtrex.hpp"
+
+#include "madspace/driver/backend.hpp"
+#include "madspace/phasespace/pdf.hpp"
+
+#include <memory>
+#include <optional>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+namespace madtrex {
+
+// pids fixes, once and for all, which incoming-parton PDG codes this PdfSet
+// can be evaluated for; evaluating at a PDG code outside that list throws. 
+// default_pids() uses gluon + 5-flavor set unless caller overrides it.
+class PdfSet {
+public:
+    PdfSet(
+        madspace::ContextPtr context,
+        const std::string& grid_file,
+        std::vector<int> pids,
+        const std::string& alpha_s_grid_file,
+        const std::string& prefix
+    );
+
+    PdfSet(
+        madspace::ContextPtr context,
+        const madspace::PdfGrid& grid,
+        std::vector<int> pids,
+        const madspace::AlphaSGrid* alpha_s_grid,
+        const std::string& existing_globals_prefix,
+        const std::string& label
+    );
+
+    bool has_alpha_s() const { return static_cast<bool>(_alpha_s_runtime); }
+    const std::vector<int>& pids() const { return _pids; }
+
+    // Evaluates the parton density for `x.size()` independent (x, q, pdg)
+    // triples in one batched call. x/q/pdgs must be the same size; every
+    // entry of pdgs must be one of pids(), or this throws.
+    std::vector<double> eval_pdf(
+        const std::vector<double>& x,
+        const std::vector<double>& q,
+        const std::vector<long int>& pdgs
+    ) const;
+
+    // Evaluates alpha_s at each entry of q. Throws if has_alpha_s() is
+    // false (this PdfSet was constructed without an alpha_s_grid_file).
+    std::vector<double> eval_alpha_s(const std::vector<double>& q) const;
+
+private:
+    madspace::ContextPtr _context;
+    std::string _label;
+    std::vector<int> _pids;
+    std::unordered_map<long int, madspace::me_int_t> _pid_index;
+
+    std::optional<madspace::PartonDensity> _density;
+    madspace::RuntimePtr _density_runtime;
+    std::optional<madspace::RunningCoupling> _alpha_s;
+    madspace::RuntimePtr _alpha_s_runtime;
+};
+
+// A scale (muF or muR) to evaluate a PdfSet/alpha_s at, either exactly (an
+// "absolute" value in GeV, the same for every event regardless of what it
+// actually generated at) or relative to each event's own recorded scale (a
+// multiplicative "factor", eg 2.0 for twice the event's own muF/muR -- the
+// pre-existing, and still default, behaviour).
+//
+// The single-argument constructor is intentionally not explicit: a bare
+// double (eg passed to PdfDriver::set_baseline()/add_hypothesis(), which
+// take ScaleChoice by value) is always taken as a *factor*, so every
+// existing call site that used to pass a plain muF_scale_factor/
+// muR_scale_factor double keeps compiling and behaving identically.
+struct ScaleChoice {
+    double value = 1.0;
+    bool absolute = false;
+
+    ScaleChoice() = default;
+    ScaleChoice(double v) : value(v), absolute(false) {}
+    ScaleChoice(double v, bool is_absolute) : value(v), absolute(is_absolute) {}
+
+    static ScaleChoice factor(double v) { return ScaleChoice(v, false); }
+    static ScaleChoice fixed(double v) { return ScaleChoice(v, true); }
+
+    // Combines this scale with a further multiplicative factor, eg a seven-
+    // point envelope's 0.5/1/2 (see PdfDriver::add_scale_variation_envelope()):
+    // a factor scale's own value is simply multiplied through as usual; an
+    // absolute scale's value is *also* just multiplied by extra_factor (ie
+    // "vary around this fixed central value"), staying absolute.
+    ScaleChoice scaled_by(double extra_factor) const {
+        return ScaleChoice(value * extra_factor, absolute);
+    }
+
+    // The concrete scale to evaluate at for one event whose own recorded
+    // muF_/muR_ is nominal: value itself if absolute, else nominal * value.
+    double resolve(double nominal) const { return absolute ? value : value * nominal; }
+};
+
+struct PdfHypothesis {
+    std::string launch_name;
+    std::shared_ptr<PdfSet> beam1_pdf_set;
+    std::shared_ptr<PdfSet> beam2_pdf_set;
+    ScaleChoice muF_scale{1.0, false};
+    ScaleChoice muR_scale{1.0, false};
+};
+
+// Which of a seven-point scale-variation envelope's weights (see
+// PdfDriver::add_scale_variation_envelope()) actually get written to the
+// output LHE by PdfDriver::write_weights(). Other hypotheses added via
+// add_hypothesis() are unaffected regardless of this setting -- it only
+// ever touches launch names registered through
+// add_scale_variation_envelope().
+enum class ScaleVariationOutput {
+    none,     // none of the seven (default): still computed (needed to find
+              // the envelope, and to feed rwgt_xSec/rwgt_xErr), just not written
+    envelope, // only the two whose total cross section (reweightor::rwgt_xSec)
+              // is smallest/largest among the seven
+    all       // all seven
+};
+
+// Builds a REX::tea::weightor evaluating hyp's PDF-luminosity factor --
+// pdf(x1, muF; pdg1) * pdf(x2, muF; pdg2), further multiplied by
+// alpha_s(muR)^alpha_s_power if alpha_s_power != 0 -- for every event 
+std::shared_ptr<REX::tea::weightor> make_pdf_weightor(
+    const PdfHypothesis& hyp, double beam1_energy, double beam2_energy, int alpha_s_power
+);
+
+class PdfDriver {
+public:
+    PdfDriver();
+    explicit PdfDriver(madspace::ContextPtr context);
+
+    PdfDriver& load_pdf_set(
+        const std::string& name,
+        const std::string& grid_file,
+        const std::string& alpha_s_grid_file = "",
+        std::vector<int> pids = default_pids()
+    );
+
+    // As load_pdf_set(), but reuses an already-parsed grid
+    PdfDriver& add_pdf_set(
+        const std::string& name,
+        const madspace::PdfGrid& grid,
+        const madspace::AlphaSGrid* alpha_s_grid = nullptr,
+        std::vector<int> pids = default_pids(),
+        const std::string& existing_globals_prefix = ""
+    );
+
+    PdfDriver& set_alpha_s_power(int power);
+
+    // Designates which (previously load_pdf_set()'d) PDF configuration and
+    // scale choice the loaded events' own weight belongs to. muF_scale/
+    // muR_scale each accept either a plain double (an implicit
+    // ScaleChoice::factor(), ie a multiplier of the event's own recorded
+    // scale -- unchanged pre-existing behaviour) or an explicit
+    // ScaleChoice::fixed(value) for an absolute scale in GeV. Single-name
+    // form uses the same PDF set for both beams (symmetric colliders).
+    PdfDriver& set_baseline(
+        const std::string& pdf_set_name,
+        ScaleChoice muF_scale = ScaleChoice::factor(1.0),
+        ScaleChoice muR_scale = ScaleChoice::factor(1.0)
+    );
+    // Two-name form uses independent PDF sets per beam (e.g. ep colliders, or
+    // asymmetric proton-ion running where beam 2 uses a nuclear PDF set).
+    PdfDriver& set_baseline(
+        const std::string& beam1_pdf_set_name,
+        const std::string& beam2_pdf_set_name,
+        ScaleChoice muF_scale = ScaleChoice::factor(1.0),
+        ScaleChoice muR_scale = ScaleChoice::factor(1.0)
+    );
+
+    // Single-name: same PDF set for both beams. See set_baseline() for
+    // muF_scale/muR_scale's absolute-vs-factor ScaleChoice semantics.
+    PdfDriver& add_hypothesis(
+        const std::string& launch_name,
+        const std::string& pdf_set_name,
+        ScaleChoice muF_scale = ScaleChoice::factor(1.0),
+        ScaleChoice muR_scale = ScaleChoice::factor(1.0)
+    );
+    // Two-name: independent PDF sets per beam.
+    PdfDriver& add_hypothesis(
+        const std::string& launch_name,
+        const std::string& beam1_pdf_set_name,
+        const std::string& beam2_pdf_set_name,
+        ScaleChoice muF_scale = ScaleChoice::factor(1.0),
+        ScaleChoice muR_scale = ScaleChoice::factor(1.0)
+    );
+
+    // Registers the standard seven-point scale-variation envelope against
+    // pdf_set_name: add_hypothesis() calls for every (muR_factor,
+    // muF_factor) pair in {0.5, 1, 2}^2 except the two "uncorrelated"
+    // extremes (0.5, 2) and (2, 0.5), each combined onto the given baseline
+    // muF_scale/muR_scale via ScaleChoice::scaled_by() -- ie the pair (1, 1)
+    // reduces to exactly (muF_scale, muR_scale), and this works the same way
+    // whether the baseline is a factor or an absolute value (see
+    // ScaleChoice::scaled_by()'s own doc comment). Launch names are
+    // group_name + "_muR<r>_muF<f>". Which of these seven actually get
+    // written to the output LHE is controlled by
+    // set_scale_variation_output(); all seven are always computed
+    // regardless (needed to determine the envelope in the first place).
+    // group_name must be unique among add_scale_variation_envelope() calls
+    // on this PdfDriver.
+    PdfDriver& add_scale_variation_envelope(
+        const std::string& group_name,
+        const std::string& pdf_set_name,
+        ScaleChoice muF_scale = ScaleChoice::factor(1.0),
+        ScaleChoice muR_scale = ScaleChoice::factor(1.0)
+    );
+    // Two-name: independent PDF sets per beam.
+    PdfDriver& add_scale_variation_envelope(
+        const std::string& group_name,
+        const std::string& beam1_pdf_set_name,
+        const std::string& beam2_pdf_set_name,
+        ScaleChoice muF_scale = ScaleChoice::factor(1.0),
+        ScaleChoice muR_scale = ScaleChoice::factor(1.0)
+    );
+
+    // Which of every add_scale_variation_envelope() group's seven weights
+    // actually get written to the output LHE by write_weights(). Defaults
+    // to ScaleVariationOutput::none. Applies uniformly to every registered
+    // group; other hypotheses added via add_hypothesis() are unaffected
+    // regardless.
+    PdfDriver& set_scale_variation_output(ScaleVariationOutput mode);
+    ScaleVariationOutput scale_variation_output() const;
+
+    // Fixes the beam energies used to derive each event's x1/x2 (see
+    // make_pdf_weightor), overriding whatever load_events() would otherwise
+    // take from the loaded lhe's own init block. Needed whenever that block
+    // doesn't carry real beam energies, eg MadGraph7's internal binary
+    // EventFile format: pass real beam energies here, or via meta's
+    // beam1_energy/beam2_energy at load_events() instead.
+    PdfDriver& set_beam_energies(double beam1_energy, double beam2_energy);
+
+    // Sniffs path for MadGraph7's internal binary EventFile format or LHEF
+    // XML, then builds the owned REX::tea::reweightor against the result.
+    // meta supplies init-block information (in particular beam energies)
+    // for the binary path; ignored when path is XML.
+    // Must be called after set_baseline() and every add_hypothesis() call it
+    // should cover; requires beam energies via set_beam_energies(), meta, or the
+    // loaded XML's own <init> block; otherwise throws. 
+    // May be called again to rebuild against a different event source.
+    PdfDriver& load_events(const std::string& path, const madspace::LHEMeta& meta = {});
+
+    bool binary_output() const { return _binary_output; }
+    PdfDriver& set_binary_output(bool binary);
+
+    // Before actually writing anything, drops whichever of every
+    // add_scale_variation_envelope() group's weight columns
+    // set_scale_variation_output() says shouldn't be written (see its own
+    // doc comment) from every event's wgts_ and from weight_ids -- so
+    // those columns are entirely absent from the output, not just zeroed.
+    // Idempotent: safe to call write_weights() more than once.
+    PdfDriver& write_weights(const std::string& path);
+
+    // Drops every currently-registered weight column whose id is *not* in
+    // keep_names, from both weight_ids and every event's wgts_ -- a public,
+    // name-driven generalization of write_weights()'s own
+    // add_scale_variation_envelope()-keyed filtering, for callers that
+    // build their own hypothesis set directly via add_hypothesis() (so have
+    // no add_scale_variation_envelope() group for that mechanism to key
+    // off) but still want only some of the resulting weights actually
+    // written. A name not currently present is silently ignored. Must be
+    // called after get().run() (so weight_ids/events are populated) and
+    // before write_weights(); idempotent, like write_weights()'s own
+    // filtering.
+    //
+    // Python callers in particular need this: the bound weight_ids
+    // property replaces its whole shared_ptr on assignment, which would
+    // desync it from each event's own (separately shared_ptr-aliased)
+    // weight_ids -- this method instead erases in place, the same way
+    // write_weights()'s internal filtering does, so the aliasing (and thus
+    // every event's own weight_ids) stays correct.
+    PdfDriver& keep_weights(const std::vector<std::string>& keep_names);
+
+    // End-to-end convenience: load_events(events_path, meta), run the
+    // reweighting (REX::tea::reweightor's full setup/run_all_iterations/
+    // finalise_reweighting loop), then write_weights(output_path).
+    PdfDriver& reweight(
+        const std::string& events_path,
+        const std::string& output_path,
+        const madspace::LHEMeta& meta = {}
+    );
+
+    madspace::ContextPtr context() const { return _context; }
+    // Access to the owned REX::tea::reweightor once load_events() has built
+    // it. Throws otherwise.
+    REX::tea::reweightor& get();
+    const REX::tea::reweightor& get() const;
+    bool built() const noexcept { return static_cast<bool>(_reweightor); }
+
+    // The muted-by-default warning logger shared with the built reweightor
+    // (and, transitively, its procReweightor -- see
+    // REX::tea::reweightor::extract_sorter()). Muted (false) from
+    // construction, independent of whether load_events() has been called
+    // yet; call verbose().set_verbose(true) to unmute. Re-linked into the
+    // reweightor's own verbose on every load_events()/rebuild.
+    REX::verbosity& verbose() { return _verbose; }
+    const REX::verbosity& verbose() const { return _verbose; }
+    // Plain-bool convenience for verbose().set_verbose(v)/is_verbose().
+    PdfDriver& set_verbose(bool v) {
+        _verbose.set_verbose(v);
+        return *this;
+    }
+    bool is_verbose() const { return _verbose.is_verbose(); }
+
+    // The standard gluon + 5-quark-flavor (and antiquarks) QCD-collider PDG
+    // set: {21, 1, -1, 2, -2, 3, -3, 4, -4, 5, -5}.
+    static std::vector<int> default_pids();
+
+private:
+    madspace::ContextPtr _context;
+    std::unordered_map<std::string, std::shared_ptr<PdfSet>> _pdf_sets;
+    int _alpha_s_power = 0;
+    std::optional<PdfHypothesis> _baseline;
+    std::vector<PdfHypothesis> _hypotheses;
+    // One entry per add_scale_variation_envelope() call, each the seven
+    // launch_names it registered into _hypotheses, in the same order.
+    std::vector<std::vector<std::string>> _scale_variation_groups;
+    ScaleVariationOutput _scale_variation_output = ScaleVariationOutput::none;
+    double _beam1_energy = 0.0, _beam2_energy = 0.0;
+    bool _beam_energies_set = false;
+    bool _binary_output = false;
+    REX::verbosity _verbose{false};
+    std::unique_ptr<REX::tea::reweightor> _reweightor;
+
+    const std::shared_ptr<PdfSet>& find_pdf_set(const std::string& name) const;
+    void filter_scale_variation_weights();
+};
+
+} // namespace madtrex
diff --git a/madtrex/include/madtrex/tupperware.hpp b/madtrex/include/madtrex/tupperware.hpp
new file mode 100644
index 0000000000..1712c81d08
--- /dev/null
+++ b/madtrex/include/madtrex/tupperware.hpp
@@ -0,0 +1,586 @@
+/* .---------------------------------------------------------. */
+/* | _____                          __        __             | */
+/* ||_   _|   _ _ __  _ __   ___ _ _\ \      / /_ _ _ __ ___ | */
+/* |  | || | | | '_ \| '_ \ / _ \ '__\ \ /\ / / _` | '__/ _ \| */
+/* |  | || |_| | |_) | |_) |  __/ |   \ V  V / (_| | | |  __/| */
+/* |  |_| \__,_| .__/| .__/ \___|_|    \_/\_/ \__,_|_|  \___|| */
+/* |           |_|   |_|                                     | */
+/* '---------------------------------------------------------' */
+/*       Tearex-Umami Pointer Passing Event Reweighting        */
+/*           With Accelerated Repetitive Execution             */
+//
+// TupperWare is the "simple" generic reweighting bridge between MadGraph7 and
+// the standalone Rex/teaRex libraries: it turns already-loaded, UMAMI-compliant
+// madspace::MatrixElementApi objects into REX::tea::weightors, ie plain
+// std::function<shared_ptr<vector<double>>(REX::process&)> callables, and combines
+// them with a caller-supplied event checker into a REX::tea::procReweightor.
+//
+
+#pragma once
+
+#include "teaRex.h"
+#include "madspace/driver/context.hpp"
+
+#include <deque>
+#include <functional>
+#include <memory>
+#include <mutex>
+#include <shared_mutex>
+#include <optional>
+#include <string>
+#include <vector>
+
+namespace madtrex {
+
+// Identifies how to reload a fresh, independent instance of a
+// MatrixElementApi's underlying UMAMI library from its owning Context --
+// same file, same parameter card, but a distinct in-memory instance with its
+// own global state (its own process_instance()s, its own set/get parameter
+// state, etc). Attached to an add_api()/set_normaliser() registration so
+// that TupperWare::build() can honor WareHouse::set_max_batch(): once
+// REX::lhe::max_batch causes a process to be split across several
+// REX::process batches (reweightor::extract_sorter()'s overflow cloning),
+// each batch's weightor reloads its own instance via this instead of every
+// batch sharing -- and having to serialize calls through, see
+// detail::api_call_mutex -- the one MatrixElementApi that add_api()/
+// set_normaliser() was originally given. Only consulted when the TupperWare
+// is actually built with reload_per_batch=true (see TupperWare::build()),
+// which WareHouse::build() does exactly when set_max_batch() is active.
+//
+// The instance is shared by the weightors of the batch (see
+// detail::reload_matrix_element()): one whose normaliser uses the same
+// library loads it once, while the normalisations are computed.
+struct ReloadSpec {
+    madspace::ContextPtr context;
+    std::string me_path;
+    std::string param_card;
+};
+
+namespace detail {
+// Internal, not part of the public TupperWare/WareHouse API. api's
+// process_instance() resolves to a per-thread instance via
+// madspace::ThreadPool::thread_index(), a thread_local that defaults to 0 on
+// any thread that isn't a madspace ThreadPool worker; REX::tea::threadPool
+// (both for procReweightor evaluation and, potentially, madtrex::ParamHandler
+// iterators run from within one) spawns its own plain std::threads, which all
+// resolve to instance slot 0. Every call into api that touches that instance
+// -- MatrixElementApi::call() in make_weightor()'s evaluate(), and
+// set_parameter()/get_parameter() from madtrex::ParamHandler -- goes through
+// this same per-api mutex so they can't race regardless of which thread pool
+// (if any) ends up invoking them.
+std::mutex& api_call_mutex(const madspace::MatrixElementApi& api);
+// Prints message to stderr the first time key is seen in this process (see
+// its definition); shared by TupperWare and Driver for conditions that change
+// what is computed, such as falling back to helicity summation.
+void warn_once(const std::string& key, const std::string& message);
+// MadMatrix libraries keep their model parameters in a process-wide
+// singleton, in the common library all subprocess libraries of a process
+// directory link against: every MatrixElementApi of the process shares them
+// (an instance reloaded from a ReloadSpec follows ParamHandler's launches
+// like the api it was reloaded for), and umami_initialize() -- run by each
+// load_matrix_element() -- resets them to the param_card. So:
+// - set_api_parameter(): api.set_parameter(), with the value recorded per
+//   library (file_name()) (madtrex::ParamHandler's launches). Callers hold
+//   api_call_mutex(api), as for any other set_parameter() call;
+// - reload_matrix_element(): the instance reloaded from reload for the batch
+//   proc, loaded on its first use (and kept in proc's device_data, shared by
+//   the batch's weightors) holding parameters_lock() exclusively, with the
+//   recorded parameter values set again right after;
+// - forget_api_parameters(): drops what is recorded for api's library, for a
+//   Driver that has just loaded it (umami_initialize() reset its parameters
+//   to the param_card): the values a previous Driver's launches set on the
+//   same library (the normaliser of a later fptype or model group is the
+//   original's library) must not follow its reloaded instances;
+// - parameters_lock(): held shared by every matrix element evaluation.
+std::shared_mutex& parameters_lock();
+void set_api_parameter(
+    const madspace::MatrixElementApi& api, const std::string& location, double real, double imag
+);
+void forget_api_parameters(const madspace::MatrixElementApi& api);
+const madspace::MatrixElementApi&
+reload_matrix_element(const ReloadSpec& reload, REX::process& proc);
+} // namespace detail
+
+// event::helicity_ of an event whose helicities are all fixed (LHE SPINUP -1,
+// 0 or +1) but whose helicity configuration is not in the helicity table of
+// its subprocess, as set by Driver's event checkers; -1 (REX's default) means
+// the event has no fixed helicities. Either is an error with helicity
+// reweighting, and irrelevant with helicity summation.
+inline constexpr int helicity_not_in_table = -2;
+
+// Whether an event lists its two incoming legs the other way round from the
+// matrix element's leg order: an event of a beam-swapped ("mirrored") flavor
+// channel, as MG7 writes them (LHE puts the +z beam first, while the matrix
+// element was evaluated in its own leg order). Such events are evaluated with
+// legs 0 and 1 swapped back -- momenta, and so everything derived from them --
+// without touching the events themselves. Driver derives this from the flavor
+// channels of subprocesses.json; empty means no event is swapped.
+using IncomingSwapped = std::function<bool(const REX::event&)>;
+
+// Converts an already-loaded madspace::MatrixElementApi into a REX::tea::weightor.
+// Input keys are auto-derived from api.required_inputs()/supported_inputs(),
+// restricted to what can actually be sourced from a REX::process: momenta
+// (UMAMI_IN_MOMENTA), the strong coupling (UMAMI_IN_ALPHA_S), and, if the api
+// supports channel pinning, the per-event flavor_/helicity_ indices
+// (UMAMI_IN_FLAVOR_INDEX / UMAMI_IN_HELICITY_INDEX). Throws if the api requires an
+// input that cannot be derived this way (eg a sampling-time random number), or if
+// it does not report supporting UMAMI_OUT_MATRIX_ELEMENT. The returned weightor
+// reads the process's momenta out of its flat SoA storage and reshapes them into
+// UMAMI's batch-contiguous call convention; all matching events in a process are
+// evaluated in a single UMAMI call.
+std::shared_ptr<REX::tea::weightor> make_weightor(const madspace::MatrixElementApi& api);
+
+// As above, but with explicit control over which UMAMI keys are used. input_keys
+// may only contain keys derivable from a REX::process (UMAMI_IN_MOMENTA,
+// UMAMI_IN_ALPHA_S, UMAMI_IN_FLAVOR_INDEX, UMAMI_IN_HELICITY_INDEX); output_keys
+// may only contain UMAMI_OUT_MATRIX_ELEMENT and UMAMI_OUT_DIAGRAM_AMP2 (the latter
+// is summed over diagrams before being folded into the result). result_keys picks
+// which of the requested output_keys are summed together into the weight returned
+// per event, and must be a subset of output_keys.
+std::shared_ptr<REX::tea::weightor> make_weightor(
+    const madspace::MatrixElementApi& api,
+    std::vector<UmamiInputKey> input_keys,
+    std::vector<UmamiOutputKey> output_keys,
+    std::vector<UmamiOutputKey> result_keys,
+    IncomingSwapped incoming_swapped = {}
+);
+
+// As the four-argument make_weightor above, but the returned weightor never
+// actually evaluates against api: the first time it is bound to a given
+// REX::process it reloads its own MatrixElementApi instance via
+// reload.context->load_matrix_element(reload.me_path, reload.param_card)
+// and reuses that instance (not api) for every later call against the same
+// process. api is only used, by the caller, to resolve input_keys/
+// output_keys when those aren't given explicitly -- it plays no further
+// role once the returned weightor actually runs. See ReloadSpec and
+// WareHouse::set_max_batch().
+std::shared_ptr<REX::tea::weightor> make_weightor(
+    const madspace::MatrixElementApi& api,
+    std::vector<UmamiInputKey> input_keys,
+    std::vector<UmamiOutputKey> output_keys,
+    std::vector<UmamiOutputKey> result_keys,
+    ReloadSpec reload,
+    IncomingSwapped incoming_swapped = {}
+);
+
+// The auto-derived key sets used by the single-argument make_weightor overload,
+// exposed so callers building a manual override can start from the defaults.
+// Throws under the same conditions as the single-argument make_weightor overload.
+//
+// UMAMI_IN_HELICITY_INDEX decides between helicity reweighting (each event
+// reweighted at its own helicity, the matrix element evaluated for that one
+// helicity) and helicity summation. It is included when helicity_reweighting
+// is true and api supports it -- with a one-time warning on stderr, and
+// helicity summation, when api doesn't -- or whenever api requires it. Off by
+// default for this low-level function; TupperWare and Driver default to
+// helicity reweighting (see TupperWare::set_helicity_reweighting()).
+// Reweighting with UMAMI_IN_HELICITY_INDEX then requires every event to
+// carry helicity information (a SPINUP of -1, 0 or +1 for every particle):
+// evaluation throws for a process with an event that doesn't.
+std::vector<UmamiInputKey> default_input_keys(
+    const madspace::MatrixElementApi& api, bool helicity_reweighting = false
+);
+std::vector<UmamiOutputKey> default_output_keys(const madspace::MatrixElementApi& api);
+
+// As make_weightor, but returns a REX::tea::deferred_weightor: the api's
+// umami_matrix_element call still runs once per procReweightor::evaluate(amp)
+// call (ie once per reweighting iteration -- its result legitimately differs
+// each time, eg because umami_set_parameter changed the couplings between
+// iterations), but the result is left device-resident rather than copied
+// back to the host immediately. The returned REX::tea::weight_puller does
+// that copy, lazily -- WareHouse/procReweightor::pull_deferred_backlog()
+// calls it exactly once per accumulated iteration, after the whole run's
+// iteration loop finishes, not after each individual call.
+//
+// Momenta/alphaS/flavor/helicity inputs are staged onto the api's device
+// exactly once for a given process (cached via REX::process::device_data,
+// so it's automatically invalidated whenever this weightor is bound to a
+// different process) and reused verbatim for the rest of the run, same as
+// make_weightor already does for repeated calls against an unchanged
+// process.
+//
+// n_iterations must equal (or exceed) the number of times the returned
+// deferred_weightor will actually be invoked over the process's lifetime:
+// each call claims its own slice of a per-output-key device buffer sized
+// for exactly that many iterations up front, since deferring the copy-back
+// means every iteration's result has to survive, un-overwritten, until
+// it's pulled -- an undercount throws rather than silently clobbering an
+// earlier iteration's not-yet-pulled result. WareHouse::build() derives
+// this automatically from its own iterators (see TupperWare::build());
+// pass it explicitly only when building a procReweightor by hand.
+//
+// Only meaningful for a cuda/hip-backed api (ie api.device_type() !=
+// DeviceType::cpu) -- callers don't need to check this themselves;
+// TupperWare::build() already routes cpu apis through make_weightor
+// instead.
+std::shared_ptr<REX::tea::deferred_weightor>
+make_deferred_weightor(const madspace::MatrixElementApi& api, std::size_t n_iterations);
+
+std::shared_ptr<REX::tea::deferred_weightor> make_deferred_weightor(
+    const madspace::MatrixElementApi& api,
+    std::vector<UmamiInputKey> input_keys,
+    std::vector<UmamiOutputKey> output_keys,
+    std::vector<UmamiOutputKey> result_keys,
+    std::size_t n_iterations,
+    IncomingSwapped incoming_swapped = {}
+);
+
+// As the five-argument make_deferred_weightor above, but reloads its own
+// MatrixElementApi instance per REX::process the same way the ReloadSpec
+// overload of make_weightor does, rather than ever evaluating against api
+// itself. See ReloadSpec and WareHouse::set_max_batch().
+std::shared_ptr<REX::tea::deferred_weightor> make_deferred_weightor(
+    const madspace::MatrixElementApi& api,
+    std::vector<UmamiInputKey> input_keys,
+    std::vector<UmamiOutputKey> output_keys,
+    std::vector<UmamiOutputKey> result_keys,
+    std::size_t n_iterations,
+    ReloadSpec reload,
+    IncomingSwapped incoming_swapped = {}
+);
+
+// The REX::event::extra / REX::process::extra key under which load_lhe_binary()
+// and open_lhe_source_binary() store each event's own generation-time squared
+// matrix element (a double), when the binary event file carries one (the
+// "matrix_element" column madspace writes on request, see EventRecord::
+// f_matrix_element). See TupperWare::set_stored_normalisation().
+inline constexpr const char* stored_matrix_element_key = "matrix_element";
+
+// TupperWare bundles what's needed to build a REX::tea::procReweightor for a single
+// physical process: one or more externally-owned MatrixElementApis (evaluated in
+// registration order as procReweightor::reweight_functions), an optional separate
+// normalisation MatrixElementApi, and an event checker defining which events belong
+// to this process. The MatrixElementApis are expected to outlive the TupperWare
+// (and the procReweightor/weightors it produces), matching how madspace::Context
+// owns them.
+class TupperWare {
+public:
+    TupperWare() = default;
+
+    // Convenience constructors for the common case of a single MatrixElementApi and
+    // event checker. The api is registered as the first reweight_function, and also
+    // as the normaliser if no separate normaliser is set later.
+    TupperWare(const madspace::MatrixElementApi& api, REX::eventBelongs checker);
+    TupperWare(const madspace::MatrixElementApi& api, REX::event_bool_fn checker);
+
+    // Registers a matrix element as the next reweight_functions amplitude
+    // (registration order determines the amp index used by
+    // procReweightor::evaluate(amp)), with auto-derived UMAMI keys.
+    TupperWare& add_api(const madspace::MatrixElementApi& api);
+    // As above, with an explicit override of the UMAMI input/output keys used for
+    // this particular api instead of auto-derivation.
+    TupperWare& add_api(
+        const madspace::MatrixElementApi& api,
+        std::vector<UmamiInputKey> input_keys,
+        std::vector<UmamiOutputKey> output_keys
+    );
+
+    // As above, additionally registering how to reload an independent
+    // instance of api for reweightor::extract_sorter()'s max_batch overflow
+    // clones. Only consulted when this TupperWare is built with
+    // reload_per_batch=true; see ReloadSpec and WareHouse::set_max_batch().
+    TupperWare& add_api(const madspace::MatrixElementApi& api, ReloadSpec reload);
+    TupperWare& add_api(
+        const madspace::MatrixElementApi& api,
+        std::vector<UmamiInputKey> input_keys,
+        std::vector<UmamiOutputKey> output_keys,
+        ReloadSpec reload
+    );
+
+    TupperWare& add_api(std::vector<const madspace::MatrixElementApi*> apis);
+    TupperWare& add_api(
+        std::vector<const madspace::MatrixElementApi*> apis,
+        std::vector<UmamiInputKey> input_keys_list,
+        std::vector<UmamiOutputKey> output_keys_list
+    );
+    TupperWare& add_api(
+        std::vector<const madspace::MatrixElementApi*> apis,
+        std::vector<std::vector<UmamiInputKey>> input_keys_list,
+        std::vector<std::vector<UmamiOutputKey>> output_keys_list
+    );
+
+    TupperWare& set_api(std::vector<const madspace::MatrixElementApi*> apis);
+    TupperWare& set_api(
+        std::vector<const madspace::MatrixElementApi*> apis,
+        std::vector<UmamiInputKey> input_keys_list,
+        std::vector<UmamiOutputKey> output_keys_list
+    );
+    TupperWare& set_api(
+        std::vector<const madspace::MatrixElementApi*> apis,
+        std::vector<std::vector<UmamiInputKey>> input_keys_list,
+        std::vector<std::vector<UmamiOutputKey>> output_keys_list
+    );
+
+    // Registers the matrix element used for normalisation
+    // (procReweightor::normaliser). If never called, build() leaves the normaliser
+    // unset, which makes procReweightor::initialise() default to the first
+    // reweight function.
+    TupperWare& set_normaliser(const madspace::MatrixElementApi& api);
+    TupperWare& set_normaliser(
+        const madspace::MatrixElementApi& api,
+        std::vector<UmamiInputKey> input_keys,
+        std::vector<UmamiOutputKey> output_keys
+    );
+    // As above, additionally registering how to reload an independent
+    // instance of the normaliser for max_batch overflow clones; see
+    // add_api(api, reload) and WareHouse::set_max_batch().
+    TupperWare& set_normaliser(const madspace::MatrixElementApi& api, ReloadSpec reload);
+    TupperWare& set_normaliser(
+        const madspace::MatrixElementApi& api,
+        std::vector<UmamiInputKey> input_keys,
+        std::vector<UmamiOutputKey> output_keys,
+        ReloadSpec reload
+    );
+
+    // Normalise against the squared matrix element each event was generated
+    // with, stored under REX::process::extra[key] (see
+    // stored_matrix_element_key), instead of evaluating the normaliser on it
+    // again. Only where that is the same quantity: build() uses it only when
+    // the normaliser is evaluated helicity-summed (no UMAMI_IN_HELICITY_INDEX
+    // among its input keys -- the generation's matrix element is summed
+    // over helicities, it only samples one for the event record), and each
+    // process falls back to evaluating the normaliser whenever any of its
+    // events has no stored value. An empty key (the default) always
+    // evaluates the normaliser. Has no effect without a normaliser.
+    TupperWare& set_stored_normalisation(std::string key);
+    const std::string& stored_normalisation() const;
+
+    // Which UMAMI output key(s) are summed into the weight returned by every
+    // weightor generated by this TupperWare, including the normaliser's. Defaults
+    // to just UMAMI_OUT_MATRIX_ELEMENT.
+    TupperWare& set_result_keys(std::vector<UmamiOutputKey> keys);
+
+    // Helicity reweighting (true, the default: each event is reweighted at
+    // its own helicity, via UMAMI_IN_HELICITY_INDEX) or helicity summation
+    // (false) for apis registered here without explicit input_keys (see
+    // default_input_keys()); apis added with explicit input_keys use exactly
+    // those. A matrix element without per-event helicity support falls back
+    // to helicity summation with a one-time warning; events without helicity
+    // information (any SPINUP other than -1, 0, +1) are an error with
+    // helicity reweighting on.
+    TupperWare& set_helicity_reweighting(bool helicity_reweighting);
+    bool helicity_reweighting() const { return _helicity_reweighting; }
+    // Former name of set_helicity_reweighting().
+    TupperWare& set_use_helicity(bool use_helicity);
+
+    // Which events list their incoming legs the other way round from the
+    // matrix elements (see IncomingSwapped); passed to every weightor built
+    // here, the normaliser's included.
+    TupperWare& set_incoming_swapped(IncomingSwapped incoming_swapped);
+
+    // Defines which events belong to this process. Setting one clears the other.
+    TupperWare& set_event_checker(REX::eventBelongs checker);
+    TupperWare& set_event_checker(std::shared_ptr<REX::eventBelongs> checker);
+    TupperWare& set_event_checker(REX::event_bool_fn checker);
+
+    // The MatrixElementApis registered via add_api()/set_api() (reweight
+    // functions only; not the separate normaliser, if any -- it is almost
+    // always one of these same apis anyway). Exposed so external code that
+    // needs direct access to the underlying apis themselves rather than the
+    // weightor::procReweightor wrapping them -- eg madtrex::ParamHandler,
+    // for driving parameter reweighting straight through
+    // MatrixElementApi::set_parameter() -- doesn't need its own separate
+    // tracking of "which apis did this WareHouse load".
+    const std::vector<const madspace::MatrixElementApi*>& apis() const { return _apis; }
+
+    // Builds the procReweightor from the registered apis and event checker.
+    // Throws if no apis or no event checker have been registered. Each
+    // registered api is routed to make_weightor() (cpu) or
+    // make_deferred_weightor() (cuda/hip) according to its own
+    // api.device_type() -- registering a mix of cpu and off-cpu apis in the
+    // same TupperWare works as expected, each amp slot serviced by whichever
+    // is appropriate for that api. n_iterations is only used by (forwarded
+    // to) any deferred apis; defaults to 1 for callers that build a
+    // TupperWare directly rather than through WareHouse::build(), which
+    // derives it from its own iterators automatically.
+    // reload_per_batch, when true, routes every api (and the normaliser)
+    // that was registered with a ReloadSpec through the reload-aware
+    // make_weightor()/make_deferred_weightor() overloads instead of the
+    // plain ones, so each of reweightor::extract_sorter()'s max_batch
+    // overflow clones of the resulting procReweightor ends up bound to its
+    // own freshly-loaded MatrixElementApi instance rather than sharing (and
+    // serializing calls through) the one instance add_api()/
+    // set_normaliser() were given. Apis registered without a ReloadSpec are
+    // unaffected either way. WareHouse::build() passes true automatically
+    // whenever WareHouse::set_max_batch() is active; false is the right
+    // choice for any other caller, since it costs an extra, otherwise
+    // pointless reload on every registered api's first evaluation.
+    //
+    // max_deferred_iterations is copied straight onto the built
+    // procReweightor's own field of the same name (see
+    // REX::tea::procReweightor::max_deferred_iterations): REX::npos (the
+    // default) means no cap, matching the pre-existing behavior of only
+    // ever copying deferred (eg GPU) results back to the host once, at the
+    // end of the whole run; any other value caps how many outer
+    // reweighting iterations' worth of them are allowed to accumulate
+    // device-side before an automatic mid-run flush. WareHouse::build()
+    // passes whatever WareHouse::set_max_deferred_iterations() was given.
+    std::shared_ptr<REX::tea::procReweightor> build(
+        std::size_t n_iterations = 1,
+        bool reload_per_batch = false,
+        std::size_t max_deferred_iterations = REX::npos
+    );
+
+private:
+    const madspace::MatrixElementApi* _normaliser = nullptr;
+    std::vector<UmamiInputKey> _normaliser_input_keys;
+    std::vector<UmamiOutputKey> _normaliser_output_keys;
+    std::optional<ReloadSpec> _normaliser_reload;
+    std::string _stored_normalisation;
+
+    std::vector<const madspace::MatrixElementApi*> _apis;
+    std::vector<std::vector<UmamiInputKey>> _input_keys_list;
+    std::vector<std::vector<UmamiOutputKey>> _output_keys_list;
+    std::vector<std::optional<ReloadSpec>> _reload_specs;
+    std::vector<UmamiOutputKey> _result_keys;
+    bool _helicity_reweighting = true;
+    IncomingSwapped _incoming_swapped;
+
+    std::vector<std::shared_ptr<REX::tea::weightor>> _weightors;
+    std::vector<std::shared_ptr<REX::tea::deferred_weightor>> _deferred_weightors;
+    std::shared_ptr<REX::eventBelongs> _event_checker = nullptr;
+    REX::event_bool_fn _event_checker_fn = nullptr;
+};
+
+// WareHouse aggregates one or more per-process TupperWare specifications together
+// with the reweightor-level plumbing (initialise/finalise/iterators/launch_names)
+// into a single REX::tea::reweightor, which it owns. Each registered TupperWare
+// becomes one entry in the resulting reweightor::reweightors (ie one
+// procReweightor per physical process, matching the "one TupperWare per process"
+// model documented on TupperWare above); the REX::lhe passed to build() supplies
+// the lhe base that reweightor::extract_sorter() reads events from and sorts into
+// those per-process procReweightors via their event checkers.
+//
+// TupperWares are retained after build(), so the same registered apis/checkers can
+// be rebuilt against a different REX::lhe by calling build() again.
+class WareHouse {
+public:
+    WareHouse() = default;
+    WareHouse(const WareHouse&) = delete;
+    WareHouse& operator=(const WareHouse&) = delete;
+    WareHouse(WareHouse&&) noexcept = default;
+    WareHouse& operator=(WareHouse&&) noexcept = default;
+
+    // Convenience: appends a new, empty TupperWare and returns a reference to it
+    // for in-place configuration, eg house.add_process().add_api(api).set_event_checker(chk).
+    TupperWare& add_process();
+
+    // Reweighting-loop plumbing, mirroring REX::tea::reweightor's own setters:
+    // per-iteration hooks (eg SLHA card rewriting between amplitude evaluations),
+    // hooks run once before/after the iteration loop, and the weight-id names
+    // recorded for each iteration's appended weights. Optional: if never called,
+    // build() leaves the corresponding reweightor default in place (a single
+    // no-op iteration, true_function initialise/finalise, no launch_names).
+    WareHouse& set_iterators(std::vector<REX::tea::iterator> iters);
+    WareHouse& add_iterator(REX::tea::iterator iter);
+    WareHouse& set_initialise(REX::tea::iterator init);
+    WareHouse& set_finalise(REX::tea::iterator fin);
+    WareHouse& set_launch_names(std::vector<std::string> names);
+    WareHouse& add_launch_name(const std::string& name);
+    const std::vector<std::string>& launch_names() const;
+
+    // Maximum number of events per REX::process (see REX::lhe::max_batch),
+    // applied to the lhe passed to build(). Left at REX::npos (unset) by
+    // default, meaning no splitting. Setting this to something else also
+    // makes build() pass reload_per_batch=true to every registered
+    // TupperWare's build() (see TupperWare::build()), so that apis
+    // registered with a ReloadSpec get their own freshly-loaded instance per
+    // batch instead of every overflow clone sharing (and serializing calls
+    // through) the same one -- which is the whole point of splitting a
+    // process into independently-runnable batches to begin with. Apis
+    // registered without a ReloadSpec are unaffected.
+    WareHouse& set_max_batch(std::size_t max_batch);
+    std::size_t max_batch() const;
+
+    // Worker threads of the built reweightor (REX::tea::reweightor::
+    // pool_threads): 0, the default, uses std::thread::hardware_concurrency().
+    // The reweightor runs one REX::process batch per thread at a time, so
+    // threads only help with as many batches; see set_min_batch().
+    WareHouse& set_threads(std::size_t threads);
+    std::size_t threads() const;
+
+    // Splitting the events across threads (see batch_size_for()): with
+    // max_batch() unset, build() splits every process with more than
+    // max(min_batch, events / threads) events into batches of that size, so
+    // that a sample dominated by one subprocess still keeps every thread
+    // busy, while no batch gets smaller than min_batch events (below which
+    // a batch's fixed cost -- its reloaded MatrixElementApi instance, its
+    // own sorting and setup -- outweighs what another thread gains).
+    // REX::npos, the default, never splits.
+    WareHouse& set_min_batch(std::size_t min_batch);
+    std::size_t min_batch() const;
+
+    // The REX::lhe::max_batch build() uses for a sample of n_events: max_batch()
+    // when set; else max(min_batch(), ceil(n_events / threads)) when
+    // min_batch() is set, more than one thread is used and the sample has
+    // at least 2 * min_batch() events; else REX::npos (no splitting).
+    std::size_t batch_size_for(std::size_t n_events) const;
+
+    // Passed straight through to every registered TupperWare's build() (see
+    // TupperWare::build()'s own doc comment) -- ie every built
+    // procReweightor's max_deferred_iterations field. Left at REX::npos
+    // (the default) meaning no cap: deferred (eg GPU) results only ever
+    // get copied back to the host once, at the end of the whole run, same
+    // as before this existed. Useful together with off-cpu apis (see
+    // TupperWare::add_api()/make_deferred_weightor()) to bound how much
+    // device-side memory a long run's worth of not-yet-pulled results can
+    // occupy at once, at the cost of more frequent, smaller device->host
+    // copies.
+    WareHouse& set_max_deferred_iterations(std::size_t max_deferred_iterations);
+    std::size_t max_deferred_iterations() const;
+
+    // The muted-by-default warning logger shared with the built reweightor
+    // (and, transitively, every procReweightor/process batch it runs --
+    // see REX::tea::reweightor::extract_sorter(), which links each batch's
+    // procReweightor::verbose to this same flag). Muted (false) from
+    // construction, independent of whether build() has been called yet;
+    // call verbose().set_verbose(true) to unmute. Re-linked into the
+    // reweightor's own verbose on every build()/rebuild.
+    REX::verbosity& verbose();
+    const REX::verbosity& verbose() const;
+    // Plain-bool convenience for verbose().set_verbose(v)/is_verbose(),
+    // for callers (eg Python bindings) that would rather not need
+    // REX::verbosity itself.
+    WareHouse& set_verbose(bool v);
+    bool is_verbose() const;
+
+    // Builds (or rebuilds) the owned reweightor from the registered TupperWares'
+    // procReweightors and the given REX::lhe, which becomes the reweightor's lhe
+    // base. Throws if no processes have been registered. Repeated calls discard
+    // any previously-built reweightor.
+    REX::tea::reweightor& build(REX::lhe&& mother);
+    REX::tea::reweightor& build(const REX::lhe& mother);
+
+    // Access to the owned reweightor. Throws if build() has not been called yet.
+    REX::tea::reweightor& get();
+    const REX::tea::reweightor& get() const;
+    bool built() const noexcept;
+
+    // Direct access to TupperWare; getting, setting, and adding
+    TupperWare& process(std::size_t index);
+    const TupperWare& process(std::size_t index) const;
+    std::deque<TupperWare>& processes();
+    const std::deque<TupperWare>& processes() const;
+    WareHouse& set_processes(std::vector<TupperWare> processes);
+    WareHouse& add_process(TupperWare process);
+    WareHouse& add_process(std::vector<TupperWare> processes);
+
+private:
+    // a deque, not a vector: add_process() hands out references to its
+    // elements (also to Python), which adding further processes must not
+    // invalidate
+    std::deque<TupperWare> _processes;
+    std::vector<REX::tea::iterator> _iterators;
+    REX::tea::iterator _initialise = REX::tea::true_function;
+    REX::tea::iterator _finalise = REX::tea::true_function;
+    std::vector<std::string> _launch_names;
+    std::size_t _max_batch = REX::npos;
+    std::size_t _threads = 0;
+    std::size_t _min_batch = REX::npos;
+    std::size_t _max_deferred_iterations = REX::npos;
+    REX::verbosity _verbose{false};
+
+    std::unique_ptr<REX::tea::reweightor> _reweightor;
+};
+
+} // namespace madtrex
diff --git a/madtrex/install.py b/madtrex/install.py
new file mode 100644
index 0000000000..08d506e674
--- /dev/null
+++ b/madtrex/install.py
@@ -0,0 +1,294 @@
+#!/usr/bin/env python3
+"""Build and install MadtRex (and the bundled Rex fork it needs) from source.
+
+MadtRex sits on top of madspace, which has to be installed first (by default
+from ../madspace/install, i.e. `python madspace/install.py`); madspace itself
+does not depend on MadtRex.
+
+Usage:
+  python install.py                 # interactive
+  python install.py --yes           # no prompts
+  python install.py --yes -j 8
+  python install.py --clean         # rebuild from scratch
+  python install.py --system        # into the current Python environment
+
+Steps:
+  1. check that madspace is importable (from ../madspace/install, or from the
+     current environment with --system);
+  2. pip install ./Rex (the `pyrex` package) into install/;
+  3. pip install this directory (the `madtrex` package) into install/,
+     built against that madspace.
+
+Rebuilds reuse the CMake tree in build/ and are therefore incremental. --clean
+deletes build/ and install/ first.
+"""
+
+import argparse
+import os
+import shutil
+import subprocess
+import sys
+import tomllib
+from pathlib import Path
+
+SCRIPT_DIR = Path(__file__).parent.resolve()
+INSTALL_DIR = SCRIPT_DIR / "install"
+BUILD_DIR = SCRIPT_DIR / "build"
+REX_DIR = SCRIPT_DIR / "Rex"
+MADSPACE_SRC_DIR = SCRIPT_DIR.parent / "madspace"
+MADSPACE_INSTALL_DIR = MADSPACE_SRC_DIR / "install"
+
+
+def ask_yes_no(prompt: str, default: bool = True) -> bool:
+    suffix = " [Y/n] " if default else " [y/N] "
+    while True:
+        try:
+            answer = input(prompt + suffix).strip().lower()
+        except EOFError:
+            return default
+        if not answer:
+            return default
+        if answer in ("y", "yes"):
+            return True
+        if answer in ("n", "no"):
+            return False
+
+
+def default_jobs() -> int:
+    """Number of compilation jobs to use when none was requested."""
+    try:
+        return len(os.sched_getaffinity(0))  # Linux: respects cgroup/taskset
+    except AttributeError:
+        return os.cpu_count() or 1
+
+
+def set_build_parallelism(env: dict, jobs: int | None) -> dict:
+    """See madspace/install.py: scikit-build-core runs `cmake --build` without
+    -j, so the job count has to come from CMAKE_BUILD_PARALLEL_LEVEL."""
+    if jobs is not None and jobs <= 0:
+        jobs = None
+    if jobs is None:
+        if env.get("CMAKE_BUILD_PARALLEL_LEVEL"):
+            return env
+        jobs = default_jobs()
+    env["CMAKE_BUILD_PARALLEL_LEVEL"] = str(jobs)
+    print(f"Compiling with {jobs} parallel job(s) (-j to change).")
+    return env
+
+
+def run(cmd: list, env: dict | None = None, cwd: Path = SCRIPT_DIR) -> None:
+    display = " ".join(str(c) for c in cmd)
+    print(f"\n$ {display}\n")
+    result = subprocess.run(cmd, cwd=cwd, env=env)
+    if result.returncode != 0:
+        sys.exit(result.returncode)
+
+
+def with_pythonpath(env: dict, *dirs: Path) -> dict:
+    parts = [str(d) for d in dirs]
+    if existing := env.get("PYTHONPATH"):
+        parts.append(existing)
+    env["PYTHONPATH"] = os.pathsep.join(parts)
+    return env
+
+
+def find_madspace(env: dict) -> Path | None:
+    """The madspace package directory the build would use, or None if
+    madspace cannot be imported with this environment."""
+    result = subprocess.run(
+        [
+            sys.executable,
+            "-c",
+            "import madspace, os; "
+            "print(os.path.dirname(os.path.abspath(madspace.__file__)))",
+        ],
+        env=env,
+        capture_output=True,
+        text=True,
+    )
+    if result.returncode != 0:
+        return None
+    return Path(result.stdout.strip().splitlines()[-1])
+
+
+def madspace_problem(madspace_dir: Path, env: dict) -> str | None:
+    """Why the madspace in madspace_dir is too old to build MadtRex against
+    (it predates the API MadtRex uses: umami.h installed with the headers,
+    MatrixElementApi.set_parameter/supported_inputs), or None."""
+    if not (madspace_dir / "include").is_dir():
+        return ("it has no headers (include/), as a precompiled copy of madspace "
+                "(e.g. in a gridpack) does")
+    if not (madspace_dir / "include" / "madspace" / "umami.h").is_file():
+        return "it does not install include/madspace/umami.h"
+    result = subprocess.run(
+        [
+            sys.executable,
+            "-c",
+            "import madspace; api = madspace.MatrixElementApi; "
+            "missing = [n for n in ('supports_set_parameter', 'set_parameter', "
+            "'supported_inputs') if not hasattr(api, n)]; print(' '.join(missing))",
+        ],
+        env=env,
+        capture_output=True,
+        text=True,
+    )
+    if result.returncode != 0:
+        return "importing it failed:\n" + result.stderr.strip()
+    missing = result.stdout.strip().splitlines()[-1:] if result.stdout.strip() else []
+    if missing and missing[0]:
+        return "its MatrixElementApi has no " + ", ".join(missing[0].split())
+    return None
+
+
+def install_build_deps(env: dict, system: bool) -> None:
+    """Install the [build-system] requires of pyproject.toml, needed because
+    madtrex is built with --no-build-isolation (so that CMake sees the same
+    Python environment, and therefore the same madspace, as the install)."""
+    with open(SCRIPT_DIR / "pyproject.toml", "rb") as f:
+        requires = tomllib.load(f).get("build-system", {}).get("requires", [])
+    if not requires:
+        return
+    print("\nInstalling build dependencies...")
+    cmd = [sys.executable, "-m", "pip", "install", "--upgrade", *requires]
+    if not system:
+        cmd.append(f"--target={INSTALL_DIR}")
+    run(cmd, env=env)
+
+
+def clean_install_dirs() -> None:
+    for target in (BUILD_DIR, INSTALL_DIR, REX_DIR / "build"):
+        if target.is_dir():
+            shutil.rmtree(target)
+            print(f"Removed {target}")
+
+
+def build_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(
+        description=__doc__,
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+    )
+    parser.add_argument(
+        "-y",
+        "--yes",
+        action="store_true",
+        default=False,
+        help="Non-interactive, no prompts.",
+    )
+    parser.add_argument(
+        "-j",
+        "--jobs",
+        type=int,
+        default=None,
+        metavar="N",
+        help="Number of parallel compilation jobs "
+        f"(default: every available core, {default_jobs()} here).",
+    )
+    parser.add_argument(
+        "--system",
+        action="store_true",
+        default=False,
+        help="Install into the current Python environment instead of the local "
+        "install/ directory; madspace must then be importable from it too.",
+    )
+    parser.add_argument(
+        "--clean",
+        action="store_true",
+        default=False,
+        help="Delete the build/ and install/ directories first.",
+    )
+    return parser
+
+
+def main(argv: list[str] | None = None) -> None:
+    args = build_parser().parse_args(argv)
+
+    if args.clean:
+        clean_install_dirs()
+
+    env = os.environ.copy()
+    if not args.system:
+        env = with_pythonpath(env, INSTALL_DIR, MADSPACE_INSTALL_DIR)
+
+    # 1. madspace
+    madspace_dir = find_madspace(env)
+    if madspace_dir is None:
+        where = (
+            "the current Python environment"
+            if args.system
+            else str(MADSPACE_INSTALL_DIR)
+        )
+        madspace_cmd = [sys.executable, str(MADSPACE_SRC_DIR / "install.py")]
+        if args.system:
+            madspace_cmd.append("--system")
+        print(f"madspace is not installed in {where}; MadtRex needs it.")
+        if args.yes or not ask_yes_no("Run madspace/install.py now?", default=True):
+            print(
+                "Install madspace first:\n    "
+                + " ".join(madspace_cmd)
+                + "\nthen re-run this script."
+            )
+            sys.exit(1)
+        run(madspace_cmd, cwd=MADSPACE_SRC_DIR)
+        madspace_dir = find_madspace(env)
+        if madspace_dir is None:
+            print("madspace is still not importable after installing it; giving up.")
+            sys.exit(1)
+    if problem := madspace_problem(madspace_dir, env):
+        print(
+            f"The madspace in {madspace_dir} is too old for MadtRex: {problem}.\n"
+            "Reinstall madspace from this checkout's sources first:\n    "
+            f"{sys.executable} {MADSPACE_SRC_DIR / 'install.py'} --source --yes\n"
+            "(or 'install madspace --source' in MadGraph), then re-run this script."
+        )
+        sys.exit(1)
+    print(f"Using madspace from {madspace_dir}")
+
+    env = set_build_parallelism(env, args.jobs)
+
+    # 2. the bundled Rex fork (pyrex)
+    rex_cmd = [
+        sys.executable,
+        "-m",
+        "pip",
+        "install",
+        "--upgrade",
+        "--no-deps",
+        "-Cbuild-dir=build",
+        ".",
+    ]
+    if not args.system:
+        rex_cmd.append(f"--target={INSTALL_DIR}")
+    run(rex_cmd, env=env, cwd=REX_DIR)
+
+    # 3. madtrex itself. --no-deps: madspace and pyrex were just checked /
+    # installed above; letting pip resolve them would pull unrelated
+    # releases from PyPI instead of the local ones.
+    install_build_deps(env, args.system)
+    cmd = [
+        sys.executable,
+        "-m",
+        "pip",
+        "install",
+        "--no-build-isolation",
+        "--no-deps",
+        "--upgrade",
+        "-Cbuild-dir=build",
+        f"-Ccmake.define.MADSPACE_DIR={madspace_dir}",
+        ".",
+    ]
+    if not args.system:
+        cmd.append(f"--target={INSTALL_DIR}")
+    run(cmd, env=env)
+
+    if args.system:
+        print("\nInstalled system-wide.")
+    else:
+        print(f"\nInstalled to: {INSTALL_DIR}")
+        print(
+            "To use it outside MadGraph, put both install directories on "
+            f"PYTHONPATH:\n    {INSTALL_DIR}{os.pathsep}{MADSPACE_INSTALL_DIR}"
+        )
+
+
+if __name__ == "__main__":
+    main()
diff --git a/madtrex/madtrex/__init__.py b/madtrex/madtrex/__init__.py
new file mode 100644
index 0000000000..fbc800c0a8
--- /dev/null
+++ b/madtrex/madtrex/__init__.py
@@ -0,0 +1,26 @@
+"""MadtRex: MatrixElementApi <-> REX::tea event reweighting for MadGraph7.
+
+Built on top of madspace and pyrex (the bundled Rex fork); madspace itself
+does not depend on this package.
+"""
+
+import os as _os
+
+from . import _compat
+
+# _madtrex_py is compiled against madspace's headers: refuse a madspace that
+# changed since (a binary mismatch would otherwise crash or misbehave).
+_madspace_dir = _compat.find_madspace_dir()
+if _madspace_dir is not None:
+    _problem = _compat.build_problem(_os.path.dirname(_os.path.abspath(__file__)),
+                                     _madspace_dir)
+    if _problem:
+        raise ImportError(_problem)
+
+# madspace first: it preloads libmadspace and registers the MatrixElementApi/
+# Context/LHEMeta types that the _madtrex_py bindings use.
+import madspace  # noqa: F401,E402
+import pyrex  # noqa: F401,E402
+import pyrex.tea  # noqa: F401,E402
+
+from ._madtrex_py import *  # noqa: F401,F403,E402
diff --git a/madtrex/madtrex/_compat.py b/madtrex/madtrex/_compat.py
new file mode 100644
index 0000000000..fbc0d4c55e
--- /dev/null
+++ b/madtrex/madtrex/_compat.py
@@ -0,0 +1,78 @@
+"""Is this madtrex build compatible with the madspace next to it?
+
+_madtrex_py is compiled against madspace's installed headers: its classes
+(MatrixElementApi, EventFile, ...) and their inline methods are baked into
+it. A madspace reinstalled with other headers after madtrex was built (a
+newer checkout, 'install madspace --bin', ...) is a silent binary mismatch,
+so CMake records a digest of the headers it compiled against in
+_build_info.json, and this module compares it with the madspace that would
+be imported.
+
+Pure Python and standard library only: MadGraph loads this file directly
+(madgraph/various/reweight_backend.py) to tell whether MadtRex can run,
+without importing madtrex, madspace or pyrex.
+"""
+
+import hashlib
+import importlib.util
+import json
+import os
+
+BUILD_INFO = "_build_info.json"
+
+REBUILD_HINT = (
+    "rebuild it: python madtrex/install.py --clean --yes "
+    "(or 'install madtrex --clean -y' in MadGraph)"
+)
+
+
+def headers_digest(include_dir):
+    """sha256 over the files under include_dir: one "<relative path>
+    <sha256>" line per file, in sorted path order. madtrex/CMakeLists.txt
+    computes the same digest at build time."""
+    lines = []
+    for root, _dirs, files in os.walk(include_dir):
+        for name in files:
+            path = os.path.join(root, name)
+            rel = os.path.relpath(path, include_dir).replace(os.sep, "/")
+            with open(path, "rb") as f:
+                lines.append((rel, hashlib.sha256(f.read()).hexdigest()))
+    text = "".join("%s %s\n" % line for line in sorted(lines))
+    return hashlib.sha256(text.encode()).hexdigest()
+
+
+def read_build_info(package_dir):
+    """The _build_info.json of the madtrex package in package_dir, or None."""
+    try:
+        with open(os.path.join(package_dir, BUILD_INFO)) as f:
+            return json.load(f)
+    except (OSError, ValueError):
+        return None
+
+
+def find_madspace_dir():
+    """The directory of the madspace package that 'import madspace' would
+    load, without importing it, or None."""
+    spec = importlib.util.find_spec("madspace")
+    if spec is None or not spec.origin:
+        return None
+    return os.path.dirname(os.path.abspath(spec.origin))
+
+
+def build_problem(package_dir, madspace_dir):
+    """Why the madtrex package in package_dir cannot be used with the
+    madspace package in madspace_dir, or None when it can (or when there is
+    nothing to compare: a build from before the digest was recorded)."""
+    info = read_build_info(package_dir)
+    if not info or not info.get("madspace_headers"):
+        return None
+    include_dir = os.path.join(madspace_dir, "include")
+    if not os.path.isdir(include_dir):
+        return None
+    if headers_digest(include_dir) == info["madspace_headers"]:
+        return None
+    return (
+        "madtrex (%s) was built against another madspace than the one in %s "
+        "(madspace was reinstalled or updated since): %s"
+        % (package_dir, madspace_dir, REBUILD_HINT)
+    )
diff --git a/madtrex/pyproject.toml b/madtrex/pyproject.toml
new file mode 100644
index 0000000000..a1e958309c
--- /dev/null
+++ b/madtrex/pyproject.toml
@@ -0,0 +1,34 @@
+[build-system]
+requires = ["scikit-build-core>=0.10", "packaging>=22"]
+build-backend = "scikit_build_core.build"
+
+[project]
+name = "madtrex"
+# Taken from the MadGraph VERSION file, like madspace: madtrex is never
+# released on its own.
+dynamic = ["version"]
+description = "MadtRex: event reweighting for MadGraph7, bridging madspace matrix elements and Rex/teaRex"
+readme = "README.md"
+authors = [
+  { name = "Zenny Wettersten" },
+]
+requires-python = ">=3.11"
+classifiers = [
+    "Programming Language :: Python :: 3",
+    "Operating System :: OS Independent",
+]
+# madspace provides MatrixElementApi/Context/LHEMeta (and libmadspace, which
+# _madtrex_py links against); pyrex is the Python package of the Rex fork
+# bundled in Rex/ (pip install ./Rex), providing the REX::tea types that
+# TupperWare/WareHouse build and consume.
+dependencies = ["madspace", "pyrex"]
+
+[tool.scikit-build]
+minimum-version = "build-system.requires"
+cmake.build-type = "Release"
+wheel.packages = ["madtrex"]
+
+[tool.scikit-build.metadata.version]
+provider = "scikit_build_core.metadata.regex"
+input = "../VERSION"
+regex = '(?m)^version\s*=\s*(?P<value>\S+)'
diff --git a/madtrex/src/binary_io.cpp b/madtrex/src/binary_io.cpp
new file mode 100644
index 0000000000..ca3cd556cd
--- /dev/null
+++ b/madtrex/src/binary_io.cpp
@@ -0,0 +1,498 @@
+#include "madtrex/binary_io.hpp"
+
+#include <algorithm>
+#include <cctype>
+#include <stdexcept>
+
+namespace madtrex {
+
+namespace {
+
+using madspace::DataLayout;
+using madspace::EventBuffer;
+using madspace::EventFile;
+using madspace::EventRecord;
+using madspace::LHEEvent;
+using madspace::LHEMeta;
+using madspace::LHEParticle;
+using madspace::ParticleRecord;
+
+// The layout madtrex writes its own binary LHE files with (open_lhe_sink_
+// binary(), save_weights_binary()'s sibling for events): always the same,
+// fixed shape, since madtrex controls that format end to end -- unlike
+// detect_lhe_event_layout() below, which reads a file this code did *not*
+// write.
+DataLayout lhe_event_layout() {
+    return DataLayout(
+        EventRecord::layout(EventRecord::f_lhe_event),
+        ParticleRecord::layout(ParticleRecord::f_lhe_particle)
+    );
+}
+
+// The layout to open path (an externally-produced binary LHE-format event
+// file, eg MadGraph7's own events.npy from EventGenerator::
+// combine_to_lhe_npy()) for *reading* with. Unlike lhe_event_layout()
+// above, this isn't a fixed shape: combine_to_lhe_npy() additionally stores
+// each event's own x1/fact_scale1 (beam 1) and x2/fact_scale2 (beam 2)
+// whenever run_card.toml's [generation] systematics is enabled, growing
+// every per-event record by that much. extract_lhe_event() never actually
+// reads those values (only process_id/weight/scale/alpha_qed/alpha_qcd/
+// particles), but the reader still needs to know they're there in order to
+// compute correct byte offsets for everything that comes after them in the
+// same record (in particular every particle's own data) -- so this peeks
+// the file's own header first and builds a layout that actually matches,
+// rather than assuming systematics was off. An MLM run's events also
+// carry each particle's cluster_scale (ParticleRecord::f_clustering), which
+// the reader skips over the same way.
+DataLayout detect_lhe_event_layout(const std::string& path) {
+    auto names = madspace::peek_npy_field_names(path);
+    auto has = [&](const std::string& name) {
+        return std::find(names.begin(), names.end(), name) != names.end();
+    };
+    int event_flags = EventRecord::f_lhe_event;
+    if (has("fact_scale1")) event_flags |= EventRecord::f_beam1;
+    if (has("fact_scale2")) event_flags |= EventRecord::f_beam2;
+    if (has("partial_weight_product")) event_flags |= EventRecord::f_partial_weights;
+    auto event_layout = EventRecord::layout(event_flags);
+    // The weight columns after those: the systematics' rwgt_<id> (written
+    // with [systematics] enabled, the default) and the reweighting's own
+    // rwgt_<name> columns (added to events.npy in place by MadtRex's
+    // MadGraph-side driver), in file order, then the bias. Missing them
+    // shifted every later field, and the file was refused as "Invalid
+    // header for event file".
+    for (const auto& name : names) {
+        if (name.rfind("rwgt_", 0) == 0) {
+            event_layout.push_back({name, madspace::FieldLayout::f64, 7});
+        }
+    }
+    if (has("bias")) {
+        event_layout.push_back({"bias", madspace::FieldLayout::f64, 8});
+    }
+    // the generation-time matrix element madspace stores on request, for
+    // stored_matrix_element_key
+    if (has("matrix_element")) {
+        event_layout.push_back({"matrix_element", madspace::FieldLayout::f64, 9});
+    }
+    int particle_flags = ParticleRecord::f_lhe_particle;
+    if (has("part1_cluster_scale")) particle_flags |= ParticleRecord::f_clustering;
+    return DataLayout(event_layout, ParticleRecord::layout(particle_flags));
+}
+
+// Extracts event index i out of buffer (laid out with EventRecord::
+// f_lhe_event | ParticleRecord::f_lhe_particle, max_particles particles per
+// event) as an LHEEvent. Particles with status_code == 0 are dropped: real
+// LHE particles never carry status 0 (incoming is -1, outgoing +1,
+// intermediate resonance 2), so a 0 marks the zero-filled padding
+// EventGenerator::combine_to_lhe_npy (or open_lhe_sink_binary(), for a
+// shard written by split_events()) appends past an event's actual particle
+// count, up to the file's fixed particle_count(). Shared between
+// read_binary_events() (whole-file) and open_lhe_source_binary()'s pull()
+// (streamed, chunk at a time) -- the two differ only in how many events'
+// worth of buffer they hand this one event at a time.
+LHEEvent extract_lhe_event(EventBuffer& buffer, std::size_t i, std::size_t max_particles) {
+    auto rec = buffer.event(i);
+    LHEEvent evt;
+    evt.process_id = rec.lhe_process_id();
+    evt.weight = rec.lhe_weight();
+    evt.scale = rec.lhe_scale();
+    evt.alpha_qed = rec.lhe_alpha_qed();
+    evt.alpha_qcd = rec.lhe_alpha_qcd();
+    evt.particles.reserve(max_particles);
+    for (std::size_t j = 0; j < max_particles; ++j) {
+        auto prec = buffer.particle(i, j);
+        int status = prec.lhe_status_code();
+        if (status == 0) break; // padding: always a trailing run
+        LHEParticle part;
+        part.pdg_id = prec.lhe_pdg_id();
+        part.status_code = status;
+        part.mother1 = prec.lhe_mother1();
+        part.mother2 = prec.lhe_mother2();
+        part.color = prec.lhe_color();
+        part.anti_color = prec.lhe_anti_color();
+        part.px = prec.lhe_px();
+        part.py = prec.lhe_py();
+        part.pz = prec.lhe_pz();
+        part.energy = prec.lhe_energy();
+        part.mass = prec.lhe_mass();
+        part.lifetime = prec.lhe_lifetime();
+        part.spin = prec.lhe_spin();
+        evt.particles.push_back(part);
+    }
+    return evt;
+}
+
+bool has_matrix_element(const DataLayout& layout) {
+    const auto& fields = layout.event_layout();
+    return std::any_of(fields.begin(), fields.end(), [](const auto& field) {
+        return field.name == "matrix_element";
+    });
+}
+
+// Reads every event out of an already-open, load-mode EventFile. Uses
+// file.layout() (whatever it was actually opened with -- see
+// detect_lhe_event_layout()) rather than assuming a fixed one, so this
+// works regardless of which optional extra fields that layout does or
+// doesn't include.
+// The file's stored matrix elements, if it has them, go to
+// matrix_elements, one per event.
+std::vector<LHEEvent> read_binary_events(EventFile& file, std::vector<double>& matrix_elements) {
+    // Named local: EventBuffer stores its DataLayout by reference, so the
+    // layout must outlive every buffer built from it -- file.layout()
+    // lives exactly as long as file does, which already outlives this
+    // function's own scope (owned by the caller).
+    const DataLayout& layout = file.layout();
+    std::size_t max_particles = file.particle_count();
+    EventBuffer buffer(0, max_particles, layout);
+
+    std::vector<LHEEvent> events;
+    events.reserve(file.event_count());
+    bool has_me = has_matrix_element(layout);
+    matrix_elements.clear();
+
+    constexpr std::size_t chunk = 4096;
+    while (file.read(buffer, chunk)) {
+        for (std::size_t i = 0; i < buffer.event_count(); ++i) {
+            events.push_back(extract_lhe_event(buffer, i, max_particles));
+            if (has_me) matrix_elements.push_back(buffer.event(i).matrix_element());
+        }
+    }
+    return events;
+}
+
+REX::parton lhe_particle_to_parton(const LHEParticle& p) {
+    REX::arr2<short int> mother{
+        static_cast<short int>(p.mother1), static_cast<short int>(p.mother2)
+    };
+    REX::arr2<short int> icol{
+        static_cast<short int>(p.color), static_cast<short int>(p.anti_color)
+    };
+    return REX::parton(
+        REX::arr4<double>{p.energy, p.px, p.py, p.pz},
+        p.mass,
+        p.lifetime,
+        p.spin,
+        static_cast<long int>(p.pdg_id),
+        static_cast<short int>(p.status_code),
+        mother,
+        icol
+    );
+}
+
+std::shared_ptr<REX::event> lhe_event_to_event(const LHEEvent& raw) {
+    std::vector<REX::parton> partons;
+    partons.reserve(raw.particles.size());
+    for (auto& p : raw.particles) partons.push_back(lhe_particle_to_parton(p));
+
+    auto evt = std::make_shared<REX::event>(std::move(partons));
+    evt->set_proc_id(raw.process_id);
+    evt->set_weight(raw.weight);
+    evt->set_scale(raw.scale);
+    evt->set_alphaEW(raw.alpha_qed);
+    evt->set_alphaS(raw.alpha_qcd);
+    return evt;
+}
+
+REX::initNode lhe_meta_to_init(const LHEMeta& meta) {
+    // Deliberately the default (nProc_ == 0, xSec_/xSecErr_/xMax_/lProc_ all
+    // empty) constructor, not initNode(nproc): that overload pre-sizes those
+    // four vectors to nproc entries, which the add_xSec()/etc. push-back loop
+    // below would then grow past nproc, tripping validate_init()'s
+    // nProc_-matches-size checks. set_nProc() records the count separately,
+    // the same way the LHEF XML reader (read_lhef_body() in Rex.cc) does it.
+    REX::initNode init;
+    init.set_nProc(static_cast<short unsigned int>(meta.processes.size()));
+    init.set_idBm(
+        static_cast<long int>(meta.beam1_pdg_id), static_cast<long int>(meta.beam2_pdg_id)
+    );
+    init.set_eBm(meta.beam1_energy, meta.beam2_energy);
+    init.set_pdfG(
+        static_cast<short int>(meta.beam1_pdf_authors),
+        static_cast<short int>(meta.beam2_pdf_authors)
+    );
+    init.set_pdfS(
+        static_cast<long int>(meta.beam1_pdf_id), static_cast<long int>(meta.beam2_pdf_id)
+    );
+    init.set_idWgt(static_cast<short int>(meta.weight_mode));
+    for (auto& proc : meta.processes) {
+        init.add_xSec(proc.cross_section);
+        init.add_xSecErr(proc.cross_section_error);
+        init.add_xMax(proc.max_weight);
+        init.add_lProc(proc.process_id);
+    }
+    return init;
+}
+
+// Inverse of lhe_particle_to_parton()/lhe_event_to_event(): reads back out
+// of a REX::event exactly the fields those two populated in, so
+// open_lhe_sink_binary()'s round trip through load_lhe_binary() is
+// lossless for everything the binary format itself carries (it never
+// carried mass_/vtim_/spin_ separately from momenta on the *event* struct
+// to begin with -- those are per-particle, read from p below).
+LHEParticle parton_to_lhe_particle(const REX::event& e, std::size_t i) {
+    LHEParticle p;
+    p.pdg_id = static_cast<int>(e.pdg_[i]);
+    p.status_code = e.status_[i];
+    p.mother1 = e.mother_[i][0];
+    p.mother2 = e.mother_[i][1];
+    p.color = e.icol_[i][0];
+    p.anti_color = e.icol_[i][1];
+    p.energy = e.momenta_[i].t();
+    p.px = e.momenta_[i].x();
+    p.py = e.momenta_[i].y();
+    p.pz = e.momenta_[i].z();
+    p.mass = e.mass_[i];
+    p.lifetime = e.vtim_[i];
+    p.spin = e.spin_[i];
+    return p;
+}
+
+std::string sanitize_token(const std::string& raw) {
+    std::string out;
+    out.reserve(raw.size());
+    for (char c : raw) {
+        out.push_back(
+            (std::isalnum(static_cast<unsigned char>(c)) || c == '_' || c == '-') ? c : '_'
+        );
+    }
+    return out.empty() ? std::string("weight") : out;
+}
+
+} // namespace
+
+std::shared_ptr<REX::lhe> load_lhe_binary(EventFile& file, const LHEMeta& meta) {
+    std::vector<double> matrix_elements;
+    auto raw_events = read_binary_events(file, matrix_elements);
+
+    REX::lheReader<LHEMeta, LHEEvent> reader;
+    reader.set_init_translator(lhe_meta_to_init);
+    reader.set_event_translator(lhe_event_to_event);
+
+    auto result = reader.read(meta, raw_events);
+    if (!matrix_elements.empty() && matrix_elements.size() == result.events.size()) {
+        for (std::size_t i = 0; i < result.events.size(); ++i) {
+            result.events[i]->extra[stored_matrix_element_key] = matrix_elements[i];
+        }
+    }
+    return std::make_shared<REX::lhe>(std::move(result));
+}
+
+std::shared_ptr<REX::lhe> load_lhe_binary(const std::string& path, const LHEMeta& meta) {
+    EventFile file(path, detect_lhe_event_layout(path), 0, EventFile::load);
+    return load_lhe_binary(file, meta);
+}
+
+REX::eventSource open_lhe_source_binary(const std::string& path, const LHEMeta& meta) {
+    // Holds the open EventFile (and, via its own layout(), the DataLayout
+    // EventBuffer only ever borrows a reference to -- see
+    // extract_lhe_event()'s doc comment) for as long as the returned
+    // eventSource's pull() closure is alive. Forward-only, same as
+    // open_lhe_source(): no rewind is offered.
+    struct State {
+        EventFile file;
+        std::size_t max_particles;
+        bool has_me;
+        explicit State(const std::string& p) :
+            file(p, detect_lhe_event_layout(p), 0, EventFile::load),
+            max_particles(file.particle_count()),
+            has_me(has_matrix_element(file.layout())) {}
+    };
+    auto state = std::make_shared<State>(path);
+
+    REX::eventSource src;
+    src.read_init = [meta]() -> REX::initNode { return lhe_meta_to_init(meta); };
+    src.pull = [state](std::size_t n) -> std::vector<std::shared_ptr<REX::event>> {
+        EventBuffer buffer(0, state->max_particles, state->file.layout());
+        if (!state->file.read(buffer, n)) {
+            return {};
+        }
+        std::vector<std::shared_ptr<REX::event>> result;
+        result.reserve(buffer.event_count());
+        for (std::size_t i = 0; i < buffer.event_count(); ++i) {
+            result.push_back(
+                lhe_event_to_event(extract_lhe_event(buffer, i, state->max_particles))
+            );
+            if (state->has_me) {
+                result.back()->extra[stored_matrix_element_key] =
+                    static_cast<double>(buffer.event(i).matrix_element());
+            }
+        }
+        return result;
+    };
+    return src;
+}
+
+void save_weights_binary(const REX::lhe& lhe, const std::string& path_prefix) {
+    if (!lhe.weight_ids || lhe.weight_ids->empty()) {
+        throw std::runtime_error("save_weights_binary: lhe has no weight_ids to write");
+    }
+    for (std::size_t k = 0; k < lhe.weight_ids->size(); ++k) {
+        EventBuffer buffer(lhe.events.size(), 0, madspace::weight_file_layout);
+        for (std::size_t i = 0; i < lhe.events.size(); ++i) {
+            const auto& evt = lhe.events[i];
+            if (!evt || evt->wgts_.size() <= k) {
+                throw std::runtime_error(
+                    "save_weights_binary: event " + std::to_string(i) +
+                    " is missing weight '" + (*lhe.weight_ids)[k] + "'"
+                );
+            }
+            buffer.event(i).weight() = evt->wgts_[k];
+        }
+        std::string file_name =
+            path_prefix + "_" + sanitize_token((*lhe.weight_ids)[k]) + ".npy";
+        EventFile out(file_name, madspace::weight_file_layout, 0, EventFile::create);
+        out.write(buffer);
+    }
+}
+
+REX::eventSink open_weights_sink_binary(
+    const std::vector<std::string>& weight_ids, const std::string& path_prefix
+) {
+    if (weight_ids.empty()) {
+        throw std::runtime_error("open_weights_sink_binary: weight_ids must not be empty");
+    }
+    // One owner (this State) of every per-weight-id EventFile, reached
+    // identically by every closure below -- same rationale as
+    // open_lhe_sink_binary()'s State: write_close()'s clear() is then the
+    // sole, deterministic trigger for finalizing every file's on-disk
+    // header, not whichever copy of a shared_ptr happens to die last.
+    struct State {
+        std::vector<std::unique_ptr<EventFile>> files;
+    };
+    auto state = std::make_shared<State>();
+
+    REX::eventSink sink;
+    sink.write_open = [state, weight_ids, path_prefix](const REX::initNode&, const std::any&) {
+        state->files.clear();
+        state->files.reserve(weight_ids.size());
+        for (auto& id : weight_ids) {
+            std::string file_name = path_prefix + "_" + sanitize_token(id) + ".npy";
+            state->files.push_back(std::make_unique<EventFile>(
+                file_name, madspace::weight_file_layout, 0, EventFile::create
+            ));
+        }
+    };
+    sink.write_event = [state, weight_ids](std::shared_ptr<REX::event> ev) {
+        if (!ev) return;
+        if (ev->wgts_.size() < weight_ids.size()) {
+            throw std::runtime_error(
+                "open_weights_sink_binary: event is missing one or more of the " +
+                std::to_string(weight_ids.size()) + " expected weights"
+            );
+        }
+        for (std::size_t k = 0; k < weight_ids.size(); ++k) {
+            EventBuffer buffer(1, 0, madspace::weight_file_layout);
+            buffer.event(0).weight() = ev->wgts_[k];
+            state->files.at(k)->write(buffer);
+        }
+    };
+    sink.write_close = [state]() {
+        state->files.clear(); // destructs every EventFile, finalizing each header
+    };
+    return sink;
+}
+
+void load_weights_binary(
+    std::vector<std::shared_ptr<REX::event>>& events,
+    const std::vector<std::string>& weight_ids,
+    const std::string& path_prefix
+) {
+    for (auto& id : weight_ids) {
+        std::string file_name = path_prefix + "_" + sanitize_token(id) + ".npy";
+        EventFile file(file_name, madspace::weight_file_layout, 0, EventFile::load);
+        if (file.event_count() != events.size()) {
+            throw std::runtime_error(
+                "load_weights_binary: " + file_name + " has " +
+                std::to_string(file.event_count()) + " rows, expected " +
+                std::to_string(events.size()) + " (one per event)"
+            );
+        }
+        EventBuffer buffer(0, 0, madspace::weight_file_layout);
+        constexpr std::size_t chunk = 4096;
+        std::size_t offset = 0;
+        while (file.read(buffer, chunk)) {
+            for (std::size_t i = 0; i < buffer.event_count(); ++i) {
+                auto& evt = events.at(offset + i);
+                if (!evt) {
+                    throw std::runtime_error(
+                        "load_weights_binary: events[" + std::to_string(offset + i) +
+                        "] is null"
+                    );
+                }
+                evt->wgts_.push_back(buffer.event(i).weight());
+            }
+            offset += buffer.event_count();
+        }
+    }
+}
+
+REX::eventSink open_lhe_sink_binary(const std::string& path, std::size_t particle_count) {
+    // Single owner (unique_ptr) of the EventFile, reached identically by
+    // every closure below through the one shared State -- unlike each
+    // closure separately capturing its own shared_ptr<EventFile> copy, this
+    // makes write_close()'s reset() the sole, deterministic trigger for
+    // EventFile::~EventFile() (which finalizes the on-disk shape/header),
+    // rather than leaving that to whenever every copy happens to go out of
+    // scope.
+    struct State {
+        std::string path;
+        std::size_t particle_count;
+        std::unique_ptr<EventFile> file;
+    };
+    auto state = std::make_shared<State>();
+    state->path = path;
+    state->particle_count = particle_count;
+
+    REX::eventSink sink;
+    sink.write_open = [](const REX::initNode&, const std::any&) {
+        // The binary format carries no init-block information; see
+        // load_lhe_binary()'s own doc comment.
+    };
+    sink.write_event = [state](std::shared_ptr<REX::event> ev) {
+        if (!ev) return;
+        std::size_t n = ev->size();
+        if (!state->file) {
+            // Establishes this shard's fixed particle-slot capacity from
+            // whichever event lands here first, if the caller didn't pin
+            // one down explicitly; see open_lhe_sink_binary()'s doc
+            // comment.
+            std::size_t pc = state->particle_count != 0 ? state->particle_count : n;
+            state->file = std::make_unique<EventFile>(
+                state->path, lhe_event_layout(), pc, EventFile::create
+            );
+            state->particle_count = pc;
+        }
+        if (n > state->particle_count) {
+            throw std::runtime_error(
+                "open_lhe_sink_binary: event has " + std::to_string(n) +
+                " particles, more than this shard's fixed capacity (" +
+                std::to_string(state->particle_count) + ")"
+            );
+        }
+
+        DataLayout layout = lhe_event_layout();
+        EventBuffer buffer(1, state->particle_count, layout);
+        LHEEvent raw;
+        raw.process_id = static_cast<int>(ev->proc_id_);
+        raw.weight = ev->weight_;
+        raw.scale = ev->scale_;
+        raw.alpha_qed = ev->alphaEW_;
+        raw.alpha_qcd = ev->alphaS_;
+        buffer.event(0).from_lhe_event(raw);
+        for (std::size_t j = 0; j < n; ++j) {
+            buffer.particle(0, j).from_lhe_particle(parton_to_lhe_particle(*ev, j));
+        }
+        // particles [n, state->particle_count) stay zero-filled
+        // (status_code == 0, among everything else), matching
+        // read_binary_events()'s own padding convention.
+        state->file->write(buffer);
+    };
+    sink.write_close = [state]() {
+        // Finalizes the on-disk header/shape (EventFile::~EventFile()); a
+        // no-op, leaving no file on disk at all, if write_event() was never
+        // called (an empty shard).
+        state->file.reset();
+    };
+    return sink;
+}
+
+} // namespace madtrex
diff --git a/madtrex/src/madtrex.cpp b/madtrex/src/madtrex.cpp
new file mode 100644
index 0000000000..384429aeb6
--- /dev/null
+++ b/madtrex/src/madtrex.cpp
@@ -0,0 +1,1029 @@
+#include "madtrex/madtrex.hpp"
+
+#include <algorithm>
+#include <cmath>
+#include <cstring>
+#include <filesystem>
+#include <format>
+#include <fstream>
+#include <set>
+#include <stdexcept>
+
+namespace madtrex {
+
+LheFormat detect_lhe_format(const std::string& path) {
+    std::ifstream in(path, std::ios::binary);
+    if (!in) {
+        throw std::runtime_error("detect_lhe_format: failed to open " + path);
+    }
+    char magic[6] = {0};
+    in.read(magic, sizeof(magic));
+    if (in.gcount() == static_cast<std::streamsize>(sizeof(magic)) &&
+        std::memcmp(magic, "\x93NUMPY", sizeof(magic)) == 0) {
+        return LheFormat::binary;
+    }
+    return LheFormat::xml;
+}
+
+std::shared_ptr<REX::lhe> load_lhe_xml(const std::string& path) {
+    return std::make_shared<REX::lhe>(REX::load_lhef(path));
+}
+
+std::shared_ptr<REX::lhe> load_lhe_xml(std::istream& stream) {
+    return std::make_shared<REX::lhe>(REX::load_lhef(stream));
+}
+
+std::shared_ptr<REX::lhe> load_lhe(const std::string& path) {
+    if (detect_lhe_format(path) == LheFormat::binary) {
+        return load_lhe_binary(path);
+    }
+    return load_lhe_xml(path);
+}
+
+SplitResult split_events(
+    const std::string& path,
+    const std::string& output_prefix,
+    std::size_t n_shards,
+    std::size_t particle_count,
+    std::size_t batch_size,
+    const madspace::LHEMeta& meta
+) {
+    if (n_shards == 0) {
+        throw std::runtime_error("split_events: n_shards must be > 0");
+    }
+
+    auto shard_path = [&](std::size_t i) {
+        return output_prefix + "_shard" + std::to_string(i) + ".npy";
+    };
+    auto make_sink = [&](std::size_t i) -> REX::eventSink {
+        return open_lhe_sink_binary(shard_path(i), particle_count);
+    };
+
+    REX::eventSource source = detect_lhe_format(path) == LheFormat::xml
+        ? REX::open_lhe_source(path)
+        : open_lhe_source_binary(path, meta);
+    std::vector<std::size_t> counts = REX::split_source(source, n_shards, make_sink, batch_size);
+
+    SplitResult result;
+    result.counts = std::move(counts);
+    result.paths.resize(n_shards);
+    for (std::size_t i = 0; i < n_shards; ++i) {
+        if (result.counts[i] > 0) {
+            result.paths[i] = shard_path(i);
+        }
+    }
+    return result;
+}
+
+REX::tea::reweightor combine_shards(
+    const std::vector<std::string>& shard_event_paths,
+    const std::vector<std::string>& shard_weight_prefixes,
+    const std::vector<std::string>& launch_names,
+    const std::string& output_path,
+    bool binary_output,
+    const madspace::LHEMeta& meta
+) {
+    if (shard_event_paths.empty()) {
+        throw std::runtime_error("combine_shards: no shards given");
+    }
+    if (shard_event_paths.size() != shard_weight_prefixes.size()) {
+        throw std::runtime_error(
+            "combine_shards: shard_event_paths/shard_weight_prefixes size mismatch"
+        );
+    }
+    if (launch_names.empty()) {
+        throw std::runtime_error("combine_shards: no launch_names given");
+    }
+
+    // meta only informs the combined lhe's own init block (beam ids/
+    // energies, xsec, PDF set); every other shard only contributes events,
+    // so it's consulted at most once below, for whichever shard seeds
+    // `combined`, and nowhere else. Shards with an empty path (see
+    // SplitResult::paths) contributed zero events at split_events() time
+    // and are silently skipped here the same way.
+    std::shared_ptr<REX::lhe> combined;
+    for (std::size_t s = 0; s < shard_event_paths.size(); ++s) {
+        if (shard_event_paths[s].empty()) continue;
+        if (!combined) {
+            combined = load_lhe_binary(shard_event_paths[s], meta);
+            load_weights_binary(combined->events, launch_names, shard_weight_prefixes[s]);
+            continue;
+        }
+        auto shard = load_lhe_binary(shard_event_paths[s]);
+        load_weights_binary(shard->events, launch_names, shard_weight_prefixes[s]);
+        combined->events.insert(
+            combined->events.end(), shard->events.begin(), shard->events.end()
+        );
+    }
+    if (!combined) {
+        throw std::runtime_error("combine_shards: every shard is empty; nothing to combine");
+    }
+
+    combined->set_weight_ids(launch_names);
+    combined->sync_weight_ids();
+
+    REX::tea::reweightor result(std::move(*combined));
+    result.set_launch_names(launch_names);
+    // The one and only place in this workflow that computes cross
+    // sections/errors -- see this function's doc comment for why that
+    // matters -- and it does so over every event from every shard at once.
+    result.calc_xSecs();
+    result.calc_xErrs();
+
+    if (binary_output) {
+        save_weights_binary(result, output_path);
+    } else {
+        REX::write_lhef(result, output_path, true);
+    }
+
+    return result;
+}
+
+namespace {
+
+std::vector<long int> read_pdg_array(const json& arr) {
+    std::vector<long int> result;
+    result.reserve(arr.size());
+    for (auto& v : arr) result.push_back(v.get<long int>());
+    return result;
+}
+
+std::vector<FlavorChannel> expand_flavor_entry(const json& entry) {
+    int index = entry.at("index").get<int>();
+    bool mirror = entry.value("mirror", false);
+
+    std::vector<FlavorChannel> result;
+    for (auto& option : entry.at("options")) {
+        auto pdgs = read_pdg_array(option);
+        result.push_back(FlavorChannel{index, pdgs, false});
+        if (mirror) {
+            if (pdgs.size() < 2) {
+                throw std::runtime_error(
+                    "parse_subprocesses_json: flavor option has mirror=true but fewer "
+                    "than 2 external legs"
+                );
+            }
+            auto swapped = pdgs;
+            std::swap(swapped[0], swapped[1]);
+            result.push_back(FlavorChannel{index, std::move(swapped), true});
+        }
+    }
+    return result;
+}
+
+// Concatenates an event's incoming-leg PDGs (status == -1, in event storage
+// order) followed by its outgoing-leg PDGs (status == +1, in event storage
+// order). Legs with any other status (eg intermediate resonances) are not
+// part of this tuple.
+std::vector<long int> ordered_pdgs(const REX::event& e) {
+    std::vector<long int> incoming, outgoing;
+    for (std::size_t i = 0; i < e.size(); ++i) {
+        if (e.status_[i] == -1) incoming.push_back(e.pdg_[i]);
+        else if (e.status_[i] == 1) outgoing.push_back(e.pdg_[i]);
+    }
+    incoming.insert(incoming.end(), outgoing.begin(), outgoing.end());
+    return incoming;
+}
+
+std::vector<double> ordered_spins(const REX::event& e) {
+    std::vector<double> incoming, outgoing;
+    for (std::size_t i = 0; i < e.size(); ++i) {
+        if (e.status_[i] == -1) incoming.push_back(e.spin_[i]);
+        else if (e.status_[i] == 1) outgoing.push_back(e.spin_[i]);
+    }
+    incoming.insert(incoming.end(), outgoing.begin(), outgoing.end());
+    return incoming;
+}
+
+const FlavorChannel* match_flavor(const SubProcessSpec& spec, const REX::event& e) {
+    auto pdgs = ordered_pdgs(e);
+    for (auto& fc : spec.flavors) {
+        if (fc.pdgs == pdgs) return &fc;
+    }
+    return nullptr;
+}
+
+// A fixed helicity in the LHE SPINUP column: -1, +1, or 0 (scalars, and the
+// longitudinal helicity of a massive vector), as in the helicity table of
+// subprocesses.json. Anything else -- 9 for "unknown/summed" -- is not.
+bool is_fixed_helicity(double spin) {
+    return std::abs(spin - std::round(spin)) < 1e-6 && std::abs(spin) < 1.5;
+}
+
+// Explicit-key fallback used when the library doesn't export the optional
+// umami_supported_inputs/required_inputs introspection symbols. Momenta and
+// alpha_s are requested unconditionally; flavor index only when the subprocess
+// distinguishes more than one channel (so single-channel libraries that lack
+// UMAMI_IN_FLAVOR_INDEX support still work). Helicity index only with
+// helicity reweighting (see Driver::set_helicity_reweighting()), and only added when the
+// subprocess actually has helicity combinations to index into: older
+// generated libraries hard-error on UMAMI_IN_HELICITY_INDEX when it isn't
+// genuinely supported, and this fallback -- unlike default_input_keys() --
+// has no supported_inputs() to check that against, since introspection
+// itself is what failed to land us here in the first place; helicity
+// reweighting is the implementation's explicit assertion that it knows better.
+// SubProcessSpec's own helicity_index (via make_event_checker) populates
+// event::helicity_ whenever fixed helicities are present regardless of this
+// flag, ready for use once helicity reweighting is turned on.
+// Why the helicity index of spec's events can't be mapped onto api's
+// helicities, or "" if it can: the subprocess needs a helicity table (from
+// subprocesses.json, which also defines MG7's SPINUP values) that enumerates
+// exactly the helicities the library does (UMAMI_META_HELICITY_COUNT), with
+// one helicity per external leg. UMAMI exposes the library's helicity count
+// but not its table, so the count is what can be checked.
+std::string helicity_map_problem(const SubProcessSpec& spec, const madspace::MatrixElementApi& api) {
+    if (spec.helicities.empty()) {
+        return "its subprocesses.json entry has no helicity table";
+    }
+    std::size_t legs = spec.incoming.size() + spec.outgoing.size();
+    for (auto& row : spec.helicities) {
+        if (row.size() != legs) {
+            return std::format(
+                "its helicity table has rows of {} helicities for {} external legs",
+                row.size(), legs
+            );
+        }
+    }
+    std::size_t library_count;
+    try {
+        library_count = api.helicity_count();
+    } catch (const std::runtime_error&) {
+        return "the library does not report its helicity count";
+    }
+    if (library_count != spec.helicities.size()) {
+        return std::format(
+            "its helicity table has {} helicity combinations, the library {}",
+            spec.helicities.size(), library_count
+        );
+    }
+    return "";
+}
+
+// Whether api accepts UMAMI_IN_HELICITY_INDEX, or true if it doesn't say (the
+// input_keys_for() fallback below then decides).
+bool accepts_helicity_index(const madspace::MatrixElementApi& api) {
+    try {
+        auto supported = api.supported_inputs();
+        return UMAMI_IN_HELICITY_INDEX < supported.size() && supported[UMAMI_IN_HELICITY_INDEX];
+    } catch (const std::runtime_error&) {
+        return true;
+    }
+}
+
+std::vector<UmamiInputKey> input_keys_for(const SubProcessSpec& spec, bool helicity_reweighting) {
+    std::vector<UmamiInputKey> keys{UMAMI_IN_MOMENTA, UMAMI_IN_ALPHA_S};
+    std::set<int> distinct_indices;
+    for (auto& fc : spec.flavors) distinct_indices.insert(fc.index);
+    if (distinct_indices.size() > 1) keys.push_back(UMAMI_IN_FLAVOR_INDEX);
+    if (helicity_reweighting && !spec.helicities.empty()) keys.push_back(UMAMI_IN_HELICITY_INDEX);
+    return keys;
+}
+
+// The subprocess of the original process directory (normaliser_directory,
+// parsed into candidates) that spec, from a process regenerated in another
+// model, is normalised against: the one describing exactly the same matrix
+// element call -- the same incoming and outgoing legs in the same order,
+// the same flavor channels (UMAMI flavor index, PDG ids in leg order,
+// mirroring) and the same helicity table -- so that an event's flavor and
+// helicity indices mean the same in both libraries. Throws if there is none.
+const SubProcessSpec& matching_normaliser_spec(
+    const SubProcessSpec& spec,
+    const std::vector<SubProcessSpec>& candidates,
+    const std::string& normaliser_directory
+) {
+    auto same_flavors = [](const SubProcessSpec& a, const SubProcessSpec& b) {
+        return std::equal(a.flavors.begin(), a.flavors.end(), b.flavors.begin(), b.flavors.end(),
+                          [](const FlavorChannel& x, const FlavorChannel& y) {
+                              return x.index == y.index && x.pdgs == y.pdgs && x.mirrored == y.mirrored;
+                          });
+    };
+    for (const auto& candidate : candidates) {
+        if (candidate.incoming == spec.incoming && candidate.outgoing == spec.outgoing &&
+            same_flavors(candidate, spec) && candidate.helicities == spec.helicities) {
+            return candidate;
+        }
+    }
+    throw std::runtime_error(std::format(
+        "Driver::load_process: subprocess {} has no identical counterpart (same legs, "
+        "flavor channels and helicities) in {}, so the matrix element its events were "
+        "generated with is not known exactly",
+        spec.me_path_template, normaliser_directory
+    ));
+}
+
+} // namespace
+
+std::vector<SubProcessSpec> parse_subprocesses_json(const std::string& subprocesses_json_path) {
+    std::ifstream in(subprocesses_json_path);
+    if (!in) {
+        throw std::runtime_error(
+            "parse_subprocesses_json: failed to open " + subprocesses_json_path
+        );
+    }
+    json data;
+    in >> data;
+    if (!data.is_array()) {
+        throw std::runtime_error(
+            "parse_subprocesses_json: expected a top-level JSON array in " +
+            subprocesses_json_path
+        );
+    }
+
+    std::vector<SubProcessSpec> specs;
+    specs.reserve(data.size());
+    for (auto& entry : data) {
+        SubProcessSpec spec;
+        spec.incoming = read_pdg_array(entry.at("incoming"));
+        spec.outgoing = read_pdg_array(entry.at("outgoing"));
+        spec.me_path_template = entry.at("me_path").get<std::string>();
+        spec.source_path = entry.value("path", std::string());
+        spec.diagram_count = entry.value("diagram_count", std::size_t(0));
+
+        for (auto& flavor_entry : entry.at("flavors")) {
+            auto expanded = expand_flavor_entry(flavor_entry);
+            spec.flavors.insert(spec.flavors.end(), expanded.begin(), expanded.end());
+        }
+
+        if (entry.contains("helicities")) {
+            for (auto& hel : entry.at("helicities")) {
+                spec.helicities.push_back(hel.get<std::vector<int>>());
+            }
+        }
+
+        specs.push_back(std::move(spec));
+    }
+    return specs;
+}
+
+REX::eventBelongs make_event_checker(const SubProcessSpec& spec) {
+    REX::eventBelongs checker;
+
+    // One template event per un-mirrored flavor option: the default
+    // external_legs_comparator matches purely on PDG multiset per status,
+    // which mirroring doesn't change, so mirrored variants don't need their
+    // own template.
+    for (auto& fc : spec.flavors) {
+        if (fc.mirrored) continue;
+        std::size_t n = fc.pdgs.size();
+        REX::event templ(n);
+        std::vector<short int> status(n, 1);
+        for (std::size_t i = 0; i < spec.incoming.size() && i < n; ++i) status[i] = -1;
+        templ.set_pdg(fc.pdgs);
+        templ.set_status(status);
+        checker.add_event(templ);
+    }
+
+    checker.flavor_index = [spec](REX::event& e) -> std::size_t {
+        auto* fc = match_flavor(spec, e);
+        if (!fc) {
+            throw std::runtime_error(
+                "make_event_checker: event matched the subprocess by PDG content but "
+                "not positionally by any known flavor channel; the assumed event leg "
+                "ordering (incoming, then outgoing, both in event storage order) may "
+                "not hold for this event"
+            );
+        }
+        return static_cast<std::size_t>(fc->index);
+    };
+
+    // The index of the event's helicity configuration in the subprocess's
+    // helicity table -- the table the matrix element enumerates its
+    // helicities in (UMAMI_IN_HELICITY_INDEX), and the one MG7 writes SPINUP
+    // from -- read from SPINUP in the order of the matched flavor channel's
+    // legs (an MG7 mirror is a rotation, keeping the leg order; a beam-
+    // swapped "mirrored" channel has its two incoming spins swapped back).
+    // REX::npos (-1) without fixed helicities, helicity_not_in_table for a
+    // configuration missing from the table; both only matter, and are
+    // reported, with helicity reweighting (see TupperWare's evaluate()).
+    if (spec.helicities.empty()) return checker; // no table, no index
+    checker.helicity_index = [spec](REX::event& e) -> std::size_t {
+        auto* fc = match_flavor(spec, e);
+        if (!fc) {
+            throw std::runtime_error(
+                "make_event_checker: event matched the subprocess by PDG content but "
+                "not positionally by any known flavor channel"
+            );
+        }
+        auto spins = ordered_spins(e);
+        if (fc->mirrored && spins.size() >= 2) std::swap(spins[0], spins[1]);
+
+        for (double s : spins) {
+            if (!is_fixed_helicity(s)) return REX::npos; // no fixed helicity: sum over helicities
+        }
+        std::vector<int> rounded;
+        rounded.reserve(spins.size());
+        for (double s : spins) rounded.push_back(static_cast<int>(std::lround(s)));
+
+        for (std::size_t i = 0; i < spec.helicities.size(); ++i) {
+            if (spec.helicities[i] == rounded) return i;
+        }
+        return static_cast<std::size_t>(helicity_not_in_table);
+    };
+
+    return checker;
+}
+
+namespace {
+
+// Scans dir for regular files whose name starts with prefix and ends with
+// suffix, returning the (possibly empty) middle portion of each match --
+// ie each one's resolved placeholder value(s), still concatenated together
+// if the caller is matching more than one placeholder at once.
+std::vector<std::string> scan_placeholder_values(
+    const std::filesystem::path& dir, const std::string& prefix, const std::string& suffix
+) {
+    namespace fs = std::filesystem;
+    std::vector<std::string> found;
+    for (auto& entry : fs::directory_iterator(dir)) {
+        if (!entry.is_regular_file()) continue;
+        std::string name = entry.path().filename().string();
+        if (name.size() < prefix.size() + suffix.size()) continue;
+        if (name.compare(0, prefix.size(), prefix) != 0) continue;
+        if (name.compare(name.size() - suffix.size(), suffix.size(), suffix) != 0) continue;
+        found.push_back(name.substr(prefix.size(), name.size() - prefix.size() - suffix.size()));
+    }
+    return found;
+}
+
+// Picks one of found (each a candidate value for one placeholder) via
+// priority: the first priority entry with a match wins; if none of the
+// priority entries match but exactly one value was found, that one is used
+// regardless; otherwise throws, listing what was found, so the caller can
+// pass an explicit priority and try again. what/setter_hint name the
+// placeholder kind and the Driver setter to use, for the error message
+// only. found must outlive the returned reference.
+const std::string& pick_priority(
+    const std::vector<std::string>& found,
+    const std::vector<std::string>& priority,
+    const std::string& what,
+    const std::string& setter_hint
+) {
+    for (auto& preferred : priority) {
+        auto it = std::find(found.begin(), found.end(), preferred);
+        if (it != found.end()) return *it;
+    }
+    if (found.size() == 1) {
+        return found.front();
+    }
+    std::string found_list;
+    for (auto& f : found) {
+        if (!found_list.empty()) found_list += ", ";
+        found_list += f;
+    }
+    throw std::runtime_error(
+        "resolve_me_path: multiple " + what + " variants found (" + found_list +
+        ") and none is listed in the " + what + " priority; call " + setter_hint +
+        " to disambiguate"
+    );
+}
+
+} // namespace
+
+std::string resolve_me_path(
+    const std::string& process_directory,
+    const std::string& me_path_template,
+    const std::vector<std::string>& device_priority,
+    const std::vector<std::string>& fptype_priority
+) {
+    namespace fs = std::filesystem;
+    fs::path full_template = fs::path(process_directory) / me_path_template;
+    std::string filename_template = full_template.filename().string();
+    fs::path dir = full_template.parent_path();
+
+    static const std::string device_placeholder = "{device}";
+    static const std::string fptype_placeholder = "{fptype}";
+    auto device_pos = filename_template.find(device_placeholder);
+    auto fptype_pos = filename_template.find(fptype_placeholder);
+
+    if (device_pos == std::string::npos && fptype_pos == std::string::npos) {
+        if (!fs::exists(full_template)) {
+            throw std::runtime_error(
+                "resolve_me_path: " + full_template.string() + " does not exist"
+            );
+        }
+        return full_template.string();
+    }
+
+    if (!fs::exists(dir) || !fs::is_directory(dir)) {
+        throw std::runtime_error(
+            "resolve_me_path: directory " + dir.string() + " does not exist"
+        );
+    }
+
+    if (device_pos != std::string::npos && fptype_pos != std::string::npos) {
+        // Both placeholders present: {device} is expected to come first,
+        // matching every me_path template MadMatrix itself ever generates
+        // (see me_lib_format in madmatrix/output.py / MADMATRIX_LIB in
+        // madmatrix.mk).
+        if (fptype_pos < device_pos) {
+            throw std::runtime_error(
+                "resolve_me_path: me_path template has {fptype} before {device}, which is "
+                "not supported: " + filename_template
+            );
+        }
+        std::string prefix = filename_template.substr(0, device_pos);
+        std::string middle_sep = filename_template.substr(
+            device_pos + device_placeholder.size(),
+            fptype_pos - (device_pos + device_placeholder.size())
+        );
+        std::string suffix = filename_template.substr(fptype_pos + fptype_placeholder.size());
+        if (middle_sep.empty()) {
+            throw std::runtime_error(
+                "resolve_me_path: {device} and {fptype} must be separated by at least one "
+                "literal character: " + filename_template
+            );
+        }
+
+        std::vector<std::pair<std::string, std::string>> found; // (device, fptype)
+        for (auto& value : scan_placeholder_values(dir, prefix, suffix)) {
+            // middle_sep (eg "_") is expected to separate the device and
+            // fptype values within value; search from the end since fptype
+            // is the short, trailing token (one of SUPPORTED_FPTYPES in
+            // madmatrix.mk: "d", "f", "m"), so this is robust even if a
+            // device name ever happened to contain middle_sep itself.
+            auto sep_pos = value.rfind(middle_sep);
+            if (sep_pos == std::string::npos) continue;
+            found.emplace_back(value.substr(0, sep_pos), value.substr(sep_pos + middle_sep.size()));
+        }
+        if (found.empty()) {
+            throw std::runtime_error(
+                "resolve_me_path: no library matching " + prefix + "*" + middle_sep + "*" +
+                suffix + " found in " + dir.string()
+            );
+        }
+
+        // Deduplicated (a device with several fptype variants on disk must
+        // not look "ambiguous" on the device axis alone -- pick_priority's
+        // found.size()==1 auto-select needs found to hold distinct values).
+        std::vector<std::string> devices;
+        std::set<std::string> seen_devices;
+        for (auto& [d, f] : found) {
+            if (seen_devices.insert(d).second) devices.push_back(d);
+        }
+        const std::string& chosen_device =
+            pick_priority(devices, device_priority, "device", "Driver::set_device_priority(...)");
+
+        std::vector<std::string> fptypes_for_device;
+        std::set<std::string> seen_fptypes;
+        for (auto& [d, f] : found) {
+            if (d == chosen_device && seen_fptypes.insert(f).second) fptypes_for_device.push_back(f);
+        }
+        const std::string& chosen_fptype = pick_priority(
+            fptypes_for_device, fptype_priority, "floating-point type",
+            "Driver::set_fptype_priority(...)"
+        );
+
+        return (dir / (prefix + chosen_device + middle_sep + chosen_fptype + suffix)).string();
+    }
+
+    // Exactly one of the two placeholders is present (either {device}
+    // alone -- the pre-existing convention -- or {fptype} alone).
+    bool is_device = device_pos != std::string::npos;
+    std::size_t pos = is_device ? device_pos : fptype_pos;
+    std::size_t placeholder_len = (is_device ? device_placeholder : fptype_placeholder).size();
+    std::string prefix = filename_template.substr(0, pos);
+    std::string suffix = filename_template.substr(pos + placeholder_len);
+
+    auto found = scan_placeholder_values(dir, prefix, suffix);
+    if (found.empty()) {
+        throw std::runtime_error(
+            "resolve_me_path: no library matching " + prefix + "*" + suffix + " found in " +
+            dir.string()
+        );
+    }
+    const std::string& chosen = is_device
+        ? pick_priority(found, device_priority, "device", "Driver::set_device_priority(...)")
+        : pick_priority(
+              found, fptype_priority, "floating-point type", "Driver::set_fptype_priority(...)"
+          );
+    return (dir / (prefix + chosen + suffix)).string();
+}
+
+Driver::Driver() : _context(madspace::default_context()) {}
+
+Driver::Driver(madspace::ContextPtr context) : _context(std::move(context)) {}
+
+Driver& Driver::set_device_priority(std::vector<std::string> priority) {
+    _device_priority = std::move(priority);
+    return *this;
+}
+
+const std::vector<std::string>& Driver::device_priority() const { return _device_priority; }
+
+Driver& Driver::set_fptype_priority(std::vector<std::string> priority) {
+    _fptype_priority = std::move(priority);
+    return *this;
+}
+
+const std::vector<std::string>& Driver::fptype_priority() const { return _fptype_priority; }
+
+Driver& Driver::set_normaliser_fptype_priority(std::vector<std::string> priority) {
+    _normaliser_fptype_priority = std::move(priority);
+    _normaliser_fptype_priority_set = true;
+    return *this;
+}
+
+const std::vector<std::string>& Driver::normaliser_fptype_priority() const {
+    return _normaliser_fptype_priority_set ? _normaliser_fptype_priority : _fptype_priority;
+}
+
+Driver& Driver::set_helicity_reweighting(bool helicity_reweighting) {
+    _helicity_reweighting = helicity_reweighting;
+    return *this;
+}
+
+bool Driver::helicity_reweighting() const { return _helicity_reweighting; }
+
+Driver& Driver::set_normaliser_process_directory(
+    const std::string& process_directory, const std::string& param_card
+) {
+    _normaliser_process_directory = process_directory;
+    _normaliser_param_card = param_card;
+    return *this;
+}
+
+const std::string& Driver::normaliser_process_directory() const {
+    return _normaliser_process_directory;
+}
+
+Driver& Driver::set_normaliser_param_card(const std::string& param_card) {
+    _normaliser_param_card = param_card;
+    return *this;
+}
+
+const std::string& Driver::normaliser_param_card() const { return _normaliser_param_card; }
+
+Driver& Driver::set_use_stored_normalisation(bool use_stored_normalisation) {
+    _use_stored_normalisation = use_stored_normalisation;
+    return *this;
+}
+
+bool Driver::use_stored_normalisation() const { return _use_stored_normalisation; }
+
+Driver& Driver::set_use_helicity(bool use_helicity) { return set_helicity_reweighting(use_helicity); }
+
+bool Driver::use_helicity() const { return _helicity_reweighting; }
+
+REX::verbosity& Driver::verbose() { return _warehouse.verbose(); }
+
+const REX::verbosity& Driver::verbose() const { return _warehouse.verbose(); }
+
+Driver& Driver::set_verbose(bool v) {
+    _warehouse.set_verbose(v);
+    return *this;
+}
+
+bool Driver::is_verbose() const { return _warehouse.is_verbose(); }
+
+Driver& Driver::set_max_batch(std::size_t max_batch) {
+    _warehouse.set_max_batch(max_batch);
+    return *this;
+}
+
+std::size_t Driver::max_batch() const { return _warehouse.max_batch(); }
+
+Driver& Driver::set_threads(std::size_t threads) {
+    _warehouse.set_threads(threads);
+    return *this;
+}
+
+std::size_t Driver::threads() const { return _warehouse.threads(); }
+
+Driver& Driver::set_min_batch(std::size_t min_batch) {
+    _warehouse.set_min_batch(min_batch);
+    return *this;
+}
+
+std::size_t Driver::min_batch() const { return _warehouse.min_batch(); }
+
+Driver& Driver::set_max_deferred_iterations(std::size_t max_deferred_iterations) {
+    _warehouse.set_max_deferred_iterations(max_deferred_iterations);
+    return *this;
+}
+
+std::size_t Driver::max_deferred_iterations() const {
+    return _warehouse.max_deferred_iterations();
+}
+
+Driver& Driver::set_streaming(bool enable, std::size_t batch_size) {
+    _streaming = enable;
+    _stream_batch_size = batch_size;
+    return *this;
+}
+
+bool Driver::streaming() const { return _streaming; }
+
+std::size_t Driver::stream_batch_size() const { return _stream_batch_size; }
+
+Driver& Driver::load_process(
+    const std::string& subprocesses_json_path, const std::string& param_card
+) {
+    namespace fs = std::filesystem;
+    fs::path json_path(subprocesses_json_path);
+    fs::path process_directory = json_path.parent_path().parent_path();
+
+    std::string resolved_param_card = param_card;
+    if (resolved_param_card.empty()) {
+        resolved_param_card = (process_directory / "Cards" / "param_card.dat").string();
+    }
+
+    auto specs = parse_subprocesses_json(subprocesses_json_path);
+    // The normaliser is the matrix element the events were generated with:
+    // this process directory's own, unless set_normaliser_process_directory()
+    // names the original one (a "change model" reweighting), whose
+    // subprocesses must then be the same ones (see matching_normaliser_spec()).
+    fs::path normaliser_directory = process_directory;
+    std::string normaliser_param_card = resolved_param_card;
+    std::vector<SubProcessSpec> normaliser_specs;
+    if (!_normaliser_process_directory.empty()) {
+        normaliser_directory = _normaliser_process_directory;
+        normaliser_param_card = _normaliser_param_card.empty()
+            ? (normaliser_directory / "Cards" / "param_card.dat").string()
+            : _normaliser_param_card;
+        normaliser_specs = parse_subprocesses_json(
+            (normaliser_directory / "SubProcesses" / "subprocesses.json").string()
+        );
+    } else if (!_normaliser_param_card.empty()) {
+        normaliser_param_card = _normaliser_param_card;
+    }
+    // See set_normaliser_fptype_priority()'s own doc comment: only load a
+    // genuinely separate normaliser api when it would actually resolve to a
+    // *different* library than the reweight-target api -- when they're
+    // equal (the common case: no explicit call, or an explicit call that
+    // happens to match), sharing one api for both roles is both correct and
+    // cheaper (skips a redundant library load/context registration).
+    const auto& normaliser_priority = normaliser_fptype_priority();
+    bool separate_normaliser = normaliser_priority != _fptype_priority ||
+                               !_normaliser_process_directory.empty() ||
+                               normaliser_param_card != resolved_param_card;
+    for (auto& spec : specs) {
+        std::string me_path = resolve_me_path(
+            process_directory.string(), spec.me_path_template, _device_priority, _fptype_priority
+        );
+        const auto& api = _context->load_matrix_element(me_path, resolved_param_card);
+        detail::forget_api_parameters(api);
+        const SubProcessSpec& normaliser_spec = _normaliser_process_directory.empty()
+            ? spec
+            : matching_normaliser_spec(spec, normaliser_specs, normaliser_directory.string());
+        // Normaliser must reflect the sample's ORIGINAL configuration
+        // (see set_normaliser_fptype_priority()), so when separate it gets
+        // its own, independently-loaded api -- a different compiled library
+        // (different fptype, or the original process directory's) --
+        // rather than reusing the reweight-target api above. Deliberately
+        // NOT passed to add_api(): it must never be touched by
+        // ParamHandler's per-launch apply_launch() overrides, always
+        // reflecting the untouched baseline every launch is reweighted
+        // relative to.
+        std::string normaliser_me_path;
+        const madspace::MatrixElementApi* normaliser_api = &api;
+        if (separate_normaliser) {
+            normaliser_me_path = resolve_me_path(
+                normaliser_directory.string(), normaliser_spec.me_path_template,
+                _device_priority, normaliser_priority
+            );
+            normaliser_api =
+                &_context->load_matrix_element(normaliser_me_path, normaliser_param_card);
+            detail::forget_api_parameters(*normaliser_api);
+        }
+        auto checker = make_event_checker(spec);
+        // Helicity reweighting for this subprocess needs both a library that
+        // accepts UMAMI_IN_HELICITY_INDEX and a helicity map from the events'
+        // SPINUP values onto the library's helicity enumeration
+        // (helicity_map_problem()) -- for the reweighted matrix element and
+        // the normaliser alike, which are always evaluated the same way:
+        // both at the event's helicity, or both summed over helicities.
+        bool helicity_reweighting = _helicity_reweighting;
+        for (const auto* lib : {&api, normaliser_api}) {
+            if (!helicity_reweighting) break;
+            std::string problem = accepts_helicity_index(*lib)
+                ? helicity_map_problem(spec, *lib)
+                : "it does not support per-event helicities (UMAMI_IN_HELICITY_INDEX)";
+            if (!problem.empty()) {
+                detail::warn_once(
+                    "no-helicity-map:" + lib->file_name(),
+                    std::format(
+                        "helicity reweighting is on, but the events of {} cannot be "
+                        "mapped onto its helicities ({}); using helicity-summed matrix "
+                        "elements for it",
+                        lib->file_name(), problem
+                    )
+                );
+                helicity_reweighting = false;
+            }
+        }
+        auto keys_for = [&](const madspace::MatrixElementApi& lib, const SubProcessSpec& lib_spec) {
+            std::pair<std::vector<UmamiInputKey>, std::vector<UmamiOutputKey>> keys;
+            try {
+                keys = {default_input_keys(lib, helicity_reweighting), default_output_keys(lib)};
+            } catch (const std::runtime_error&) {
+                keys = {input_keys_for(lib_spec, helicity_reweighting), {UMAMI_OUT_MATRIX_ELEMENT}};
+            }
+            return keys;
+        };
+        auto [input_keys, output_keys] = keys_for(api, spec);
+        // Same (context, me_path, resolved_param_card) that api was just
+        // loaded from, so TupperWare::build() can reload an independent
+        // instance per max_batch overflow batch instead of every batch
+        // sharing this one -- but only once set_max_batch() actually asks
+        // for that (see WareHouse::set_max_batch()/TupperWare::build());
+        // registering this unconditionally here costs nothing when it's
+        // not.
+        ReloadSpec reload{reload_context(), me_path, resolved_param_card};
+
+        auto& tw = _warehouse.add_process()
+            .add_api(api, input_keys, output_keys, reload)
+            .set_event_checker(std::move(checker));
+        if (_use_stored_normalisation) {
+            tw.set_stored_normalisation(stored_matrix_element_key);
+        }
+        // Beam-swapped events of mirrored flavor channels go back into the
+        // matrix element's leg order (see IncomingSwapped).
+        if (std::any_of(spec.flavors.begin(), spec.flavors.end(), [](auto& fc) { return fc.mirrored; })) {
+            tw.set_incoming_swapped([spec](const REX::event& e) {
+                auto* fc = match_flavor(spec, e);
+                return fc != nullptr && fc->mirrored;
+            });
+        }
+
+        if (separate_normaliser) {
+            auto [normaliser_input_keys, normaliser_output_keys] =
+                keys_for(*normaliser_api, normaliser_spec);
+            auto has_helicity = [](const std::vector<UmamiInputKey>& keys) {
+                return std::find(keys.begin(), keys.end(), UMAMI_IN_HELICITY_INDEX) != keys.end();
+            };
+            if (has_helicity(normaliser_input_keys) != has_helicity(input_keys)) {
+                throw std::runtime_error(std::format(
+                    "Driver::load_process: {} and its normaliser {} would not be evaluated "
+                    "with the same helicity treatment",
+                    api.file_name(), normaliser_api->file_name()
+                ));
+            }
+            ReloadSpec normaliser_reload{reload_context(), normaliser_me_path, normaliser_param_card};
+            tw.set_normaliser(*normaliser_api, normaliser_input_keys, normaliser_output_keys, normaliser_reload);
+        } else {
+            // Explicit normaliser, using the same resolved keys as
+            // add_api(): procReweightor::initialise() can only fall back to
+            // an *immediate* (already-computed) slot of reweight_functions
+            // when no normaliser is set, and a cuda/hip api registered here
+            // lands in a genuinely deferred slot instead (see
+            // TupperWare::build()), which can't stand in as an implicit
+            // normaliser (see teaRex.cc). Passing the same input_keys/
+            // output_keys explicitly also avoids set_normaliser()
+            // re-deriving them via default_input_keys()/
+            // default_output_keys(), which would throw for real MadMatrix
+            // libraries that don't implement the optional UMAMI
+            // introspection symbols -- the same reason add_api() above
+            // needed the try/catch fallback in the first place.
+            tw.set_normaliser(api, input_keys, output_keys, reload);
+        }
+        _specs.push_back(std::move(spec));
+    }
+    return *this;
+}
+
+Driver& Driver::load_process_directory(
+    const std::string& process_directory, const std::string& param_card
+) {
+    namespace fs = std::filesystem;
+    fs::path json_path = fs::path(process_directory) / "SubProcesses" / "subprocesses.json";
+    return load_process(json_path.string(), param_card);
+}
+
+Driver& Driver::load_param_reweighting(const std::string& rwgt_path) {
+    _param_handler = ParamHandler();
+    _param_handler.add_apis(_warehouse).read_rwgt_card(rwgt_path);
+    _warehouse.set_iterators(_param_handler.iterators());
+    _warehouse.set_launch_names(_param_handler.launch_names());
+    return *this;
+}
+
+ParamHandler& Driver::param_handler() { return _param_handler; }
+
+const ParamHandler& Driver::param_handler() const { return _param_handler; }
+
+Driver& Driver::load_events(const std::string& path, const madspace::LHEMeta& meta) {
+    LheFormat format = detect_lhe_format(path);
+    if (_streaming) {
+        auto source = std::make_shared<REX::eventSource>(
+            format == LheFormat::binary ? open_lhe_source_binary(path, meta)
+                                         : REX::open_lhe_source(path)
+        );
+        _warehouse.build(REX::lhe::open_streaming(source, _stream_batch_size));
+        _binary_output = (format == LheFormat::binary);
+        return *this;
+    }
+    std::shared_ptr<REX::lhe> loaded =
+        format == LheFormat::binary ? load_lhe_binary(path, meta) : load_lhe_xml(path);
+    _warehouse.build(std::move(*loaded));
+    _binary_output = (format == LheFormat::binary);
+    return *this;
+}
+
+bool Driver::binary_output() const { return _binary_output; }
+
+Driver& Driver::set_binary_output(bool binary) {
+    _binary_output = binary;
+    return *this;
+}
+
+Driver& Driver::write_weights(const std::string& path) {
+    if (!_warehouse.built()) {
+        throw std::runtime_error(
+            "Driver::write_weights: WareHouse has not been built yet; call load_events() first"
+        );
+    }
+    auto& built = _warehouse.get();
+    if (_binary_output) {
+        save_weights_binary(built, path);
+    } else {
+        // the stored matrix elements of binary input are MadtRex's own, not
+        // part of the events (REX::event::print() writes every extra)
+        for (auto& event : built.events) {
+            if (event) event->extra.erase(stored_matrix_element_key);
+        }
+        REX::write_lhef(built, path, true);
+    }
+    return *this;
+}
+
+Driver& Driver::run_streaming(const std::string& output_path, bool include_ids) {
+    if (!_streaming) {
+        throw std::runtime_error(
+            "Driver::run_streaming: streaming is not enabled; call set_streaming(true) "
+            "before load_events()"
+        );
+    }
+    if (!_warehouse.built()) {
+        throw std::runtime_error(
+            "Driver::run_streaming: WareHouse has not been built yet; call load_events() "
+            "first"
+        );
+    }
+    if (_binary_output) {
+        // A streaming source never carries a header (see
+        // open_lhe_source_binary()/open_lhe_source()'s own header-less
+        // XML case), so reweightor::run_streaming()'s own weight_ids
+        // resolution -- pre-existing header ids, if any, plus
+        // launch_names, in that order -- reduces to just launch_names
+        // here; open_weights_sink_binary() needs that same final list up
+        // front, unlike REX::open_lhe_sink()'s XML sink, which discovers
+        // it from write_open()'s own initNode/header argument instead.
+        const auto& launch_names = _warehouse.launch_names();
+        if (launch_names.empty()) {
+            throw std::runtime_error(
+                "Driver::run_streaming: binary_output() requires launch_names to be set "
+                "(eg via load_param_reweighting() or warehouse().set_launch_names()), "
+                "since binary weight files are named per weight id"
+            );
+        }
+        auto sink = open_weights_sink_binary(launch_names, output_path);
+        _warehouse.get().run_streaming(sink, include_ids);
+        return *this;
+    }
+    auto sink = REX::open_lhe_sink(output_path, include_ids);
+    _warehouse.get().run_streaming(sink, include_ids);
+    return *this;
+}
+
+Driver& Driver::reweight(
+    const std::string& process_directory,
+    const std::string& events_path,
+    const std::string& output_path,
+    const std::string& rwgt_path,
+    const std::string& param_card,
+    const madspace::LHEMeta& meta
+) {
+    load_process_directory(process_directory, param_card);
+    if (!rwgt_path.empty()) {
+        load_param_reweighting(rwgt_path);
+    }
+    load_events(events_path, meta);
+    if (_streaming) {
+        run_streaming(output_path);
+    } else {
+        _warehouse.get().run();
+        write_weights(output_path);
+    }
+    return *this;
+}
+
+madspace::ContextPtr Driver::context() const { return _context; }
+
+madspace::ContextPtr Driver::reload_context() {
+    // The instances reloaded per batch (ReloadSpec) are only evaluated from
+    // REX's worker threads, ie through process instance 0: a context of one
+    // thread initializes just that one per reload, rather than one per
+    // thread of _context
+    if (!_reload_context) {
+        _reload_context = std::make_shared<madspace::Context>(_context->device(), 1);
+    }
+    return _reload_context;
+}
+
+WareHouse& Driver::warehouse() { return _warehouse; }
+
+const WareHouse& Driver::warehouse() const { return _warehouse; }
+
+const std::vector<SubProcessSpec>& Driver::specs() const { return _specs; }
+
+} // namespace madtrex
diff --git a/madtrex/src/param_handler.cpp b/madtrex/src/param_handler.cpp
new file mode 100644
index 0000000000..5f01a9c934
--- /dev/null
+++ b/madtrex/src/param_handler.cpp
@@ -0,0 +1,103 @@
+#include "madtrex/param_handler.hpp"
+
+#include <fstream>
+#include <mutex>
+#include <stdexcept>
+
+namespace madtrex {
+
+ParamHandler& ParamHandler::add_api(const madspace::MatrixElementApi& api) {
+    if (!api.supports_set_parameter()) {
+        throw std::runtime_error(
+            "ParamHandler::add_api: matrix element " + api.file_name() +
+            " does not implement umami_set_parameter; use REX::tea::param_rwgt's "
+            "file-based parameter reweighting for this library instead"
+        );
+    }
+    for (auto* existing : _apis) {
+        if (existing == &api) return *this;
+    }
+    _apis.push_back(&api);
+    return *this;
+}
+
+ParamHandler& ParamHandler::add_apis(const std::vector<const madspace::MatrixElementApi*>& apis) {
+    for (auto* api : apis) add_api(*api);
+    return *this;
+}
+
+ParamHandler& ParamHandler::add_apis(const WareHouse& house) {
+    for (auto& tw : house.processes()) add_apis(tw.apis());
+    return *this;
+}
+
+ParamHandler& ParamHandler::read_rwgt_card(std::istream& rwgt_in) {
+    _card_iter.parse_rwgt_card(rwgt_in);
+    return *this;
+}
+
+ParamHandler& ParamHandler::read_rwgt_card(const std::string& rwgt_path) {
+    std::ifstream rwgt_in(rwgt_path);
+    if (!rwgt_in) {
+        throw std::runtime_error("ParamHandler::read_rwgt_card: failed to open " + rwgt_path);
+    }
+    return read_rwgt_card(rwgt_in);
+}
+
+std::size_t ParamHandler::launch_count() const { return _card_iter.cards.size(); }
+
+bool ParamHandler::apply_launch(std::size_t idx) {
+    if (idx >= _card_iter.cards.size()) {
+        throw std::out_of_range("ParamHandler::apply_launch: launch index out of range");
+    }
+    if (_apis.empty()) {
+        throw std::runtime_error("ParamHandler::apply_launch: no matrix element apis registered");
+    }
+
+    for (auto& [api, values] : _nominal) {
+        std::lock_guard<std::mutex> guard(detail::api_call_mutex(*api));
+        for (auto& [location, nominal] : values) {
+            detail::set_api_parameter(*api, location, nominal.first, nominal.second);
+        }
+    }
+
+    auto set_everywhere = [this](const std::string& location, double value) {
+        for (auto* api : _apis) {
+            std::lock_guard<std::mutex> guard(detail::api_call_mutex(*api));
+            auto& cache = _nominal[api];
+            if (cache.find(location) == cache.end()) {
+                cache[location] = api->get_parameter(location);
+            }
+            detail::set_api_parameter(*api, location, value, 0.0);
+        }
+    };
+    for (auto& [block_name, block] : _card_iter.cards[idx].blocks) {
+        for (auto [param_id, value] : block.params) {
+            set_everywhere(block_name + " " + std::to_string(param_id), value);
+        }
+        // multi-index entries: the UMAMI location is "BLOCK I J ..."
+        for (const auto& [param_ids, value] : block.multi_params) {
+            std::string location = block_name;
+            for (int param_id : param_ids) location += " " + std::to_string(param_id);
+            set_everywhere(location, value);
+        }
+    }
+    return true;
+}
+
+std::vector<REX::tea::iterator> ParamHandler::iterators() {
+    std::vector<REX::tea::iterator> result;
+    result.reserve(_card_iter.cards.size());
+    for (std::size_t i = 0; i < _card_iter.cards.size(); ++i) {
+        result.push_back([this, i]() { return this->apply_launch(i); });
+    }
+    return result;
+}
+
+std::vector<std::string> ParamHandler::launch_names() { return _card_iter.get_launch_names(); }
+
+std::vector<std::string> ParamHandler::weight_context() {
+    return _card_iter.get_rwgt_commands();
+}
+
+} // namespace madtrex
diff --git a/madtrex/src/pdfware.cpp b/madtrex/src/pdfware.cpp
new file mode 100644
index 0000000000..84b0a2734c
--- /dev/null
+++ b/madtrex/src/pdfware.cpp
@@ -0,0 +1,674 @@
+#include "madtrex/pdfware.hpp"
+
+#include <algorithm>
+#include <cmath>
+#include <format>
+#include <optional>
+#include <set>
+#include <stdexcept>
+
+namespace madtrex {
+
+PdfSet::PdfSet(
+    madspace::ContextPtr context,
+    const std::string& grid_file,
+    std::vector<int> pids,
+    const std::string& alpha_s_grid_file,
+    const std::string& prefix
+) :
+    _context(std::move(context)), _label(prefix), _pids(std::move(pids)) {
+    if (_pids.empty()) {
+        throw std::invalid_argument("PdfSet: pids must not be empty");
+    }
+    for (std::size_t i = 0; i < _pids.size(); ++i) {
+        _pid_index[_pids[i]] = static_cast<madspace::me_int_t>(i);
+    }
+
+    madspace::PdfGrid grid(grid_file);
+    grid.initialize_globals(_context, prefix);
+    _density.emplace(grid, _pids, /*dynamic_pid=*/true, prefix);
+    _density_runtime =
+        madspace::build_runtime(_density->function(), _context, /*concurrent=*/false);
+
+    if (!alpha_s_grid_file.empty()) {
+        madspace::AlphaSGrid alpha_s_grid(alpha_s_grid_file);
+        alpha_s_grid.initialize_globals(_context, prefix);
+        _alpha_s.emplace(alpha_s_grid, prefix);
+        _alpha_s_runtime =
+            madspace::build_runtime(_alpha_s->function(), _context, /*concurrent=*/false);
+    }
+}
+
+PdfSet::PdfSet(
+    madspace::ContextPtr context,
+    const madspace::PdfGrid& grid,
+    std::vector<int> pids,
+    const madspace::AlphaSGrid* alpha_s_grid,
+    const std::string& existing_globals_prefix,
+    const std::string& label
+) :
+    _context(std::move(context)), _label(label), _pids(std::move(pids)) {
+    if (_pids.empty()) {
+        throw std::invalid_argument("PdfSet: pids must not be empty");
+    }
+    for (std::size_t i = 0; i < _pids.size(); ++i) {
+        _pid_index[_pids[i]] = static_cast<madspace::me_int_t>(i);
+    }
+
+    // No initialize_globals() call here: existing_globals_prefix's globals
+    // are assumed to already be registered on _context by the caller.
+    _density.emplace(grid, _pids, /*dynamic_pid=*/true, existing_globals_prefix);
+    _density_runtime =
+        madspace::build_runtime(_density->function(), _context, /*concurrent=*/false);
+
+    if (alpha_s_grid) {
+        _alpha_s.emplace(*alpha_s_grid, existing_globals_prefix);
+        _alpha_s_runtime =
+            madspace::build_runtime(_alpha_s->function(), _context, /*concurrent=*/false);
+    }
+}
+
+std::vector<double> PdfSet::eval_pdf(
+    const std::vector<double>& x,
+    const std::vector<double>& q,
+    const std::vector<long int>& pdgs
+) const {
+    std::size_t count = x.size();
+    if (q.size() != count || pdgs.size() != count) {
+        throw std::runtime_error("PdfSet::eval_pdf: x/q/pdgs size mismatch");
+    }
+    if (count == 0) {
+        return {};
+    }
+
+    // Built host-side (on the CPU device) since the values below come from a
+    // plain C++ loop over host memory; moved onto the context's own device
+    // afterward if that isn't the CPU (matching how Tensor::cpu() moves the
+    // other way for reading results back below).
+    madspace::Tensor x_tensor(madspace::DataType::dt_float, madspace::Sizes{count});
+    madspace::Tensor q_tensor(madspace::DataType::dt_float, madspace::Sizes{count});
+    madspace::Tensor idx_tensor(madspace::DataType::dt_int, madspace::Sizes{count});
+    auto x_view = x_tensor.view<double, 1>();
+    auto q_view = q_tensor.view<double, 1>();
+    auto idx_view = idx_tensor.view<madspace::me_int_t, 1>();
+    for (std::size_t i = 0; i < count; ++i) {
+        x_view[i] = x[i];
+        q_view[i] = q[i];
+        auto found = _pid_index.find(pdgs[i]);
+        if (found == _pid_index.end()) {
+            throw std::runtime_error(std::format(
+                "PdfSet::eval_pdf: PDG code {} is not among the PIDs '{}' was built for",
+                pdgs[i],
+                _label
+            ));
+        }
+        idx_view[i] = found->second;
+    }
+
+    auto device = _context->device();
+    if (device != madspace::cpu_device()) {
+        x_tensor = x_tensor.copy(*device);
+        q_tensor = q_tensor.copy(*device);
+        idx_tensor = idx_tensor.copy(*device);
+    }
+
+    auto outputs = _density_runtime->run({x_tensor, q_tensor, idx_tensor});
+    madspace::Tensor pdf_cpu = outputs.at(0).cpu();
+    auto pdf_view = pdf_cpu.view<double, 1>();
+    std::vector<double> result(count);
+    for (std::size_t i = 0; i < count; ++i) {
+        result[i] = pdf_view[i];
+    }
+    return result;
+}
+
+std::vector<double> PdfSet::eval_alpha_s(const std::vector<double>& q) const {
+    if (!has_alpha_s()) {
+        throw std::runtime_error(
+            std::format("PdfSet::eval_alpha_s: '{}' was loaded without an alpha_s grid", _label)
+        );
+    }
+    std::size_t count = q.size();
+    if (count == 0) {
+        return {};
+    }
+    madspace::Tensor q_tensor(madspace::DataType::dt_float, madspace::Sizes{count});
+    auto q_view = q_tensor.view<double, 1>();
+    for (std::size_t i = 0; i < count; ++i) {
+        q_view[i] = q[i];
+    }
+    auto device = _context->device();
+    if (device != madspace::cpu_device()) {
+        q_tensor = q_tensor.copy(*device);
+    }
+    auto outputs = _alpha_s_runtime->run({q_tensor});
+    madspace::Tensor alpha_s_cpu = outputs.at(0).cpu();
+    auto as_view = alpha_s_cpu.view<double, 1>();
+    std::vector<double> result(count);
+    for (std::size_t i = 0; i < count; ++i) {
+        result[i] = as_view[i];
+    }
+    return result;
+}
+
+namespace {
+
+struct IncomingPartons {
+    double x1, x2;
+    long int pdg1, pdg2;
+};
+
+// Reads off the two incoming (status==-1) partons of every event in proc
+// and converts their momenta into Bjorken x via the standard along-beam-axis
+// relations for massless, collinear beams. Which of the two is "beam 1" (the
+// +z-moving hadron) is decided by the sign of pz.
+std::vector<IncomingPartons> extract_incoming(
+    REX::process& proc, double beam1_energy, double beam2_energy
+) {
+    std::vector<IncomingPartons> result;
+    result.reserve(proc.events.size());
+    for (auto& ev_ptr : proc.events) {
+        auto& ev = *ev_ptr;
+        std::optional<std::size_t> idx1, idx2;
+        for (std::size_t i = 0; i < ev.size(); ++i) {
+            if (ev.status_[i] == -1) {
+                if (!idx1) {
+                    idx1 = i;
+                } else if (!idx2) {
+                    idx2 = i;
+                }
+            }
+        }
+        if (!idx1 || !idx2) {
+            throw std::runtime_error(
+                "make_pdf_weightor: event does not have exactly two incoming "
+                "(status==-1) partons"
+            );
+        }
+        auto p1 = ev.momenta_[*idx1];
+        auto p2 = ev.momenta_[*idx2];
+        bool first_is_beam1 = p1.z() >= p2.z();
+        auto& beam1_p = first_is_beam1 ? p1 : p2;
+        auto& beam2_p = first_is_beam1 ? p2 : p1;
+        long int beam1_pdg = ev.pdg_[first_is_beam1 ? *idx1 : *idx2];
+        long int beam2_pdg = ev.pdg_[first_is_beam1 ? *idx2 : *idx1];
+        result.push_back(
+            {(beam1_p.t() + beam1_p.z()) / (2.0 * beam1_energy),
+             (beam2_p.t() - beam2_p.z()) / (2.0 * beam2_energy),
+             beam1_pdg,
+             beam2_pdg}
+        );
+    }
+    return result;
+}
+
+} // namespace
+
+std::shared_ptr<REX::tea::weightor> make_pdf_weightor(
+    const PdfHypothesis& hyp, double beam1_energy, double beam2_energy, int alpha_s_power
+) {
+    if (!hyp.beam1_pdf_set || !hyp.beam2_pdf_set) {
+        throw std::runtime_error("make_pdf_weightor: hypothesis has no PdfSet");
+    }
+    if (beam1_energy <= 0.0 || beam2_energy <= 0.0) {
+        throw std::runtime_error(
+            "make_pdf_weightor: beam energies must be positive; call "
+            "PdfDriver::set_beam_energies(), or supply them via the LHEMeta passed to "
+            "load_events()"
+        );
+    }
+
+    auto beam1_pdf_set = hyp.beam1_pdf_set;
+    auto beam2_pdf_set = hyp.beam2_pdf_set;
+    ScaleChoice muF_scale = hyp.muF_scale;
+    ScaleChoice muR_scale = hyp.muR_scale;
+
+    return std::make_shared<REX::tea::weightor>(
+        [beam1_pdf_set, beam2_pdf_set, muF_scale, muR_scale, alpha_s_power, beam1_energy,
+         beam2_energy](REX::process& proc) -> std::shared_ptr<std::vector<double>> {
+            auto incoming = extract_incoming(proc, beam1_energy, beam2_energy);
+            std::size_t count = incoming.size();
+            auto result = std::make_shared<std::vector<double>>(count, 0.0);
+            if (count == 0) {
+                return result;
+            }
+
+            auto& muF_vec = proc.get_muF();
+            std::vector<double> x1(count), x2(count), muF(count);
+            std::vector<long int> pdg1(count), pdg2(count);
+            for (std::size_t i = 0; i < count; ++i) {
+                x1[i] = incoming[i].x1;
+                x2[i] = incoming[i].x2;
+                pdg1[i] = incoming[i].pdg1;
+                pdg2[i] = incoming[i].pdg2;
+                muF[i] = muF_scale.resolve(muF_vec.at(i));
+            }
+            auto pdf1 = beam1_pdf_set->eval_pdf(x1, muF, pdg1);
+            auto pdf2 = beam2_pdf_set->eval_pdf(x2, muF, pdg2);
+
+            std::vector<double> alpha_s_vals;
+            if (alpha_s_power != 0) {
+                // Use beam1's alpha_s grid if available, otherwise beam2's.
+                const auto& as_set = beam1_pdf_set->has_alpha_s() ? beam1_pdf_set : beam2_pdf_set;
+                if (!as_set->has_alpha_s()) {
+                    throw std::runtime_error(
+                        "make_pdf_weightor: alpha_s_power != 0 but neither beam's PdfSet was "
+                        "loaded with an alpha_s grid"
+                    );
+                }
+                auto& muR_vec = proc.get_muR();
+                std::vector<double> muR(count);
+                for (std::size_t i = 0; i < count; ++i) {
+                    muR[i] = muR_scale.resolve(muR_vec.at(i));
+                }
+                alpha_s_vals = as_set->eval_alpha_s(muR);
+            }
+
+            for (std::size_t i = 0; i < count; ++i) {
+                double w = pdf1[i] * pdf2[i];
+                if (alpha_s_power != 0) {
+                    w *= std::pow(alpha_s_vals[i], alpha_s_power);
+                }
+                (*result)[i] = w;
+            }
+            return result;
+        }
+    );
+}
+
+PdfDriver::PdfDriver() : _context(madspace::default_context()) {}
+
+PdfDriver::PdfDriver(madspace::ContextPtr context) : _context(std::move(context)) {}
+
+PdfDriver& PdfDriver::load_pdf_set(
+    const std::string& name,
+    const std::string& grid_file,
+    const std::string& alpha_s_grid_file,
+    std::vector<int> pids
+) {
+    if (_pdf_sets.contains(name)) {
+        throw std::runtime_error(
+            std::format("PdfDriver::load_pdf_set: a PDF set named '{}' is already loaded", name)
+        );
+    }
+    _pdf_sets[name] =
+        std::make_shared<PdfSet>(_context, grid_file, std::move(pids), alpha_s_grid_file, name);
+    return *this;
+}
+
+PdfDriver& PdfDriver::add_pdf_set(
+    const std::string& name,
+    const madspace::PdfGrid& grid,
+    const madspace::AlphaSGrid* alpha_s_grid,
+    std::vector<int> pids,
+    const std::string& existing_globals_prefix
+) {
+    if (_pdf_sets.contains(name)) {
+        throw std::runtime_error(
+            std::format("PdfDriver::add_pdf_set: a PDF set named '{}' is already loaded", name)
+        );
+    }
+    _pdf_sets[name] = std::make_shared<PdfSet>(
+        _context, grid, std::move(pids), alpha_s_grid, existing_globals_prefix, name
+    );
+    return *this;
+}
+
+PdfDriver& PdfDriver::set_alpha_s_power(int power) {
+    _alpha_s_power = power;
+    return *this;
+}
+
+const std::shared_ptr<PdfSet>& PdfDriver::find_pdf_set(const std::string& name) const {
+    auto found = _pdf_sets.find(name);
+    if (found == _pdf_sets.end()) {
+        throw std::runtime_error(std::format(
+            "PdfDriver: no PDF set named '{}' has been loaded; call load_pdf_set() first", name
+        ));
+    }
+    return found->second;
+}
+
+PdfDriver& PdfDriver::set_baseline(
+    const std::string& pdf_set_name, ScaleChoice muF_scale, ScaleChoice muR_scale
+) {
+    auto ps = find_pdf_set(pdf_set_name);
+    _baseline = PdfHypothesis{"", ps, ps, muF_scale, muR_scale};
+    return *this;
+}
+
+PdfDriver& PdfDriver::set_baseline(
+    const std::string& beam1_pdf_set_name,
+    const std::string& beam2_pdf_set_name,
+    ScaleChoice muF_scale,
+    ScaleChoice muR_scale
+) {
+    _baseline = PdfHypothesis{
+        "", find_pdf_set(beam1_pdf_set_name), find_pdf_set(beam2_pdf_set_name),
+        muF_scale, muR_scale};
+    return *this;
+}
+
+PdfDriver& PdfDriver::add_hypothesis(
+    const std::string& launch_name,
+    const std::string& pdf_set_name,
+    ScaleChoice muF_scale,
+    ScaleChoice muR_scale
+) {
+    auto ps = find_pdf_set(pdf_set_name);
+    _hypotheses.push_back({launch_name, ps, ps, muF_scale, muR_scale});
+    return *this;
+}
+
+PdfDriver& PdfDriver::add_hypothesis(
+    const std::string& launch_name,
+    const std::string& beam1_pdf_set_name,
+    const std::string& beam2_pdf_set_name,
+    ScaleChoice muF_scale,
+    ScaleChoice muR_scale
+) {
+    _hypotheses.push_back(
+        {launch_name, find_pdf_set(beam1_pdf_set_name), find_pdf_set(beam2_pdf_set_name),
+         muF_scale, muR_scale}
+    );
+    return *this;
+}
+
+namespace {
+
+// The standard "seven-point" scale-variation set: every (muR_factor,
+// muF_factor) pair in {0.5, 1, 2}^2 except the two "uncorrelated" extremes
+// (0.5, 2) and (2, 0.5), whose factor-of-4 simultaneous muR/muF shift is
+// generally considered unphysically large to include.
+const std::vector<std::pair<double, double>>& seven_point_factors() {
+    static const std::vector<std::pair<double, double>> factors = {
+        {1.0, 1.0}, {2.0, 1.0}, {0.5, 1.0}, {1.0, 2.0},
+        {1.0, 0.5}, {2.0, 2.0}, {0.5, 0.5}
+    };
+    return factors;
+}
+
+// Formats a scale factor into a launch-name-safe token; every
+// seven_point_factors() entry is one of these three.
+std::string factor_token(double factor) {
+    if (factor == 0.5) return "0p5";
+    if (factor == 1.0) return "1";
+    if (factor == 2.0) return "2";
+    return std::format("{:g}", factor);
+}
+
+} // namespace
+
+PdfDriver& PdfDriver::add_scale_variation_envelope(
+    const std::string& group_name,
+    const std::string& pdf_set_name,
+    ScaleChoice muF_scale,
+    ScaleChoice muR_scale
+) {
+    std::vector<std::string> names;
+    names.reserve(seven_point_factors().size());
+    for (auto& [muR_factor, muF_factor] : seven_point_factors()) {
+        std::string name =
+            group_name + "_muR" + factor_token(muR_factor) + "_muF" + factor_token(muF_factor);
+        add_hypothesis(
+            name, pdf_set_name, muF_scale.scaled_by(muF_factor), muR_scale.scaled_by(muR_factor)
+        );
+        names.push_back(std::move(name));
+    }
+    _scale_variation_groups.push_back(std::move(names));
+    return *this;
+}
+
+PdfDriver& PdfDriver::add_scale_variation_envelope(
+    const std::string& group_name,
+    const std::string& beam1_pdf_set_name,
+    const std::string& beam2_pdf_set_name,
+    ScaleChoice muF_scale,
+    ScaleChoice muR_scale
+) {
+    std::vector<std::string> names;
+    names.reserve(seven_point_factors().size());
+    for (auto& [muR_factor, muF_factor] : seven_point_factors()) {
+        std::string name =
+            group_name + "_muR" + factor_token(muR_factor) + "_muF" + factor_token(muF_factor);
+        add_hypothesis(
+            name, beam1_pdf_set_name, beam2_pdf_set_name, muF_scale.scaled_by(muF_factor),
+            muR_scale.scaled_by(muR_factor)
+        );
+        names.push_back(std::move(name));
+    }
+    _scale_variation_groups.push_back(std::move(names));
+    return *this;
+}
+
+PdfDriver& PdfDriver::set_scale_variation_output(ScaleVariationOutput mode) {
+    _scale_variation_output = mode;
+    return *this;
+}
+
+ScaleVariationOutput PdfDriver::scale_variation_output() const {
+    return _scale_variation_output;
+}
+
+PdfDriver& PdfDriver::set_beam_energies(double beam1_energy, double beam2_energy) {
+    _beam1_energy = beam1_energy;
+    _beam2_energy = beam2_energy;
+    _beam_energies_set = true;
+    return *this;
+}
+
+PdfDriver& PdfDriver::load_events(const std::string& path, const madspace::LHEMeta& meta) {
+    if (!_baseline) {
+        throw std::runtime_error(
+            "PdfDriver::load_events: no baseline set; call set_baseline() first"
+        );
+    }
+    if (_hypotheses.empty()) {
+        throw std::runtime_error(
+            "PdfDriver::load_events: no hypotheses registered; call add_hypothesis() first"
+        );
+    }
+
+    LheFormat format = detect_lhe_format(path);
+    std::shared_ptr<REX::lhe> loaded =
+        format == LheFormat::binary ? load_lhe_binary(path, meta) : load_lhe_xml(path);
+
+    double beam1 = _beam1_energy, beam2 = _beam2_energy;
+    if (!_beam_energies_set) {
+        beam1 = loaded->eBm_[0];
+        beam2 = loaded->eBm_[1];
+    }
+    if (beam1 <= 0.0 || beam2 <= 0.0) {
+        throw std::runtime_error(
+            "PdfDriver::load_events: beam energies are not known; call "
+            "set_beam_energies() explicitly, or supply them via meta.beam1_energy/"
+            "beam2_energy for binary event files"
+        );
+    }
+
+    auto proc_rwgt = std::make_shared<REX::tea::procReweightor>();
+    proc_rwgt->set_event_checker(REX::event_bool_fn([](REX::event&) { return true; }));
+    proc_rwgt->set_normaliser(*make_pdf_weightor(*_baseline, beam1, beam2, _alpha_s_power));
+
+    std::vector<REX::tea::weightor> rwgts;
+    std::vector<std::string> names;
+    rwgts.reserve(_hypotheses.size());
+    names.reserve(_hypotheses.size());
+    for (auto& hyp : _hypotheses) {
+        rwgts.push_back(*make_pdf_weightor(hyp, beam1, beam2, _alpha_s_power));
+        names.push_back(hyp.launch_name);
+    }
+    proc_rwgt->set_reweight_functions(rwgts);
+
+    _reweightor = std::make_unique<REX::tea::reweightor>(
+        std::move(*loaded), std::vector<std::shared_ptr<REX::tea::procReweightor>>{proc_rwgt}
+    );
+    _reweightor->set_launch_names(names);
+    // Muted by default (see _verbose's member initializer); linking here
+    // shares the flag both ways, so verbose().set_verbose(true) unmutes the
+    // reweightor after the fact too.
+    _reweightor->verbose.link(_verbose);
+    _binary_output = (format == LheFormat::binary);
+    return *this;
+}
+
+PdfDriver& PdfDriver::set_binary_output(bool binary) {
+    _binary_output = binary;
+    return *this;
+}
+
+void PdfDriver::filter_scale_variation_weights() {
+    if (_scale_variation_groups.empty() || _scale_variation_output == ScaleVariationOutput::all) {
+        return;
+    }
+    if (!_reweightor) return;
+
+    auto& weight_ids = *_reweightor->weight_ids;
+    std::set<std::size_t> to_drop;
+
+    for (auto& group : _scale_variation_groups) {
+        std::vector<std::size_t> indices;
+        indices.reserve(group.size());
+        for (auto& name : group) {
+            auto it = std::find(weight_ids.begin(), weight_ids.end(), name);
+            if (it != weight_ids.end()) {
+                indices.push_back(static_cast<std::size_t>(std::distance(weight_ids.begin(), it)));
+            }
+        }
+        // Nothing (left) of this group in weight_ids -- either
+        // write_weights() has already dropped it (repeat call, harmless
+        // no-op) or run() hasn't populated weight_ids yet.
+        if (indices.empty()) continue;
+
+        std::vector<std::size_t> keep;
+        if (_scale_variation_output == ScaleVariationOutput::envelope) {
+            std::size_t min_idx = indices.front(), max_idx = indices.front();
+            for (auto idx : indices) {
+                if (idx >= _reweightor->rwgt_xSec.size()) continue;
+                if (_reweightor->rwgt_xSec[idx] < _reweightor->rwgt_xSec[min_idx]) min_idx = idx;
+                if (_reweightor->rwgt_xSec[idx] > _reweightor->rwgt_xSec[max_idx]) max_idx = idx;
+            }
+            keep.push_back(min_idx);
+            if (max_idx != min_idx) keep.push_back(max_idx);
+        }
+        // ScaleVariationOutput::none: keep stays empty, so every index of
+        // this group below gets dropped.
+
+        for (auto idx : indices) {
+            if (std::find(keep.begin(), keep.end(), idx) == keep.end()) {
+                to_drop.insert(idx);
+            }
+        }
+    }
+
+    if (to_drop.empty()) return;
+
+    // Highest index first, so erasing one never invalidates the positions
+    // still queued for removal.
+    std::vector<std::size_t> drop_sorted(to_drop.begin(), to_drop.end());
+    std::sort(drop_sorted.rbegin(), drop_sorted.rend());
+
+    for (auto idx : drop_sorted) {
+        if (idx < weight_ids.size()) weight_ids.erase(weight_ids.begin() + idx);
+    }
+    for (auto& ev : _reweightor->events) {
+        if (!ev) continue;
+        for (auto idx : drop_sorted) {
+            if (idx < ev->wgts_.size()) ev->wgts_.erase(ev->wgts_.begin() + idx);
+        }
+    }
+    for (auto idx : drop_sorted) {
+        if (idx < _reweightor->rwgt_xSec.size()) {
+            _reweightor->rwgt_xSec.erase(_reweightor->rwgt_xSec.begin() + idx);
+        }
+        if (idx < _reweightor->rwgt_xErr.size()) {
+            _reweightor->rwgt_xErr.erase(_reweightor->rwgt_xErr.begin() + idx);
+        }
+    }
+}
+
+PdfDriver& PdfDriver::keep_weights(const std::vector<std::string>& keep_names) {
+    if (!_reweightor) {
+        throw std::runtime_error(
+            "PdfDriver::keep_weights: reweightor has not been built yet; call load_events() "
+            "first"
+        );
+    }
+    auto& weight_ids = *_reweightor->weight_ids;
+    std::vector<std::size_t> drop_sorted;
+    for (std::size_t i = 0; i < weight_ids.size(); ++i) {
+        if (std::find(keep_names.begin(), keep_names.end(), weight_ids[i]) == keep_names.end()) {
+            drop_sorted.push_back(i);
+        }
+    }
+    if (drop_sorted.empty()) return *this;
+
+    // Highest index first, so erasing one never invalidates the positions
+    // still queued for removal (drop_sorted is already ascending from the
+    // loop above, so this just reverses it).
+    std::sort(drop_sorted.rbegin(), drop_sorted.rend());
+
+    for (auto idx : drop_sorted) {
+        weight_ids.erase(weight_ids.begin() + idx);
+    }
+    for (auto& ev : _reweightor->events) {
+        if (!ev) continue;
+        for (auto idx : drop_sorted) {
+            if (idx < ev->wgts_.size()) ev->wgts_.erase(ev->wgts_.begin() + idx);
+        }
+    }
+    for (auto idx : drop_sorted) {
+        if (idx < _reweightor->rwgt_xSec.size()) {
+            _reweightor->rwgt_xSec.erase(_reweightor->rwgt_xSec.begin() + idx);
+        }
+        if (idx < _reweightor->rwgt_xErr.size()) {
+            _reweightor->rwgt_xErr.erase(_reweightor->rwgt_xErr.begin() + idx);
+        }
+    }
+    return *this;
+}
+
+PdfDriver& PdfDriver::write_weights(const std::string& path) {
+    if (!_reweightor) {
+        throw std::runtime_error(
+            "PdfDriver::write_weights: reweightor has not been built yet; call "
+            "load_events() first"
+        );
+    }
+    filter_scale_variation_weights();
+    if (_binary_output) {
+        save_weights_binary(*_reweightor, path);
+    } else {
+        REX::write_lhef(*_reweightor, path, true);
+    }
+    return *this;
+}
+
+PdfDriver& PdfDriver::reweight(
+    const std::string& events_path, const std::string& output_path, const madspace::LHEMeta& meta
+) {
+    load_events(events_path, meta);
+    _reweightor->run();
+    write_weights(output_path);
+    return *this;
+}
+
+REX::tea::reweightor& PdfDriver::get() {
+    if (!_reweightor) {
+        throw std::runtime_error("PdfDriver::get: reweightor has not been built yet");
+    }
+    return *_reweightor;
+}
+
+const REX::tea::reweightor& PdfDriver::get() const {
+    if (!_reweightor) {
+        throw std::runtime_error("PdfDriver::get: reweightor has not been built yet");
+    }
+    return *_reweightor;
+}
+
+std::vector<int> PdfDriver::default_pids() {
+    return {21, 1, -1, 2, -2, 3, -3, 4, -4, 5, -5};
+}
+
+} // namespace madtrex
diff --git a/madtrex/src/python/madtrex.cpp b/madtrex/src/python/madtrex.cpp
new file mode 100644
index 0000000000..6aeb8e9c73
--- /dev/null
+++ b/madtrex/src/python/madtrex.cpp
@@ -0,0 +1,880 @@
+#include <pybind11/functional.h>
+#include <pybind11/pybind11.h>
+#include <pybind11/stl.h>
+
+#include "madtrex/madtrex.hpp"
+#include "madtrex/pdfware.hpp"
+
+namespace py = pybind11;
+using namespace madtrex;
+using madspace::LHEMeta;
+using madspace::MatrixElementApi;
+
+PYBIND11_MODULE(_madtrex_py, m) {
+    m.doc() = "Python bindings for madtrex::TupperWare/WareHouse, the generic "
+               "MatrixElementApi <-> REX::tea reweighting bridge";
+
+    // madtrex is a separate extension module layered on top of madspace: the
+    // MatrixElementApi/Context/LHEMeta types its bindings take and return are
+    // registered by madspace's own _madspace_py, so import it first (pybind11
+    // shares registered types across modules built against the same pybind11
+    // version, see CMakeLists.txt) rather than binding them again here.
+    py::module_::import("madspace");
+
+    // TupperWare/WareHouse build REX::tea::procReweightor/reweightor objects and
+    // consume REX::eventBelongs/event_bool_fn event checkers -- all types already
+    // bound by the separately-built `pyrex` package (pyrex / pyrex.tea; named
+    // "pyrex" rather than "rex" because "rex" is already an established,
+    // unrelated package on the index). Import them first so those types are
+    // registered in the process before any madtrex binding below references
+    // them, and do NOT re-register py::classh<...> for any of them here:
+    // pybind11 raises "generic_type: type ... is already registered!" the
+    // moment both `pyrex` and `madtrex` are imported into the same process
+    // and both sides independently declare a class binding for the same C++
+    // type. Cross-module reuse (verified empirically: an independently
+    // compiled procReweightor, returned from a second, unrelated extension
+    // module, casts cleanly to the already-registered pyrex.tea.ProcReweightor
+    // Python type) is what makes objects built here actually usable with
+    // pyrex.tea.Reweightor.run() etc.
+    py::module_::import("pyrex");
+    py::module_::import("pyrex.tea");
+
+    // ---- UmamiInputKey / UmamiOutputKey ----
+    // Plain C enums from madspace/umami.h; not bound anywhere else in
+    // madspace itself. Every value is exposed (not just the subset TupperWare can
+    // auto-derive from a REX::process) since the enum also indexes
+    // MatrixElementApi.supported_inputs()/required_inputs()/supported_outputs().
+    py::enum_<UmamiInputKey>(m, "UmamiInputKey")
+        .value("MOMENTA", UMAMI_IN_MOMENTA)
+        .value("ALPHA_S", UMAMI_IN_ALPHA_S)
+        .value("FLAVOR_INDEX", UMAMI_IN_FLAVOR_INDEX)
+        .value("RANDOM_COLOR", UMAMI_IN_RANDOM_COLOR)
+        .value("RANDOM_HELICITY", UMAMI_IN_RANDOM_HELICITY)
+        .value("RANDOM_DIAGRAM", UMAMI_IN_RANDOM_DIAGRAM)
+        .value("HELICITY_INDEX", UMAMI_IN_HELICITY_INDEX)
+        .value("DIAGRAM_INDEX", UMAMI_IN_DIAGRAM_INDEX)
+        .value("CHANNEL_INDEX", UMAMI_IN_CHANNEL_INDEX);
+
+    py::enum_<UmamiOutputKey>(m, "UmamiOutputKey")
+        .value("MATRIX_ELEMENT", UMAMI_OUT_MATRIX_ELEMENT)
+        .value("DIAGRAM_AMP2", UMAMI_OUT_DIAGRAM_AMP2)
+        .value("COLOR_INDEX", UMAMI_OUT_COLOR_INDEX)
+        .value("HELICITY_INDEX", UMAMI_OUT_HELICITY_INDEX)
+        .value("DIAGRAM_INDEX", UMAMI_OUT_DIAGRAM_INDEX)
+        .value("GPU_STREAM", UMAMI_OUT_GPU_STREAM);
+
+    // default_input_keys/default_output_keys are exposed as the documented
+    // starting point for a manual add_api/set_normaliser override.
+    // make_weightor() itself is intentionally not bound: it returns a raw
+    // shared_ptr<REX::tea::weightor>, which -- like the plain weightor type in
+    // tearex.cpp -- isn't something any bound API can consume back from Python
+    // (procReweightor's Python constructors only take Python-callable
+    // std::functions, not a pre-built weightor handle). TupperWare/WareHouse's
+    // add_api()/build() cover the actual construction path end to end.
+    m.def(
+        "default_input_keys",
+        &default_input_keys,
+        py::arg("api"),
+        py::arg("helicity_reweighting") = false
+    );
+    m.def("default_output_keys", &default_output_keys, py::arg("api"));
+
+    // ---- ReloadSpec ----
+    // Plain data: how TupperWare::add_api()/set_normaliser() can reload an
+    // independent MatrixElementApi instance per WareHouse.set_max_batch()
+    // overflow batch instead of every batch sharing the one instance it was
+    // registered with -- see WareHouse.set_max_batch() below. Driver builds
+    // and attaches one of these automatically for every api it loads, so
+    // Python code driving reweighting through Driver never needs to touch
+    // this directly; it's exposed for TupperWare/WareHouse used standalone.
+    py::classh<ReloadSpec>(m, "ReloadSpec")
+        .def(
+            py::init<madspace::ContextPtr, std::string, std::string>(),
+            py::arg("context"),
+            py::arg("me_path"),
+            py::arg("param_card")
+        )
+        .def_readwrite("context", &ReloadSpec::context)
+        .def_readwrite("me_path", &ReloadSpec::me_path)
+        .def_readwrite("param_card", &ReloadSpec::param_card);
+
+    // ---- TupperWare ----
+    // set_event_checker is bound only in its shared_ptr<eventBelongs> form, not
+    // the sibling by-value eventBelongs overload: a Python EventBelongs object
+    // converts to shared_ptr<...> transparently via classh either way, and
+    // pybind11 can't otherwise distinguish which of the two overloads a given
+    // EventBelongs argument was meant for (same rationale as
+    // procReweightor's selector argument in Rex's own tearex.cpp). The
+    // convenience constructor taking a plain eventBelongs has no shared_ptr
+    // sibling to conflict with, so it's bound as-is.
+    //
+    // Lifetimes: a TupperWare only keeps raw pointers to its MatrixElementApi
+    // objects, so every method taking an api (or a list of them) keeps that
+    // argument alive as long as the TupperWare (keep_alive<1, 2>; a Python
+    // MatrixElementApi in turn keeps its madspace Context, which owns the
+    // loaded library, alive). The builder methods return *this
+    // (reference_internal), so tw.add_api(a).set_normaliser(b) configures tw
+    // itself rather than a detached copy.
+    py::classh<TupperWare>(m, "TupperWare")
+        .def(py::init<>())
+        .def(
+            py::init<const MatrixElementApi&, REX::eventBelongs>(),
+            py::arg("api"),
+            py::arg("checker"),
+            py::keep_alive<1, 2>()
+        )
+        .def(
+            py::init<const MatrixElementApi&, REX::event_bool_fn>(),
+            py::arg("api"),
+            py::arg("checker"),
+            py::keep_alive<1, 2>()
+        )
+        .def(
+            "add_api",
+            py::overload_cast<const MatrixElementApi&>(&TupperWare::add_api),
+            py::arg("api"),
+            py::keep_alive<1, 2>(),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "add_api",
+            py::overload_cast<
+                const MatrixElementApi&,
+                std::vector<UmamiInputKey>,
+                std::vector<UmamiOutputKey>>(&TupperWare::add_api),
+            py::arg("api"),
+            py::arg("input_keys"),
+            py::arg("output_keys"),
+            py::keep_alive<1, 2>(),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "add_api",
+            py::overload_cast<const MatrixElementApi&, ReloadSpec>(&TupperWare::add_api),
+            py::arg("api"),
+            py::arg("reload"),
+            py::keep_alive<1, 2>(),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "add_api",
+            py::overload_cast<
+                const MatrixElementApi&,
+                std::vector<UmamiInputKey>,
+                std::vector<UmamiOutputKey>,
+                ReloadSpec>(&TupperWare::add_api),
+            py::arg("api"),
+            py::arg("input_keys"),
+            py::arg("output_keys"),
+            py::arg("reload"),
+            py::keep_alive<1, 2>(),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "add_api",
+            py::overload_cast<std::vector<const MatrixElementApi*>>(&TupperWare::add_api),
+            py::arg("apis"),
+            py::keep_alive<1, 2>(),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "add_api",
+            py::overload_cast<
+                std::vector<const MatrixElementApi*>,
+                std::vector<UmamiInputKey>,
+                std::vector<UmamiOutputKey>>(&TupperWare::add_api),
+            py::arg("apis"),
+            py::arg("input_keys"),
+            py::arg("output_keys"),
+            py::keep_alive<1, 2>(),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "add_api",
+            py::overload_cast<
+                std::vector<const MatrixElementApi*>,
+                std::vector<std::vector<UmamiInputKey>>,
+                std::vector<std::vector<UmamiOutputKey>>>(&TupperWare::add_api),
+            py::arg("apis"),
+            py::arg("input_keys_list"),
+            py::arg("output_keys_list"),
+            py::keep_alive<1, 2>(),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "set_api",
+            py::overload_cast<std::vector<const MatrixElementApi*>>(&TupperWare::set_api),
+            py::arg("apis"),
+            py::keep_alive<1, 2>(),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "set_api",
+            py::overload_cast<
+                std::vector<const MatrixElementApi*>,
+                std::vector<UmamiInputKey>,
+                std::vector<UmamiOutputKey>>(&TupperWare::set_api),
+            py::arg("apis"),
+            py::arg("input_keys"),
+            py::arg("output_keys"),
+            py::keep_alive<1, 2>(),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "set_api",
+            py::overload_cast<
+                std::vector<const MatrixElementApi*>,
+                std::vector<std::vector<UmamiInputKey>>,
+                std::vector<std::vector<UmamiOutputKey>>>(&TupperWare::set_api),
+            py::arg("apis"),
+            py::arg("input_keys_list"),
+            py::arg("output_keys_list"),
+            py::keep_alive<1, 2>(),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "set_normaliser",
+            py::overload_cast<const MatrixElementApi&>(&TupperWare::set_normaliser),
+            py::arg("api"),
+            py::keep_alive<1, 2>(),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "set_normaliser",
+            py::overload_cast<
+                const MatrixElementApi&,
+                std::vector<UmamiInputKey>,
+                std::vector<UmamiOutputKey>>(&TupperWare::set_normaliser),
+            py::arg("api"),
+            py::arg("input_keys"),
+            py::arg("output_keys"),
+            py::keep_alive<1, 2>(),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "set_normaliser",
+            py::overload_cast<const MatrixElementApi&, ReloadSpec>(&TupperWare::set_normaliser),
+            py::arg("api"),
+            py::arg("reload"),
+            py::keep_alive<1, 2>(),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "set_normaliser",
+            py::overload_cast<
+                const MatrixElementApi&,
+                std::vector<UmamiInputKey>,
+                std::vector<UmamiOutputKey>,
+                ReloadSpec>(&TupperWare::set_normaliser),
+            py::arg("api"),
+            py::arg("input_keys"),
+            py::arg("output_keys"),
+            py::arg("reload"),
+            py::keep_alive<1, 2>(),
+            py::return_value_policy::reference_internal
+        )
+        .def("set_result_keys", &TupperWare::set_result_keys, py::arg("keys"), py::return_value_policy::reference_internal)
+        .def(
+            "set_helicity_reweighting",
+            &TupperWare::set_helicity_reweighting,
+            py::arg("helicity_reweighting"),
+            py::return_value_policy::reference_internal
+        )
+        .def("helicity_reweighting", &TupperWare::helicity_reweighting)
+        .def(
+            "set_stored_normalisation",
+            &TupperWare::set_stored_normalisation,
+            py::arg("key"),
+            py::return_value_policy::reference_internal
+        )
+        .def("stored_normalisation", &TupperWare::stored_normalisation)
+        .def("set_use_helicity", &TupperWare::set_use_helicity, py::arg("use_helicity"), py::return_value_policy::reference_internal)
+        .def(
+            "set_event_checker",
+            py::overload_cast<std::shared_ptr<REX::eventBelongs>>(
+                &TupperWare::set_event_checker
+            ),
+            py::arg("checker"),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "set_event_checker",
+            py::overload_cast<REX::event_bool_fn>(&TupperWare::set_event_checker),
+            py::arg("checker"),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "build",
+            &TupperWare::build,
+            py::arg("n_iterations") = 1,
+            py::arg("reload_per_batch") = false,
+            py::arg("max_deferred_iterations") = REX::npos
+        );
+
+    // ---- WareHouse ----
+    // Move-only (copy ctor/assignment deleted in C++, mirroring
+    // REX::tea::reweightor's own single-owner semantics); py::classh's
+    // smart_holder handles that without needing a copy constructor. The
+    // lhe&&-taking build() overload is skipped, same rationale as Reweightor's
+    // own constructor in tearex.cpp: Python has no rvalue distinction, and the
+    // const lhe& overload covers every practical call (it just wraps its
+    // argument in a local REX::lhe copy before delegating).
+    py::classh<WareHouse>(m, "WareHouse")
+        .def(py::init<>())
+        .def(
+            "add_process",
+            py::overload_cast<TupperWare>(&WareHouse::add_process),
+            py::arg("process"),
+            py::keep_alive<1, 2>(),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "add_process",
+            py::overload_cast<std::vector<TupperWare>>(&WareHouse::add_process),
+            py::arg("processes"),
+            py::keep_alive<1, 2>(),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "add_process",
+            py::overload_cast<>(&WareHouse::add_process),
+            py::return_value_policy::reference_internal
+        )
+        .def("set_processes", &WareHouse::set_processes, py::arg("processes"), py::keep_alive<1, 2>(), py::return_value_policy::reference_internal)
+        .def(
+            "process",
+            py::overload_cast<std::size_t>(&WareHouse::process),
+            py::arg("index"),
+            py::return_value_policy::reference_internal
+        )
+        .def("processes", py::overload_cast<>(&WareHouse::processes))
+        .def("set_iterators", &WareHouse::set_iterators, py::arg("iterators"), py::return_value_policy::reference_internal)
+        .def("add_iterator", &WareHouse::add_iterator, py::arg("iterator"), py::return_value_policy::reference_internal)
+        .def("set_initialise", &WareHouse::set_initialise, py::arg("init"), py::return_value_policy::reference_internal)
+        .def("set_finalise", &WareHouse::set_finalise, py::arg("fin"), py::return_value_policy::reference_internal)
+        .def("set_launch_names", &WareHouse::set_launch_names, py::arg("names"), py::return_value_policy::reference_internal)
+        .def("add_launch_name", &WareHouse::add_launch_name, py::arg("name"), py::return_value_policy::reference_internal)
+        .def("launch_names", &WareHouse::launch_names)
+        .def("set_max_batch", &WareHouse::set_max_batch, py::arg("max_batch"), py::return_value_policy::reference_internal)
+        .def("max_batch", &WareHouse::max_batch)
+        .def("set_threads", &WareHouse::set_threads, py::arg("threads"), py::return_value_policy::reference_internal)
+        .def("threads", &WareHouse::threads)
+        .def("set_min_batch", &WareHouse::set_min_batch, py::arg("min_batch"), py::return_value_policy::reference_internal)
+        .def("min_batch", &WareHouse::min_batch)
+        .def("batch_size_for", &WareHouse::batch_size_for, py::arg("n_events"))
+        .def(
+            "set_max_deferred_iterations",
+            &WareHouse::set_max_deferred_iterations,
+            py::arg("max_deferred_iterations"),
+            py::return_value_policy::reference_internal
+        )
+        .def("max_deferred_iterations", &WareHouse::max_deferred_iterations)
+        .def("set_verbose", &WareHouse::set_verbose, py::arg("v"), py::return_value_policy::reference_internal)
+        .def("is_verbose", &WareHouse::is_verbose)
+        .def(
+            "build",
+            py::overload_cast<const REX::lhe&>(&WareHouse::build),
+            py::arg("mother"),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "get",
+            py::overload_cast<>(&WareHouse::get),
+            py::return_value_policy::reference_internal
+        )
+        .def("built", &WareHouse::built);
+
+    // ---- LHE loading/writing, LheFormat ----
+    // detect_lhe_format()/load_lhe() sniff a path's leading bytes (NumPy magic
+    // vs LHEF XML) rather than trusting the extension; see madtrex.hpp/
+    // binary_io.hpp. The std::istream-taking load_lhe_xml() and
+    // madspace::EventFile-taking load_lhe_binary() overloads aren't bound --
+    // Python callers have a path string, not an already-open C++ stream/file
+    // handle.
+    py::enum_<LheFormat>(m, "LheFormat")
+        .value("XML", LheFormat::xml)
+        .value("BINARY", LheFormat::binary);
+
+    m.def("detect_lhe_format", &detect_lhe_format, py::arg("path"));
+    m.def(
+        "load_lhe_xml",
+        py::overload_cast<const std::string&>(&load_lhe_xml),
+        py::arg("path")
+    );
+    m.def("load_lhe", &load_lhe, py::arg("path"));
+    m.def(
+        "load_lhe_binary",
+        py::overload_cast<const std::string&, const LHEMeta&>(&load_lhe_binary),
+        py::arg("path"),
+        py::arg("meta") = LHEMeta{}
+    );
+    m.def(
+        "save_weights_binary", &save_weights_binary, py::arg("lhe"), py::arg("path_prefix")
+    );
+
+    // ---- Splitting a job across independent shards, and recombining them ----
+    // The generic REX::split_source() itself is intentionally not bound
+    // (would need REX::eventSource/eventSink exposed to Python too, which
+    // pyrex doesn't currently do); split_events()/combine_shards() are the
+    // two madtrex-level entry points that actually matter for driving this
+    // workflow from Python -- see madtrex.hpp for the full picture.
+    py::classh<SplitResult>(m, "SplitResult")
+        .def_readonly("paths", &SplitResult::paths)
+        .def_readonly("counts", &SplitResult::counts);
+
+    m.def(
+        "split_events",
+        &split_events,
+        py::arg("path"),
+        py::arg("output_prefix"),
+        py::arg("n_shards"),
+        py::arg("particle_count") = 0,
+        py::arg("batch_size") = 10000,
+        py::arg("meta") = LHEMeta{}
+    );
+    m.def(
+        "combine_shards",
+        &combine_shards,
+        py::arg("shard_event_paths"),
+        py::arg("shard_weight_prefixes"),
+        py::arg("launch_names"),
+        py::arg("output_path"),
+        py::arg("binary_output") = false,
+        py::arg("meta") = LHEMeta{}
+    );
+
+    // ---- SubProcessSpec / FlavorChannel ----
+    // Read-only introspection views: both are produced by
+    // Driver.load_process()/load_process_directory() parsing a
+    // subprocesses.json (see madtrex.hpp), not meant to be built from
+    // scratch in Python.
+    py::classh<FlavorChannel>(m, "FlavorChannel")
+        .def_readonly("index", &FlavorChannel::index)
+        .def_readonly("pdgs", &FlavorChannel::pdgs)
+        .def_readonly("mirrored", &FlavorChannel::mirrored);
+
+    py::classh<SubProcessSpec>(m, "SubProcessSpec")
+        .def_readonly("incoming", &SubProcessSpec::incoming)
+        .def_readonly("outgoing", &SubProcessSpec::outgoing)
+        .def_readonly("me_path_template", &SubProcessSpec::me_path_template)
+        .def_readonly("source_path", &SubProcessSpec::source_path)
+        .def_readonly("diagram_count", &SubProcessSpec::diagram_count)
+        .def_readonly("flavors", &SubProcessSpec::flavors)
+        .def_readonly("helicities", &SubProcessSpec::helicities);
+
+    // ---- ParamHandler ----
+    // Builder-style methods are bound with reference_internal (as
+    // TupperWare's and WareHouse's above): ParamHandler is meant to be
+    // configured incrementally and its identity matters once its iterators()
+    // are handed to a WareHouse, so returning *this by reference rather than
+    // a detached copy keeps chaining behave like the C++ side.
+    py::classh<ParamHandler>(m, "ParamHandler")
+        .def(py::init<>())
+        .def(
+            "add_api",
+            &ParamHandler::add_api,
+            py::arg("api"),
+            py::keep_alive<1, 2>(),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "add_apis",
+            py::overload_cast<const std::vector<const MatrixElementApi*>&>(
+                &ParamHandler::add_apis
+            ),
+            py::arg("apis"),
+            py::keep_alive<1, 2>(),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "add_apis",
+            py::overload_cast<const WareHouse&>(&ParamHandler::add_apis),
+            py::arg("house"),
+            py::keep_alive<1, 2>(),
+            py::return_value_policy::reference_internal
+        )
+        .def("apis", &ParamHandler::apis)
+        .def(
+            "read_rwgt_card",
+            py::overload_cast<const std::string&>(&ParamHandler::read_rwgt_card),
+            py::arg("rwgt_path"),
+            py::return_value_policy::reference_internal
+        )
+        .def("launch_count", &ParamHandler::launch_count)
+        .def("apply_launch", &ParamHandler::apply_launch, py::arg("idx"))
+        .def("iterators", &ParamHandler::iterators)
+        .def("launch_names", &ParamHandler::launch_names)
+        .def("weight_context", &ParamHandler::weight_context);
+
+    // ---- Driver ----
+    // Driver is not copy-constructible (it owns a WareHouse, which isn't
+    // copyable either), so every method below that returns Driver& must use
+    // reference_internal explicitly -- the default policy would otherwise
+    // try to copy-construct the return value and fail to compile.
+    py::classh<Driver>(m, "Driver")
+        .def(py::init<>())
+        .def(py::init<madspace::ContextPtr>(), py::arg("context"))
+        .def(
+            "set_device_priority",
+            &Driver::set_device_priority,
+            py::arg("priority"),
+            py::return_value_policy::reference_internal
+        )
+        .def("device_priority", &Driver::device_priority)
+        .def(
+            "set_fptype_priority",
+            &Driver::set_fptype_priority,
+            py::arg("priority"),
+            py::return_value_policy::reference_internal
+        )
+        .def("fptype_priority", &Driver::fptype_priority)
+        .def(
+            "set_normaliser_fptype_priority",
+            &Driver::set_normaliser_fptype_priority,
+            py::arg("priority"),
+            py::return_value_policy::reference_internal
+        )
+        .def("normaliser_fptype_priority", &Driver::normaliser_fptype_priority)
+        .def(
+            "set_helicity_reweighting",
+            &Driver::set_helicity_reweighting,
+            py::arg("helicity_reweighting"),
+            py::return_value_policy::reference_internal
+        )
+        .def("helicity_reweighting", &Driver::helicity_reweighting)
+        .def(
+            "set_normaliser_process_directory",
+            &Driver::set_normaliser_process_directory,
+            py::arg("process_directory"),
+            py::arg("param_card") = "",
+            py::return_value_policy::reference_internal
+        )
+        .def("normaliser_process_directory", &Driver::normaliser_process_directory)
+        .def(
+            "set_normaliser_param_card",
+            &Driver::set_normaliser_param_card,
+            py::arg("param_card"),
+            py::return_value_policy::reference_internal
+        )
+        .def("normaliser_param_card", &Driver::normaliser_param_card)
+        .def(
+            "set_use_stored_normalisation",
+            &Driver::set_use_stored_normalisation,
+            py::arg("use_stored_normalisation"),
+            py::return_value_policy::reference_internal
+        )
+        .def("use_stored_normalisation", &Driver::use_stored_normalisation)
+        .def(
+            "set_use_helicity",
+            &Driver::set_use_helicity,
+            py::arg("use_helicity"),
+            py::return_value_policy::reference_internal
+        )
+        .def("use_helicity", &Driver::use_helicity)
+        .def(
+            "set_verbose",
+            &Driver::set_verbose,
+            py::arg("v"),
+            py::return_value_policy::reference_internal
+        )
+        .def("is_verbose", &Driver::is_verbose)
+        .def(
+            "set_max_batch",
+            &Driver::set_max_batch,
+            py::arg("max_batch"),
+            py::return_value_policy::reference_internal
+        )
+        .def("max_batch", &Driver::max_batch)
+        .def(
+            "set_threads",
+            &Driver::set_threads,
+            py::arg("threads"),
+            py::return_value_policy::reference_internal
+        )
+        .def("threads", &Driver::threads)
+        .def(
+            "set_min_batch",
+            &Driver::set_min_batch,
+            py::arg("min_batch"),
+            py::return_value_policy::reference_internal
+        )
+        .def("min_batch", &Driver::min_batch)
+        .def(
+            "set_max_deferred_iterations",
+            &Driver::set_max_deferred_iterations,
+            py::arg("max_deferred_iterations"),
+            py::return_value_policy::reference_internal
+        )
+        .def("max_deferred_iterations", &Driver::max_deferred_iterations)
+        .def(
+            "set_streaming",
+            &Driver::set_streaming,
+            py::arg("enable"),
+            py::arg("batch_size") = 10000,
+            py::return_value_policy::reference_internal
+        )
+        .def("streaming", &Driver::streaming)
+        .def("stream_batch_size", &Driver::stream_batch_size)
+        .def(
+            "load_process",
+            &Driver::load_process,
+            py::arg("subprocesses_json_path"),
+            py::arg("param_card") = "",
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "load_process_directory",
+            &Driver::load_process_directory,
+            py::arg("process_directory"),
+            py::arg("param_card") = "",
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "load_param_reweighting",
+            &Driver::load_param_reweighting,
+            py::arg("rwgt_path"),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "param_handler",
+            py::overload_cast<>(&Driver::param_handler),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "load_events",
+            &Driver::load_events,
+            py::arg("path"),
+            py::arg("meta") = LHEMeta{},
+            py::return_value_policy::reference_internal
+        )
+        .def("binary_output", &Driver::binary_output)
+        .def(
+            "set_binary_output",
+            &Driver::set_binary_output,
+            py::arg("binary"),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "write_weights",
+            &Driver::write_weights,
+            py::arg("path"),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "run_streaming",
+            &Driver::run_streaming,
+            py::arg("output_path"),
+            py::arg("include_ids") = false,
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "reweight",
+            &Driver::reweight,
+            py::arg("process_directory"),
+            py::arg("events_path"),
+            py::arg("output_path"),
+            py::arg("rwgt_path") = "",
+            py::arg("param_card") = "",
+            py::arg("meta") = LHEMeta{},
+            py::return_value_policy::reference_internal
+        )
+        .def("context", &Driver::context)
+        .def(
+            "warehouse",
+            py::overload_cast<>(&Driver::warehouse),
+            py::return_value_policy::reference_internal
+        )
+        .def("specs", &Driver::specs);
+
+    // ---- ScaleChoice ----
+    // Either an absolute scale (GeV) or a multiplicative factor of an
+    // event's own recorded muF/muR -- see its own doc comment in
+    // pdfware.hpp. A plain Python float passed where a ScaleChoice is
+    // expected (set_baseline()/add_hypothesis()/add_scale_variation_envelope()
+    // below) is implicitly treated as ScaleChoice.factor(...), matching the
+    // C++ side's own non-explicit single-double constructor -- the
+    // pre-existing, default behaviour.
+    py::classh<ScaleChoice> scale_choice(m, "ScaleChoice");
+    scale_choice
+        .def(py::init<>())
+        .def(py::init<double>(), py::arg("value"))
+        .def(py::init<double, bool>(), py::arg("value"), py::arg("absolute"))
+        .def_readwrite("value", &ScaleChoice::value)
+        .def_readwrite("absolute", &ScaleChoice::absolute)
+        .def_static("factor", &ScaleChoice::factor, py::arg("value"))
+        .def_static("fixed", &ScaleChoice::fixed, py::arg("value"))
+        .def("scaled_by", &ScaleChoice::scaled_by, py::arg("extra_factor"))
+        .def("resolve", &ScaleChoice::resolve, py::arg("nominal"));
+    py::implicitly_convertible<double, ScaleChoice>();
+    py::implicitly_convertible<int, ScaleChoice>();
+
+    // ---- ScaleVariationOutput ----
+    py::enum_<ScaleVariationOutput>(m, "ScaleVariationOutput")
+        .value("NONE", ScaleVariationOutput::none)
+        .value("ENVELOPE", ScaleVariationOutput::envelope)
+        .value("ALL", ScaleVariationOutput::all);
+
+    // ---- PdfDriver ----
+    // PdfDriver builds/runs a single REX::tea::reweightor doing PDF-only (not
+    // matrix-element) reweighting -- see pdfware.hpp. PdfSet/PdfHypothesis/
+    // make_pdf_weightor are internal building blocks (same "not bound" reasoning
+    // as make_weightor above -- load_pdf_set()/set_baseline()/add_hypothesis()
+    // cover construction end to end from Python) and aren't exposed here.
+    // PdfDriver isn't copy-constructible either (it owns a
+    // unique_ptr<REX::tea::reweightor>), so every PdfDriver&-returning method
+    // uses reference_internal, same as Driver above.
+    py::classh<PdfDriver>(m, "PdfDriver")
+        .def(py::init<>())
+        .def(py::init<madspace::ContextPtr>(), py::arg("context"))
+        .def(
+            "load_pdf_set",
+            &PdfDriver::load_pdf_set,
+            py::arg("name"),
+            py::arg("grid_file"),
+            py::arg("alpha_s_grid_file") = "",
+            py::arg("pids") = PdfDriver::default_pids(),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "add_pdf_set",
+            &PdfDriver::add_pdf_set,
+            py::arg("name"),
+            py::arg("grid"),
+            py::arg("alpha_s_grid") = nullptr,
+            py::arg("pids") = PdfDriver::default_pids(),
+            py::arg("existing_globals_prefix") = "",
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "set_alpha_s_power",
+            &PdfDriver::set_alpha_s_power,
+            py::arg("power"),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "set_baseline",
+            py::overload_cast<const std::string&, ScaleChoice, ScaleChoice>(
+                &PdfDriver::set_baseline),
+            py::arg("pdf_set_name"),
+            py::arg("muF_scale") = ScaleChoice::factor(1.0),
+            py::arg("muR_scale") = ScaleChoice::factor(1.0),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "set_baseline",
+            py::overload_cast<const std::string&, const std::string&, ScaleChoice, ScaleChoice>(
+                &PdfDriver::set_baseline),
+            py::arg("beam1_pdf_set_name"),
+            py::arg("beam2_pdf_set_name"),
+            py::arg("muF_scale") = ScaleChoice::factor(1.0),
+            py::arg("muR_scale") = ScaleChoice::factor(1.0),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "add_hypothesis",
+            py::overload_cast<const std::string&, const std::string&, ScaleChoice, ScaleChoice>(
+                &PdfDriver::add_hypothesis),
+            py::arg("launch_name"),
+            py::arg("pdf_set_name"),
+            py::arg("muF_scale") = ScaleChoice::factor(1.0),
+            py::arg("muR_scale") = ScaleChoice::factor(1.0),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "add_hypothesis",
+            py::overload_cast<
+                const std::string&, const std::string&, const std::string&, ScaleChoice,
+                ScaleChoice>(&PdfDriver::add_hypothesis),
+            py::arg("launch_name"),
+            py::arg("beam1_pdf_set_name"),
+            py::arg("beam2_pdf_set_name"),
+            py::arg("muF_scale") = ScaleChoice::factor(1.0),
+            py::arg("muR_scale") = ScaleChoice::factor(1.0),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "add_scale_variation_envelope",
+            py::overload_cast<const std::string&, const std::string&, ScaleChoice, ScaleChoice>(
+                &PdfDriver::add_scale_variation_envelope),
+            py::arg("group_name"),
+            py::arg("pdf_set_name"),
+            py::arg("muF_scale") = ScaleChoice::factor(1.0),
+            py::arg("muR_scale") = ScaleChoice::factor(1.0),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "add_scale_variation_envelope",
+            py::overload_cast<
+                const std::string&, const std::string&, const std::string&, ScaleChoice,
+                ScaleChoice>(&PdfDriver::add_scale_variation_envelope),
+            py::arg("group_name"),
+            py::arg("beam1_pdf_set_name"),
+            py::arg("beam2_pdf_set_name"),
+            py::arg("muF_scale") = ScaleChoice::factor(1.0),
+            py::arg("muR_scale") = ScaleChoice::factor(1.0),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "set_scale_variation_output",
+            &PdfDriver::set_scale_variation_output,
+            py::arg("mode"),
+            py::return_value_policy::reference_internal
+        )
+        .def("scale_variation_output", &PdfDriver::scale_variation_output)
+        .def(
+            "set_verbose",
+            &PdfDriver::set_verbose,
+            py::arg("v"),
+            py::return_value_policy::reference_internal
+        )
+        .def("is_verbose", &PdfDriver::is_verbose)
+        .def(
+            "set_beam_energies",
+            &PdfDriver::set_beam_energies,
+            py::arg("beam1_energy"),
+            py::arg("beam2_energy"),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "load_events",
+            &PdfDriver::load_events,
+            py::arg("path"),
+            py::arg("meta") = LHEMeta{},
+            py::return_value_policy::reference_internal
+        )
+        .def("binary_output", &PdfDriver::binary_output)
+        .def(
+            "set_binary_output",
+            &PdfDriver::set_binary_output,
+            py::arg("binary"),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "write_weights",
+            &PdfDriver::write_weights,
+            py::arg("path"),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "keep_weights",
+            &PdfDriver::keep_weights,
+            py::arg("keep_names"),
+            py::return_value_policy::reference_internal
+        )
+        .def(
+            "reweight",
+            &PdfDriver::reweight,
+            py::arg("events_path"),
+            py::arg("output_path"),
+            py::arg("meta") = LHEMeta{},
+            py::return_value_policy::reference_internal
+        )
+        .def("context", &PdfDriver::context)
+        .def(
+            "get",
+            py::overload_cast<>(&PdfDriver::get),
+            py::return_value_policy::reference_internal
+        )
+        .def("built", &PdfDriver::built)
+        .def_static("default_pids", &PdfDriver::default_pids);
+}
diff --git a/madtrex/src/tupperware.cpp b/madtrex/src/tupperware.cpp
new file mode 100644
index 0000000000..a059f33654
--- /dev/null
+++ b/madtrex/src/tupperware.cpp
@@ -0,0 +1,1599 @@
+#include "madtrex/tupperware.hpp"
+
+#include <algorithm>
+#include <any>
+#include <format>
+#include <iostream>
+#include <map>
+#include <mutex>
+#include <shared_mutex>
+#include <optional>
+#include <set>
+#include <stdexcept>
+#include <thread>
+#include <unordered_map>
+
+namespace madtrex {
+
+namespace detail {
+
+std::mutex& api_call_mutex(const madspace::MatrixElementApi& api) {
+    static std::mutex registry_mutex;
+    static std::unordered_map<const madspace::MatrixElementApi*, std::unique_ptr<std::mutex>>
+        registry;
+    std::lock_guard<std::mutex> lock(registry_mutex);
+    auto& slot = registry[&api];
+    if (!slot) {
+        slot = std::make_unique<std::mutex>();
+    }
+    return *slot;
+}
+
+void warn_once(const std::string& key, const std::string& message) {
+    static std::mutex mutex;
+    static std::set<std::string> seen;
+    std::lock_guard<std::mutex> lock(mutex);
+    if (seen.insert(key).second) {
+        std::cerr << "MadtRex warning: " << message << std::endl;
+    }
+}
+
+namespace {
+
+// Keyed by MatrixElementApi::file_name(), the library: the values set on it
+// (see parameters_lock() in tupperware.hpp for why they are kept).
+struct LibraryParameters {
+    std::map<std::string, std::pair<double, double>> values;
+};
+
+std::mutex parameters_mutex;
+std::unordered_map<std::string, LibraryParameters> parameters_registry;
+
+LibraryParameters library_parameters(const std::string& library) {
+    std::lock_guard<std::mutex> lock(parameters_mutex);
+    auto found = parameters_registry.find(library);
+    return found == parameters_registry.end() ? LibraryParameters{} : found->second;
+}
+
+void apply_parameters(const madspace::MatrixElementApi& target, const LibraryParameters& parameters) {
+    for (auto& [location, value] : parameters.values) {
+        target.set_parameter(location, value.first, value.second);
+    }
+}
+
+} // namespace
+
+std::shared_mutex& parameters_lock() {
+    static std::shared_mutex mutex;
+    return mutex;
+}
+
+void set_api_parameter(
+    const madspace::MatrixElementApi& api, const std::string& location, double real, double imag
+) {
+    api.set_parameter(location, real, imag);
+    std::lock_guard<std::mutex> lock(parameters_mutex);
+    auto& entry = parameters_registry[api.file_name()];
+    entry.values[location] = {real, imag};
+}
+
+void forget_api_parameters(const madspace::MatrixElementApi& api) {
+    std::lock_guard<std::mutex> lock(parameters_mutex);
+    parameters_registry.erase(api.file_name());
+}
+
+const madspace::MatrixElementApi&
+reload_matrix_element(const ReloadSpec& reload, REX::process& proc) {
+    std::string key = std::format(
+        "madtrex::reload::{}::{}::{}", static_cast<const void*>(reload.context.get()),
+        reload.me_path, reload.param_card
+    );
+    if (proc.has_device_data(key)) {
+        return *std::any_cast<const madspace::MatrixElementApi*>(proc.get_device_data(key).tag);
+    }
+    const madspace::MatrixElementApi* api;
+    {
+        // no evaluation runs while the parameters are reset to the
+        // param_card and set back (and a madspace::Context is not safe to
+        // load into concurrently)
+        std::unique_lock<std::shared_mutex> exclusive(parameters_lock());
+        api = &reload.context->load_matrix_element(reload.me_path, reload.param_card);
+        apply_parameters(*api, library_parameters(api->file_name()));
+    }
+    proc.set_device_data(key, nullptr, 0, std::any(api));
+    return *api;
+}
+
+} // namespace detail
+
+namespace {
+
+using madspace::MatrixElementApi;
+
+// umami_supported_inputs/umami_required_inputs/umami_supported_outputs are optional
+// symbols: MatrixElementApi falls back to a stub that reports
+// UMAMI_ERROR_NOT_IMPLEMENTED, which the query methods turn into a thrown
+// runtime_error. Treat "not implemented" as "unknown" rather than propagating.
+std::optional<std::vector<bool>> try_supported_inputs(const MatrixElementApi& api) {
+    try {
+        return api.supported_inputs();
+    } catch (const std::exception&) {
+        return std::nullopt;
+    }
+}
+
+std::optional<std::vector<bool>> try_required_inputs(const MatrixElementApi& api) {
+    try {
+        return api.required_inputs();
+    } catch (const std::exception&) {
+        return std::nullopt;
+    }
+}
+
+std::optional<std::vector<bool>> try_supported_outputs(const MatrixElementApi& api) {
+    try {
+        return api.supported_outputs();
+    } catch (const std::exception&) {
+        return std::nullopt;
+    }
+}
+
+// LHECompleter (madspace/src/driver/lhe_output.cpp, complete_event_data())
+// can insert status==2 intermediate resonance lines into events whose
+// propagator falls inside its Breit-Wigner window. Those lines matter for
+// LHE consumers but are not part of what a compiled matrix element expects
+// as external legs (api.particle_count() only counts incoming+outgoing
+// partons), so the per-event view fed to UMAMI must exclude them. Builds a
+// fresh process from copies of proc.events with per-event indices
+// restricted to non-intermediate (status != 2) partons, leaving proc and
+// its underlying (possibly shared) events untouched.
+REX::process external_legs_process(REX::process& proc, const IncomingSwapped* incoming_swapped) {
+    std::vector<REX::event> filtered_events;
+    filtered_events.reserve(proc.events.size());
+    for (auto& ev_ptr : proc.events) {
+        REX::event copy = *ev_ptr;
+        std::vector<std::size_t> external_indices;
+        external_indices.reserve(copy.n_);
+        for (std::size_t i = 0; i < copy.n_; ++i) {
+            if (copy.status_[i] != 2) external_indices.push_back(i);
+        }
+        // the matrix element's leg order for a beam-swapped event (see
+        // IncomingSwapped): the two incoming legs come first either way
+        if (incoming_swapped && *incoming_swapped && external_indices.size() >= 2 &&
+            (*incoming_swapped)(*ev_ptr)) {
+            std::swap(external_indices[0], external_indices[1]);
+        }
+        copy.set_indices(external_indices);
+        filtered_events.push_back(std::move(copy));
+    }
+    return REX::process(std::move(filtered_events), /*filter_partons=*/true, /*column_major=*/false);
+}
+
+void check_particle_counts(const MatrixElementApi& api, REX::process& proc) {
+    std::size_t expected = api.particle_count();
+    for (std::size_t i = 0; i < proc.size(); ++i) {
+        std::size_t begin = (i == 0) ? 0 : proc.n_summed[i - 1];
+        std::size_t end = proc.n_summed[i];
+        if (end - begin != expected) {
+            throw std::runtime_error(std::format(
+                "make_weightor: event {} has {} particles, but matrix element {} "
+                "expects {}",
+                i,
+                end - begin,
+                api.file_name(),
+                expected
+            ));
+        }
+    }
+}
+
+// Check whether the process has been transposed to the UMAMI
+// memory layout, and if not, transpose it
+// Also checks that the number of particles in the process
+// matches the number expected by the matrix element
+bool gather_momenta(const MatrixElementApi& api, REX::process& proc) {
+    std::size_t count = proc.size();
+    std::size_t particles = api.particle_count();
+    if (proc.umami_momenta().size() == 0) proc.to_umami();
+    if (proc.umami_momenta().size() != count * particles * 4) {
+        throw std::runtime_error(std::format(
+            "make_weightor: process has {} events, but matrix element {} expects "
+            "{} particles per event",
+            count,
+            api.file_name(),
+            particles
+        ));
+    }
+    return true;
+}
+
+// Owns one device-side scratch allocation used to stage a single input or
+// output across to a CUDA/HIP matrix element library: unlike CPU UMAMI
+// implementations, those expect device-resident pointers, not the host
+// memory REX::process stores its data in. Freed through the same Device
+// regardless of how evaluate() exits (including via check_umami_status's
+// exceptions), so callers don't need their own cleanup path.
+class DeviceScratch {
+public:
+    DeviceScratch() = default;
+    DeviceScratch(madspace::DevicePtr device, std::size_t bytes) : _device(device) {
+        if (bytes > 0) _ptr = _device->allocate(bytes, madspace::AllocHint::normal).first;
+    }
+    ~DeviceScratch() {
+        if (_ptr) _device->free(_ptr);
+    }
+    DeviceScratch(const DeviceScratch&) = delete;
+    DeviceScratch& operator=(const DeviceScratch&) = delete;
+    DeviceScratch(DeviceScratch&& other) noexcept : _device(other._device), _ptr(other._ptr) {
+        other._ptr = nullptr;
+    }
+    DeviceScratch& operator=(DeviceScratch&& other) noexcept {
+        if (this != &other) {
+            if (_ptr) _device->free(_ptr);
+            _device = other._device;
+            _ptr = other._ptr;
+            other._ptr = nullptr;
+        }
+        return *this;
+    }
+    void* get() const { return _ptr; }
+
+private:
+    madspace::DevicePtr _device = nullptr;
+    void* _ptr = nullptr;
+};
+
+using detail::warn_once;
+
+// The input keys passed for proc, after checking that proc can be reweighted
+// with them: with UMAMI_IN_HELICITY_INDEX (helicity reweighting) every event
+// of proc must carry a helicity index (event::helicity_, set by the
+// process's event checker from the LHE SPINUP column), ie a SPINUP of -1, 0
+// or +1 for every particle. A sample without that helicity information (eg
+// SPINUP 9, "unknown/summed") cannot be reweighted helicity by helicity:
+// throws, rather than silently summing over helicities instead.
+const std::vector<UmamiInputKey>& keys_for_process(
+    const MatrixElementApi& api,
+    const REX::process& proc,
+    std::size_t count,
+    const std::vector<UmamiInputKey>& input_keys
+) {
+    if (std::find(input_keys.begin(), input_keys.end(), UMAMI_IN_HELICITY_INDEX) ==
+        input_keys.end()) {
+        return input_keys;
+    }
+    const auto& helicity = proc.helicity();
+    if (helicity.size() != count) {
+        throw std::runtime_error(std::format(
+            "helicity reweighting: the events matched to matrix element {} carry no "
+            "helicity indices; reweight with helicity summation instead "
+            "(set helicity_reweighting false in the reweight card, or "
+            "--no-helicity-reweighting)",
+            api.file_name()
+        ));
+    }
+    auto missing = std::find_if(helicity.begin(), helicity.end(), [](int h) { return h < 0; });
+    if (missing != helicity.end() && *missing == helicity_not_in_table) {
+        throw std::runtime_error(std::format(
+            "helicity reweighting: the helicity configuration (LHE SPINUP) of event {} "
+            "(0-based) of the {} events matched to matrix element {} is not in the "
+            "helicity table of its subprocess, so it has no helicity index in the "
+            "matrix element; the sample does not match this process directory",
+            missing - helicity.begin(),
+            count,
+            api.file_name()
+        ));
+    }
+    if (missing != helicity.end()) {
+        throw std::runtime_error(std::format(
+            "helicity reweighting: event {} (0-based) of the {} events matched to "
+            "matrix element {} has no helicity information (a SPINUP other than -1, "
+            "0 or +1, eg 9, "
+            "for at least one particle); reweight this sample with helicity "
+            "summation instead (set helicity_reweighting false in the reweight card, "
+            "or --no-helicity-reweighting)",
+            missing - helicity.begin(),
+            count,
+            api.file_name()
+        ));
+    }
+    return input_keys;
+}
+
+// Per-weightor-closure cache of everything evaluate() derives from a
+// REX::process: momenta/alphaS/flavor/helicity (and, on cuda/hip, their
+// device-resident copies) don't change between reweighting iterations --
+// only UMAMI parameters do, via umami_set_parameter -- so gathering them
+// (and, off-cpu, copying them to the device) only needs to happen once per
+// process this weightor ends up bound to, not on every evaluate() call.
+// Output storage (host buffers, and on-device their staging buffers) is
+// also kept here so its *allocation* is reused across iterations, even
+// though its *contents* are still overwritten by every call, since the
+// matrix element's value does change iteration to iteration.
+//
+// Copying an EvalCache -- which happens whenever its owning weightor
+// (and thus its owning procReweightor) is itself copied, eg
+// TupperWare::build()'s rwgts.push_back(*weightor_ptr), or
+// reweightor::extract_sorter()'s max_batch overflow cloning -- deliberately
+// resets to a blank cache rather than duplicating the live buffers: the two
+// copies always end up bound to different processes (or, for max_batch
+// clones, may even run concurrently on different threads), so sharing
+// state across them would be wrong or unsafe. Each repopulates
+// independently on its own first call.
+struct EvalCache {
+    EvalCache() = default;
+    EvalCache(const EvalCache&) : EvalCache() {}
+    EvalCache& operator=(const EvalCache&) {
+        *this = EvalCache();
+        return *this;
+    }
+    EvalCache(EvalCache&&) noexcept = default;
+    EvalCache& operator=(EvalCache&&) noexcept = default;
+
+    const REX::process* proc_id = nullptr; // identifies which process this cache is valid for
+    std::size_t count = 0;
+
+    // The MatrixElementApi instance actually evaluated against for proc_id:
+    // either the api the owning weightor was built with, or -- if built via
+    // the ReloadSpec overload -- a freshly-loaded instance obtained the
+    // first time this cache populated for proc_id. Resolved once per
+    // process (see evaluate()) and reused for every later call against the
+    // same process; resets to nullptr, like the rest of EvalCache, whenever
+    // this weightor's closure is copied (see the copy ctor below), so each
+    // copy resolves (and, with a ReloadSpec, reloads) independently on its
+    // own first use.
+    const MatrixElementApi* active_api = nullptr;
+
+    // Backs input_ptrs[UMAMI_IN_MOMENTA] on cpu; must outlive the cache entry.
+    REX::process external_proc;
+    std::vector<UmamiInputKey> active_inputs;
+    std::vector<void const*> input_ptrs;
+    std::vector<DeviceScratch> input_scratch; // keeps device input buffers alive; empty on cpu
+
+    std::unordered_map<UmamiOutputKey, std::vector<double>> output_buffers;
+    std::vector<void*> output_ptrs; // positional, aligned to output_keys
+    std::vector<DeviceScratch> output_scratch; // reused device output allocation; empty on cpu
+};
+
+std::shared_ptr<std::vector<double>> evaluate(
+    const MatrixElementApi& api,
+    REX::process& proc,
+    const std::vector<UmamiInputKey>& input_keys,
+    const std::vector<UmamiOutputKey>& output_keys,
+    const std::vector<UmamiOutputKey>& result_keys,
+    EvalCache& cache,
+    const ReloadSpec* reload = nullptr,
+    const IncomingSwapped* incoming_swapped = nullptr
+) {
+    std::size_t count = proc.size();
+    if (count == 0) {
+        return std::make_shared<std::vector<double>>();
+    }
+
+    // Which MatrixElementApi instance this call actually evaluates against:
+    // whichever one was already resolved into the cache for this process
+    // (reused verbatim, cache hit or not -- reloading again on every call
+    // would defeat the purpose of a per-process reload), or, the first time
+    // this cache sees a given process, either a fresh instance (reload set)
+    // or api itself (no reload spec, the pre-existing behavior).
+    const MatrixElementApi* resolved = cache.active_api;
+    bool new_process = cache.proc_id != &proc || cache.count != count;
+    if (!resolved || new_process) {
+        resolved = reload ? &detail::reload_matrix_element(*reload, proc) : &api;
+    }
+    const MatrixElementApi& active_api = *resolved;
+
+    // REX::process always stores its data on the host, but a cuda/hip matrix
+    // element's umami_matrix_element expects device-resident input/output
+    // pointers. When that's the case, inputs are staged through DeviceScratch
+    // buffers copied from the host data below, and outputs are read back into
+    // the host-side output_buffers after the call; on cpu, host pointers are
+    // passed straight through as before.
+    bool on_device = active_api.device_type() != madspace::DeviceType::cpu;
+    madspace::DevicePtr device = on_device ? active_api.device() : nullptr;
+
+    if (new_process) {
+        cache = EvalCache{};
+        cache.proc_id = &proc;
+        cache.count = count;
+        cache.active_api = resolved;
+
+        // external_proc drops any status==2 intermediate resonance lines
+        // LHECompleter may have inserted, so particle counts/momenta match
+        // what the matrix element expects (its external legs); event-level
+        // quantities (alphaS/flavor/helicity below) are unaffected by that
+        // filtering and still come from proc directly.
+        cache.external_proc = external_legs_process(proc, incoming_swapped);
+        check_particle_counts(active_api, cache.external_proc);
+
+        const auto& process_keys = keys_for_process(active_api, proc, count, input_keys);
+        cache.active_inputs.reserve(process_keys.size());
+        cache.input_ptrs.reserve(process_keys.size());
+
+        auto stage_input = [&](UmamiInputKey key, const void* host_ptr, std::size_t bytes) {
+            cache.active_inputs.push_back(key);
+            if (on_device) {
+                DeviceScratch scratch(device, bytes);
+                device->memcpy(scratch.get(), const_cast<void*>(host_ptr), bytes);
+                cache.input_ptrs.push_back(scratch.get());
+                cache.input_scratch.push_back(std::move(scratch));
+            } else {
+                cache.input_ptrs.push_back(host_ptr);
+            }
+        };
+
+        for (auto key : process_keys) {
+            switch (key) {
+            case UMAMI_IN_MOMENTA:
+                if (!gather_momenta(active_api, cache.external_proc)) {
+                    throw std::runtime_error(
+                        "make_weightor: failed to gather momenta for UMAMI"
+                    );
+                }
+                stage_input(
+                    key,
+                    cache.external_proc.umami_momenta().data(),
+                    cache.external_proc.umami_momenta().size() * sizeof(double)
+                );
+                break;
+            case UMAMI_IN_ALPHA_S:
+                if (proc.alphaS().size() != count) {
+                    throw std::runtime_error(
+                        "make_weightor: process does not carry a per-event alpha_s "
+                        "value for every event"
+                    );
+                }
+                stage_input(key, proc.alphaS().data(), count * sizeof(double));
+                break;
+            case UMAMI_IN_FLAVOR_INDEX:
+                if (proc.flavor().size() != count) {
+                    throw std::runtime_error(
+                        "make_weightor: process does not carry a per-event flavor "
+                        "index for every event"
+                    );
+                }
+                stage_input(key, proc.flavor().data(), count * sizeof(int));
+                break;
+            case UMAMI_IN_HELICITY_INDEX:
+                if (proc.helicity().size() != count) {
+                    throw std::runtime_error(
+                        "make_weightor: process does not carry a per-event helicity "
+                        "index for every event"
+                    );
+                }
+                stage_input(key, proc.helicity().data(), count * sizeof(int));
+                break;
+            default:
+                throw std::runtime_error(
+                    "make_weightor: input key cannot be derived from a REX::process"
+                );
+            }
+        }
+
+        // On device, one DeviceScratch per output_keys entry (positionally
+        // aligned with output_ptrs/output_keys) rather than per unique key,
+        // so it's read back into the right host buffer below regardless of
+        // call order. Allocated once here and reused (overwritten) by every
+        // subsequent call for this process, not freed/reallocated per call.
+        cache.output_ptrs.reserve(output_keys.size());
+        for (auto key : output_keys) {
+            std::size_t buf_size;
+            switch (key) {
+            case UMAMI_OUT_MATRIX_ELEMENT:
+                buf_size = count;
+                break;
+            case UMAMI_OUT_DIAGRAM_AMP2:
+                buf_size = active_api.diagram_count() * count;
+                break;
+            default:
+                throw std::runtime_error(
+                    "make_weightor: output key is not a reweighting-usable value"
+                );
+            }
+            auto [it, inserted] =
+                cache.output_buffers.emplace(key, std::vector<double>(buf_size, 0.0));
+            if (on_device) {
+                cache.output_scratch.emplace_back(device, buf_size * sizeof(double));
+                cache.output_ptrs.push_back(cache.output_scratch.back().get());
+            } else {
+                cache.output_ptrs.push_back(it->second.data());
+            }
+        }
+    }
+
+    {
+        std::shared_lock<std::shared_mutex> parameters(detail::parameters_lock());
+        std::lock_guard<std::mutex> guard(detail::api_call_mutex(active_api));
+        active_api.call(
+            active_api.process_instance(),
+            count,
+            count,
+            0,
+            cache.active_inputs.size(),
+            cache.active_inputs.data(),
+            cache.input_ptrs.data(),
+            output_keys.size(),
+            output_keys.data(),
+            cache.output_ptrs.data()
+        );
+        if (on_device) {
+            // No caller here ever requests UMAMI_OUT_GPU_STREAM, so
+            // umami_matrix_element is expected to have fully finished (not
+            // just launched) by the time it returns, making it safe to read
+            // the device outputs back immediately. Unlike the inputs above,
+            // this copy-back runs on every call (cache hit or not): the
+            // matrix element's value legitimately changes each iteration.
+            for (std::size_t i = 0; i < output_keys.size(); ++i) {
+                auto& host_buf = cache.output_buffers.at(output_keys[i]);
+                device->memcpy(
+                    host_buf.data(), cache.output_scratch[i].get(),
+                    host_buf.size() * sizeof(double)
+                );
+            }
+        }
+    }
+
+    auto result = std::make_shared<std::vector<double>>(count, 0.0);
+    for (auto key : result_keys) {
+        auto found = cache.output_buffers.find(key);
+        if (found == cache.output_buffers.end()) {
+            throw std::runtime_error(
+                "make_weightor: result key was not among the requested output keys"
+            );
+        }
+        if (key == UMAMI_OUT_MATRIX_ELEMENT) {
+            for (std::size_t e = 0; e < count; ++e) {
+                (*result)[e] += found->second[e];
+            }
+        } else { // UMAMI_OUT_DIAGRAM_AMP2
+            std::size_t diagrams = active_api.diagram_count();
+            for (std::size_t e = 0; e < count; ++e) {
+                for (std::size_t d = 0; d < diagrams; ++d) {
+                    (*result)[e] += found->second[d * count + e];
+                }
+            }
+        }
+    }
+    return result;
+}
+
+// Per-(api, process) device-resident cache for the deferred/cuda-hip
+// evaluation path, attached to the process itself via REX::process's
+// generic, opaque device_data slot (Rex.h) rather than living inside the
+// weightor closure -- unlike EvalCache above. This ties its lifetime
+// directly to the process it was derived from (found/populated fresh
+// whenever a process doesn't already have one under this api's key,
+// destroyed automatically when proc is), and is exactly what
+// process::device_data is documented for: "a standard place to stash and
+// retrieve raw pointers/handles... for off-CPU reweighting functions."
+// A pleasant consequence: unlike EvalCache, the closure that uses this
+// needs no mutable captured state and no reset-on-copy trick to stay safe
+// under procReweightor cloning (eg reweightor::extract_sorter()'s
+// max_batch overflow clones) -- a cloned closure just looks up
+// device_data on whatever process it ends up bound to, which naturally
+// starts out unpopulated for a different process.
+//
+// Only ever used for cuda/hip apis (see make_deferred_weightor), so unlike
+// EvalCache it needs no cpu fallback: every input is always staged to
+// device, and every output always lives in an N-iteration-wide device
+// buffer -- one contiguous allocation per output key, sized for exactly
+// n_iterations calls, since deferring the copy-back means each iteration's
+// result must survive, un-overwritten, until it's actually pulled at
+// pull_deferred_backlog() time.
+struct DeferredDeviceCache {
+    std::size_t count = 0;
+    std::size_t n_iterations = 0;
+    std::size_t next_iter = 0;
+
+    // As EvalCache::active_api: the MatrixElementApi instance actually
+    // evaluated against for this process -- either the api the owning
+    // deferred_weightor was built with, or, if built via the ReloadSpec
+    // overload, a freshly-loaded instance obtained when this cache was
+    // first populated. Since this cache already lives per-process (via
+    // proc.device_data), no separate "is this a new process" bookkeeping is
+    // needed here the way EvalCache needs it: a fresh process simply has no
+    // device_data yet, so its cache -- and this field -- start blank.
+    const MatrixElementApi* active_api = nullptr;
+
+    std::vector<UmamiInputKey> active_inputs;
+    std::vector<void const*> input_ptrs; // device pointers
+    std::vector<DeviceScratch> input_scratch;
+
+    std::unordered_map<UmamiOutputKey, DeviceScratch> output_scratch; // one N-iteration-wide buffer per key
+    std::unordered_map<UmamiOutputKey, std::size_t> per_iter_elems;
+};
+
+// Unique per api instance (pointer identity, stable for the api's
+// lifetime -- MatrixElementApis are owned indefinitely by their Context),
+// so multiple deferred apis evaluating the same process don't collide on
+// the same device_data slot.
+std::string deferred_cache_key(const MatrixElementApi& api) {
+    return std::format("madtrex::deferred_weightor::{}", static_cast<const void*>(&api));
+}
+
+REX::tea::weight_puller deferred_evaluate(
+    const MatrixElementApi& api,
+    REX::process& proc,
+    const std::vector<UmamiInputKey>& input_keys,
+    const std::vector<UmamiOutputKey>& output_keys,
+    const std::vector<UmamiOutputKey>& result_keys,
+    std::size_t n_iterations,
+    const ReloadSpec* reload = nullptr,
+    const IncomingSwapped* incoming_swapped = nullptr
+) {
+    std::size_t count = proc.size();
+    if (count == 0) {
+        return []() { return std::make_shared<std::vector<double>>(); };
+    }
+
+    std::string cache_key = deferred_cache_key(api);
+
+    std::shared_ptr<DeferredDeviceCache> cache;
+    if (proc.has_device_data(cache_key)) {
+        cache = std::any_cast<std::shared_ptr<DeferredDeviceCache>>(proc.get_device_data(cache_key).tag);
+    } else {
+        cache = std::make_shared<DeferredDeviceCache>();
+        cache->count = count;
+        cache->n_iterations = n_iterations;
+        cache->active_api = reload ? &detail::reload_matrix_element(*reload, proc) : &api;
+        const MatrixElementApi& active_api = *cache->active_api;
+        madspace::DevicePtr device = active_api.device();
+
+        // Only needed transiently to compute the initial host-side momenta
+        // for the one-time copy to device below; unlike EvalCache's
+        // external_proc, nothing here keeps it around afterward, since the
+        // deferred path never hands out host pointers.
+        REX::process external_proc = external_legs_process(proc, incoming_swapped);
+        check_particle_counts(active_api, external_proc);
+
+        const auto& process_keys = keys_for_process(active_api, proc, count, input_keys);
+        cache->active_inputs.reserve(process_keys.size());
+        cache->input_ptrs.reserve(process_keys.size());
+
+        auto stage_input = [&](UmamiInputKey key, const void* host_ptr, std::size_t bytes) {
+            cache->active_inputs.push_back(key);
+            DeviceScratch scratch(device, bytes);
+            device->memcpy(scratch.get(), const_cast<void*>(host_ptr), bytes);
+            cache->input_ptrs.push_back(scratch.get());
+            cache->input_scratch.push_back(std::move(scratch));
+        };
+
+        for (auto key : process_keys) {
+            switch (key) {
+            case UMAMI_IN_MOMENTA:
+                if (!gather_momenta(active_api, external_proc)) {
+                    throw std::runtime_error(
+                        "make_deferred_weightor: failed to gather momenta for UMAMI"
+                    );
+                }
+                stage_input(
+                    key,
+                    external_proc.umami_momenta().data(),
+                    external_proc.umami_momenta().size() * sizeof(double)
+                );
+                break;
+            case UMAMI_IN_ALPHA_S:
+                if (proc.alphaS().size() != count) {
+                    throw std::runtime_error(
+                        "make_deferred_weightor: process does not carry a per-event "
+                        "alpha_s value for every event"
+                    );
+                }
+                stage_input(key, proc.alphaS().data(), count * sizeof(double));
+                break;
+            case UMAMI_IN_FLAVOR_INDEX:
+                if (proc.flavor().size() != count) {
+                    throw std::runtime_error(
+                        "make_deferred_weightor: process does not carry a per-event "
+                        "flavor index for every event"
+                    );
+                }
+                stage_input(key, proc.flavor().data(), count * sizeof(int));
+                break;
+            case UMAMI_IN_HELICITY_INDEX:
+                if (proc.helicity().size() != count) {
+                    throw std::runtime_error(
+                        "make_deferred_weightor: process does not carry a per-event "
+                        "helicity index for every event"
+                    );
+                }
+                stage_input(key, proc.helicity().data(), count * sizeof(int));
+                break;
+            default:
+                throw std::runtime_error(
+                    "make_deferred_weightor: input key cannot be derived from a "
+                    "REX::process"
+                );
+            }
+        }
+
+        for (auto key : output_keys) {
+            std::size_t per_iter;
+            switch (key) {
+            case UMAMI_OUT_MATRIX_ELEMENT:
+                per_iter = count;
+                break;
+            case UMAMI_OUT_DIAGRAM_AMP2:
+                per_iter = active_api.diagram_count() * count;
+                break;
+            default:
+                throw std::runtime_error(
+                    "make_deferred_weightor: output key is not a reweighting-usable "
+                    "value"
+                );
+            }
+            cache->per_iter_elems[key] = per_iter;
+            cache->output_scratch.emplace(
+                key, DeviceScratch(device, per_iter * n_iterations * sizeof(double))
+            );
+        }
+
+        proc.set_device_data(cache_key, nullptr, 0, std::any(cache));
+    }
+
+    // Whether this call populated the cache above or found it already
+    // populated (ie a later iteration against the same process), active_api
+    // is the instance that was resolved -- and, from here on, is the only
+    // one ever evaluated against for this process.
+    const MatrixElementApi& active_api = *cache->active_api;
+    madspace::DevicePtr device = active_api.device();
+
+    if (cache->count != count) {
+        throw std::runtime_error(
+            "make_deferred_weightor: process's event count changed between calls; a "
+            "deferred weightor's device_data cache is only valid for a process whose "
+            "events never change after its first evaluation"
+        );
+    }
+    if (cache->next_iter >= cache->n_iterations) {
+        throw std::runtime_error(
+            "make_deferred_weightor: invoked more times than the reweighting run's "
+            "declared iteration count; the n_iterations passed to "
+            "TupperWare::build()/WareHouse must match how many times the reweightor "
+            "actually iterates"
+        );
+    }
+    std::size_t iter = cache->next_iter++;
+
+    std::vector<void*> output_ptrs;
+    output_ptrs.reserve(output_keys.size());
+    for (auto key : output_keys) {
+        auto& scratch = cache->output_scratch.at(key);
+        std::size_t per_iter = cache->per_iter_elems.at(key);
+        char* base = static_cast<char*>(scratch.get());
+        output_ptrs.push_back(base + iter * per_iter * sizeof(double));
+    }
+
+    {
+        std::shared_lock<std::shared_mutex> parameters(detail::parameters_lock());
+        std::lock_guard<std::mutex> guard(detail::api_call_mutex(active_api));
+        active_api.call(
+            active_api.process_instance(),
+            count,
+            count,
+            0,
+            cache->active_inputs.size(),
+            cache->active_inputs.data(),
+            cache->input_ptrs.data(),
+            output_keys.size(),
+            output_keys.data(),
+            output_ptrs.data()
+        );
+        // Deliberately no copy-back here: this iteration's outputs stay on
+        // device, in their own dedicated slice, until the puller below is
+        // invoked -- by pull_deferred_backlog(), once every iteration has
+        // run, not after this one specifically.
+    }
+
+    std::size_t diagrams = active_api.diagram_count();
+    std::vector<std::pair<UmamiOutputKey, void*>> this_iter_output_ptrs;
+    this_iter_output_ptrs.reserve(output_keys.size());
+    for (std::size_t i = 0; i < output_keys.size(); ++i) {
+        this_iter_output_ptrs.emplace_back(output_keys[i], output_ptrs[i]);
+    }
+
+    // cache is captured too, purely as a keep-alive guard: its device
+    // buffers are already guaranteed to outlive this puller via proc's own
+    // device_data (proc stays alive at least through
+    // reweightor::finalise_reweighting(), which is what invokes this), but
+    // holding the shared_ptr here as well costs nothing and removes any
+    // reliance on that reasoning staying true if this code is ever reused
+    // in a context with different lifetime guarantees.
+    return [device, count, diagrams, result_keys,
+            this_iter_output_ptrs = std::move(this_iter_output_ptrs),
+            cache]() -> std::shared_ptr<std::vector<double>> {
+        std::unordered_map<UmamiOutputKey, std::vector<double>> host_buffers;
+        for (auto& [key, dev_ptr] : this_iter_output_ptrs) {
+            std::size_t elems = (key == UMAMI_OUT_DIAGRAM_AMP2) ? diagrams * count : count;
+            std::vector<double> host_buf(elems, 0.0);
+            device->memcpy(host_buf.data(), dev_ptr, elems * sizeof(double));
+            host_buffers.emplace(key, std::move(host_buf));
+        }
+        auto result = std::make_shared<std::vector<double>>(count, 0.0);
+        for (auto key : result_keys) {
+            auto found = host_buffers.find(key);
+            if (found == host_buffers.end()) {
+                throw std::runtime_error(
+                    "make_deferred_weightor: result key was not among the requested "
+                    "output keys"
+                );
+            }
+            if (key == UMAMI_OUT_MATRIX_ELEMENT) {
+                for (std::size_t e = 0; e < count; ++e) {
+                    (*result)[e] += found->second[e];
+                }
+            } else { // UMAMI_OUT_DIAGRAM_AMP2
+                for (std::size_t e = 0; e < count; ++e) {
+                    for (std::size_t d = 0; d < diagrams; ++d) {
+                        (*result)[e] += found->second[d * count + e];
+                    }
+                }
+            }
+        }
+        return result;
+    };
+}
+
+// The normaliser of TupperWare::set_stored_normalisation(): each event's
+// stored matrix element, process.extra[key], or evaluate's result for any
+// process with an event that has none (another source, a file without them).
+REX::tea::weightor stored_normaliser(std::string key, REX::tea::weightor evaluate) {
+    return [key = std::move(key), evaluate = std::move(evaluate)](REX::process& proc) {
+        auto it = proc.extra.find(key);
+        if (it != proc.extra.end() && it->second.size() == proc.weight_.size()) {
+            auto stored = std::make_shared<std::vector<double>>();
+            stored->reserve(it->second.size());
+            for (const auto& value : it->second) {
+                const double* me = std::any_cast<double>(&value);
+                if (me == nullptr) {
+                    stored = nullptr;
+                    break;
+                }
+                stored->push_back(*me);
+            }
+            if (stored) return stored;
+        }
+        return evaluate(proc);
+    };
+}
+
+} // namespace
+
+std::vector<UmamiInputKey> default_input_keys(const MatrixElementApi& api, bool helicity_reweighting) {
+    auto supported = try_supported_inputs(api);
+    if (!supported) {
+        throw std::runtime_error(std::format(
+            "make_weightor: matrix element {} does not report supported inputs; "
+            "pass input keys explicitly",
+            api.file_name()
+        ));
+    }
+    auto required = try_required_inputs(api);
+
+    static constexpr UmamiInputKey derivable[] = {
+        UMAMI_IN_MOMENTA, UMAMI_IN_ALPHA_S, UMAMI_IN_FLAVOR_INDEX, UMAMI_IN_HELICITY_INDEX
+    };
+
+    std::vector<UmamiInputKey> keys;
+    for (std::size_t i = 0; i < UMAMI_INPUT_KEY_COUNT; ++i) {
+        auto key = static_cast<UmamiInputKey>(i);
+        bool is_derivable = std::find(std::begin(derivable), std::end(derivable), key) !=
+                             std::end(derivable);
+        bool is_required = required && (*required)[i];
+        if (is_required && !is_derivable) {
+            throw std::runtime_error(std::format(
+                "make_weightor: matrix element {} requires input key {}, which "
+                "cannot be derived from a REX::process; pass input keys explicitly",
+                api.file_name(),
+                i
+            ));
+        }
+        if (key == UMAMI_IN_HELICITY_INDEX && helicity_reweighting && !(*supported)[i]) {
+            warn_once(
+                "no-helicity-input:" + api.file_name(),
+                std::format(
+                    "helicity reweighting is on, but matrix element {} does not support "
+                    "per-event helicities (UMAMI_IN_HELICITY_INDEX); using "
+                    "helicity-summed matrix elements for it",
+                    api.file_name()
+                )
+            );
+        }
+        if (!is_derivable || !(*supported)[i]) {
+            continue;
+        }
+        // UMAMI_IN_HELICITY_INDEX only with helicity reweighting (see
+        // default_input_keys()'s doc comment), or when the matrix element
+        // requires it outright.
+        if (key == UMAMI_IN_HELICITY_INDEX && !is_required && !helicity_reweighting) {
+            continue;
+        }
+        keys.push_back(key);
+    }
+    return keys;
+}
+
+std::vector<UmamiOutputKey> default_output_keys(const MatrixElementApi& api) {
+    auto supported = try_supported_outputs(api);
+    if (!supported || !(*supported)[UMAMI_OUT_MATRIX_ELEMENT]) {
+        throw std::runtime_error(std::format(
+            "make_weightor: matrix element {} does not report supporting "
+            "UMAMI_OUT_MATRIX_ELEMENT; pass output/result keys explicitly",
+            api.file_name()
+        ));
+    }
+    return {UMAMI_OUT_MATRIX_ELEMENT};
+}
+
+std::shared_ptr<REX::tea::weightor> make_weightor(const MatrixElementApi& api) {
+    return make_weightor(
+        api, default_input_keys(api), default_output_keys(api), {UMAMI_OUT_MATRIX_ELEMENT}
+    );
+}
+
+std::shared_ptr<REX::tea::weightor> make_weightor(
+    const MatrixElementApi& api,
+    std::vector<UmamiInputKey> input_keys,
+    std::vector<UmamiOutputKey> output_keys,
+    std::vector<UmamiOutputKey> result_keys,
+    IncomingSwapped incoming_swapped
+) {
+    if (result_keys.empty()) {
+        result_keys = {UMAMI_OUT_MATRIX_ELEMENT};
+    }
+    for (auto key : result_keys) {
+        if (std::find(output_keys.begin(), output_keys.end(), key) == output_keys.end()) {
+            throw std::runtime_error(
+                "make_weightor: result key must be among the requested output keys"
+            );
+        }
+    }
+
+    const MatrixElementApi* api_ptr = &api;
+    return std::make_shared<REX::tea::weightor>(
+        [api_ptr,
+         input_keys = std::move(input_keys),
+         output_keys = std::move(output_keys),
+         result_keys = std::move(result_keys),
+         incoming_swapped = std::move(incoming_swapped),
+         cache = EvalCache{}](REX::process& proc) mutable {
+            return evaluate(
+                *api_ptr, proc, input_keys, output_keys, result_keys, cache, nullptr,
+                &incoming_swapped
+            );
+        }
+    );
+}
+
+std::shared_ptr<REX::tea::weightor> make_weightor(
+    const MatrixElementApi& api,
+    std::vector<UmamiInputKey> input_keys,
+    std::vector<UmamiOutputKey> output_keys,
+    std::vector<UmamiOutputKey> result_keys,
+    ReloadSpec reload,
+    IncomingSwapped incoming_swapped
+) {
+    if (result_keys.empty()) {
+        result_keys = {UMAMI_OUT_MATRIX_ELEMENT};
+    }
+    for (auto key : result_keys) {
+        if (std::find(output_keys.begin(), output_keys.end(), key) == output_keys.end()) {
+            throw std::runtime_error(
+                "make_weightor: result key must be among the requested output keys"
+            );
+        }
+    }
+
+    const MatrixElementApi* api_ptr = &api;
+    return std::make_shared<REX::tea::weightor>(
+        [api_ptr,
+         input_keys = std::move(input_keys),
+         output_keys = std::move(output_keys),
+         result_keys = std::move(result_keys),
+         reload = std::move(reload),
+         incoming_swapped = std::move(incoming_swapped),
+         cache = EvalCache{}](REX::process& proc) mutable {
+            return evaluate(
+                *api_ptr, proc, input_keys, output_keys, result_keys, cache, &reload,
+                &incoming_swapped
+            );
+        }
+    );
+}
+
+std::shared_ptr<REX::tea::deferred_weightor>
+make_deferred_weightor(const MatrixElementApi& api, std::size_t n_iterations) {
+    return make_deferred_weightor(
+        api, default_input_keys(api), default_output_keys(api), {UMAMI_OUT_MATRIX_ELEMENT},
+        n_iterations
+    );
+}
+
+std::shared_ptr<REX::tea::deferred_weightor> make_deferred_weightor(
+    const MatrixElementApi& api,
+    std::vector<UmamiInputKey> input_keys,
+    std::vector<UmamiOutputKey> output_keys,
+    std::vector<UmamiOutputKey> result_keys,
+    std::size_t n_iterations,
+    IncomingSwapped incoming_swapped
+) {
+    if (result_keys.empty()) {
+        result_keys = {UMAMI_OUT_MATRIX_ELEMENT};
+    }
+    for (auto key : result_keys) {
+        if (std::find(output_keys.begin(), output_keys.end(), key) == output_keys.end()) {
+            throw std::runtime_error(
+                "make_deferred_weightor: result key must be among the requested output "
+                "keys"
+            );
+        }
+    }
+
+    const MatrixElementApi* api_ptr = &api;
+    // Unlike make_weightor's closure, this one needs no mutable captured
+    // state (and hence no `mutable`, no reset-on-copy handling): its cache
+    // lives on the process passed in, not inside the closure -- see
+    // DeferredDeviceCache's doc comment above.
+    return std::make_shared<REX::tea::deferred_weightor>(
+        [api_ptr,
+         input_keys = std::move(input_keys),
+         output_keys = std::move(output_keys),
+         result_keys = std::move(result_keys),
+         n_iterations,
+         incoming_swapped = std::move(incoming_swapped)](REX::process& proc) -> REX::tea::weight_puller {
+            return deferred_evaluate(
+                *api_ptr, proc, input_keys, output_keys, result_keys, n_iterations, nullptr,
+                &incoming_swapped
+            );
+        }
+    );
+}
+
+std::shared_ptr<REX::tea::deferred_weightor> make_deferred_weightor(
+    const MatrixElementApi& api,
+    std::vector<UmamiInputKey> input_keys,
+    std::vector<UmamiOutputKey> output_keys,
+    std::vector<UmamiOutputKey> result_keys,
+    std::size_t n_iterations,
+    ReloadSpec reload,
+    IncomingSwapped incoming_swapped
+) {
+    if (result_keys.empty()) {
+        result_keys = {UMAMI_OUT_MATRIX_ELEMENT};
+    }
+    for (auto key : result_keys) {
+        if (std::find(output_keys.begin(), output_keys.end(), key) == output_keys.end()) {
+            throw std::runtime_error(
+                "make_deferred_weightor: result key must be among the requested output "
+                "keys"
+            );
+        }
+    }
+
+    const MatrixElementApi* api_ptr = &api;
+    return std::make_shared<REX::tea::deferred_weightor>(
+        [api_ptr,
+         input_keys = std::move(input_keys),
+         output_keys = std::move(output_keys),
+         result_keys = std::move(result_keys),
+         n_iterations,
+         reload = std::move(reload),
+         incoming_swapped = std::move(incoming_swapped)](REX::process& proc) -> REX::tea::weight_puller {
+            return deferred_evaluate(
+                *api_ptr, proc, input_keys, output_keys, result_keys, n_iterations, &reload,
+                &incoming_swapped
+            );
+        }
+    );
+}
+
+TupperWare::TupperWare(const madspace::MatrixElementApi& api, REX::eventBelongs checker) {
+    add_api(api);
+    set_normaliser(api);
+    set_event_checker(std::move(checker));
+}
+
+TupperWare::TupperWare(const madspace::MatrixElementApi& api, REX::event_bool_fn checker) {
+    add_api(api);
+    set_normaliser(api);
+    set_event_checker(std::move(checker));
+}
+
+TupperWare& TupperWare::add_api(const madspace::MatrixElementApi& api) {
+    return add_api(api, {}, {});
+}
+
+TupperWare& TupperWare::add_api(
+    const madspace::MatrixElementApi& api,
+    std::vector<UmamiInputKey> input_keys,
+    std::vector<UmamiOutputKey> output_keys
+) {
+    _apis.push_back(&api);
+    _input_keys_list.push_back(std::move(input_keys));
+    _output_keys_list.push_back(std::move(output_keys));
+    _reload_specs.push_back(std::nullopt);
+    return *this;
+}
+
+TupperWare& TupperWare::add_api(const madspace::MatrixElementApi& api, ReloadSpec reload) {
+    return add_api(api, {}, {}, std::move(reload));
+}
+
+TupperWare& TupperWare::add_api(
+    const madspace::MatrixElementApi& api,
+    std::vector<UmamiInputKey> input_keys,
+    std::vector<UmamiOutputKey> output_keys,
+    ReloadSpec reload
+) {
+    add_api(api, std::move(input_keys), std::move(output_keys));
+    _reload_specs.back() = std::move(reload);
+    return *this;
+}
+
+TupperWare& TupperWare::add_api(std::vector<const madspace::MatrixElementApi*> apis) {
+    for (const auto* api : apis) {
+        add_api(*api);
+    }
+    return *this;
+}
+
+TupperWare& TupperWare::add_api(
+    std::vector<const madspace::MatrixElementApi*> apis,
+    std::vector<UmamiInputKey> input_keys_list,
+    std::vector<UmamiOutputKey> output_keys_list
+) {
+    for (std::size_t i = 0; i < apis.size(); ++i) {
+        add_api(*apis[i], input_keys_list, output_keys_list);
+    }
+    return *this;
+}
+
+TupperWare& TupperWare::add_api(
+    std::vector<const madspace::MatrixElementApi*> apis,
+    std::vector<std::vector<UmamiInputKey>> input_keys_list,
+    std::vector<std::vector<UmamiOutputKey>> output_keys_list
+) {
+    for (std::size_t i = 0; i < apis.size(); ++i) {
+        std::vector<UmamiInputKey> input_keys;
+        if (i < input_keys_list.size()) {
+            input_keys = std::move(input_keys_list[i]);
+        }
+        std::vector<UmamiOutputKey> output_keys;
+        if (i < output_keys_list.size()) {
+            output_keys = std::move(output_keys_list[i]);
+        }
+        add_api(*apis[i], input_keys, output_keys);
+    }
+    return *this;
+}
+
+TupperWare& TupperWare::set_api(std::vector<const madspace::MatrixElementApi*> apis) {
+    _apis.clear();
+    _input_keys_list.clear();
+    _output_keys_list.clear();
+    _reload_specs.clear();
+    return add_api(std::move(apis));
+}
+
+TupperWare& TupperWare::set_api(
+    std::vector<const madspace::MatrixElementApi*> apis,
+    std::vector<UmamiInputKey> input_keys_list,
+    std::vector<UmamiOutputKey> output_keys_list
+) {
+    _apis.clear();
+    _input_keys_list.clear();
+    _output_keys_list.clear();
+    _reload_specs.clear();
+    return add_api(std::move(apis), std::move(input_keys_list), std::move(output_keys_list));
+}
+
+TupperWare& TupperWare::set_api(
+    std::vector<const madspace::MatrixElementApi*> apis,
+    std::vector<std::vector<UmamiInputKey>> input_keys_list,
+    std::vector<std::vector<UmamiOutputKey>> output_keys_list
+) {
+    _apis.clear();
+    _input_keys_list.clear();
+    _output_keys_list.clear();
+    _reload_specs.clear();
+    return add_api(std::move(apis), std::move(input_keys_list), std::move(output_keys_list));
+}
+
+TupperWare& TupperWare::set_stored_normalisation(std::string key) {
+    _stored_normalisation = std::move(key);
+    return *this;
+}
+
+const std::string& TupperWare::stored_normalisation() const { return _stored_normalisation; }
+
+TupperWare& TupperWare::set_normaliser(const madspace::MatrixElementApi& api) {
+    return set_normaliser(api, {}, {});
+}
+
+TupperWare& TupperWare::set_normaliser(
+    const madspace::MatrixElementApi& api,
+    std::vector<UmamiInputKey> input_keys,
+    std::vector<UmamiOutputKey> output_keys
+) {
+    _normaliser = &api;
+    _normaliser_input_keys = std::move(input_keys);
+    _normaliser_output_keys = std::move(output_keys);
+    _normaliser_reload.reset();
+    return *this;
+}
+
+TupperWare& TupperWare::set_normaliser(const madspace::MatrixElementApi& api, ReloadSpec reload) {
+    return set_normaliser(api, {}, {}, std::move(reload));
+}
+
+TupperWare& TupperWare::set_normaliser(
+    const madspace::MatrixElementApi& api,
+    std::vector<UmamiInputKey> input_keys,
+    std::vector<UmamiOutputKey> output_keys,
+    ReloadSpec reload
+) {
+    set_normaliser(api, std::move(input_keys), std::move(output_keys));
+    _normaliser_reload = std::move(reload);
+    return *this;
+}
+
+TupperWare& TupperWare::set_result_keys(std::vector<UmamiOutputKey> keys) {
+    _result_keys = std::move(keys);
+    return *this;
+}
+
+TupperWare& TupperWare::set_helicity_reweighting(bool helicity_reweighting) {
+    _helicity_reweighting = helicity_reweighting;
+    return *this;
+}
+
+TupperWare& TupperWare::set_use_helicity(bool use_helicity) {
+    return set_helicity_reweighting(use_helicity);
+}
+
+TupperWare& TupperWare::set_incoming_swapped(IncomingSwapped incoming_swapped) {
+    _incoming_swapped = std::move(incoming_swapped);
+    return *this;
+}
+
+TupperWare& TupperWare::set_event_checker(REX::eventBelongs checker) {
+    _event_checker = std::make_shared<REX::eventBelongs>(std::move(checker));
+    _event_checker_fn = nullptr;
+    return *this;
+}
+
+TupperWare& TupperWare::set_event_checker(std::shared_ptr<REX::eventBelongs> checker) {
+    _event_checker = std::move(checker);
+    _event_checker_fn = nullptr;
+    return *this;
+}
+
+TupperWare& TupperWare::set_event_checker(REX::event_bool_fn checker) {
+    _event_checker = nullptr;
+    _event_checker_fn = std::move(checker);
+    return *this;
+}
+
+std::shared_ptr<REX::tea::procReweightor> TupperWare::build(
+    std::size_t n_iterations, bool reload_per_batch, std::size_t max_deferred_iterations
+) {
+    if (_apis.empty()) {
+        throw std::runtime_error("TupperWare::build: no matrix elements registered");
+    }
+    if (!_event_checker && !_event_checker_fn) {
+        throw std::runtime_error("TupperWare::build: no event checker set");
+    }
+
+    auto result_keys_for = [this]() -> std::vector<UmamiOutputKey> {
+        return _result_keys.empty() ? std::vector<UmamiOutputKey>{UMAMI_OUT_MATRIX_ELEMENT}
+                                     : _result_keys;
+    };
+
+    // Each api is routed to make_weightor() (cpu) or make_deferred_weightor()
+    // (cuda/hip) by its own device_type() and registered onto proc_rwgt's
+    // single, unified reweight_functions list via add_reweight_function()/
+    // add_deferred_reweight_function(), in registration order -- so amp
+    // slot i (see procReweightor::evaluate(amp)) ends up serviced by
+    // whichever of the two an api resolves to. This is what lets a single
+    // TupperWare mix cpu and off-cpu apis correctly; unlike the old
+    // dual-array procReweightor, there's no longer a separate list per
+    // case to build up and hand over as a batch, so each api is added to
+    // proc_rwgt directly as it's resolved rather than staged into a
+    // same-sized rwgts/deferred_rwgts pair first.
+    _weightors.clear();
+    _weightors.reserve(_apis.size());
+    _deferred_weightors.clear();
+    _deferred_weightors.reserve(_apis.size());
+    auto proc_rwgt = std::make_shared<REX::tea::procReweightor>();
+    // Helicity reweighting needs, besides a matrix element accepting
+    // UMAMI_IN_HELICITY_INDEX, an event checker that can assign each event
+    // its index in the matrix element's helicity table (helicity_index, as
+    // Driver's event checkers do from subprocesses.json): without one there
+    // is no index to pass, so the auto-derived keys sum over helicities.
+    bool helicity_reweighting = _helicity_reweighting;
+    bool auto_keys = (_normaliser && _normaliser_input_keys.empty()) ||
+                     std::any_of(_input_keys_list.begin(), _input_keys_list.end(),
+                                 [](auto& keys) { return keys.empty(); });
+    if (helicity_reweighting && auto_keys && !(_event_checker && _event_checker->helicity_index)) {
+        warn_once(
+            "tupperware-no-helicity-index",
+            "helicity reweighting is on, but a TupperWare's event checker assigns no "
+            "helicity indices (REX::eventBelongs::helicity_index), so its events "
+            "cannot be mapped onto the matrix element's helicities; using "
+            "helicity-summed matrix elements for it"
+        );
+        helicity_reweighting = false;
+    }
+    for (std::size_t i = 0; i < _apis.size(); ++i) {
+        const auto& api = *_apis[i];
+        auto input_keys = _input_keys_list[i].empty() ? default_input_keys(api, helicity_reweighting)
+                                                        : _input_keys_list[i];
+        auto output_keys =
+            _output_keys_list[i].empty() ? default_output_keys(api) : _output_keys_list[i];
+        const ReloadSpec* reload =
+            (reload_per_batch && _reload_specs[i]) ? &*_reload_specs[i] : nullptr;
+        if (api.device_type() == madspace::DeviceType::cpu) {
+            auto weightor_ptr = reload
+                ? make_weightor(api, input_keys, output_keys, result_keys_for(), *reload, _incoming_swapped)
+                : make_weightor(api, input_keys, output_keys, result_keys_for(), _incoming_swapped);
+            _weightors.push_back(weightor_ptr);
+            proc_rwgt->add_reweight_function(*weightor_ptr);
+        } else {
+            auto deferred_ptr = reload
+                ? make_deferred_weightor(
+                      api, input_keys, output_keys, result_keys_for(), n_iterations, *reload,
+                      _incoming_swapped
+                  )
+                : make_deferred_weightor(
+                      api, input_keys, output_keys, result_keys_for(), n_iterations,
+                      _incoming_swapped
+                  );
+            _deferred_weightors.push_back(deferred_ptr);
+            proc_rwgt->add_deferred_reweight_function(*deferred_ptr);
+        }
+    }
+
+    if (_event_checker) {
+        proc_rwgt->event_checker = _event_checker;
+        proc_rwgt->event_checker_fn = _event_checker->get_event_bool();
+    } else {
+        proc_rwgt->set_event_checker(_event_checker_fn);
+    }
+
+    if (_normaliser) {
+        auto input_keys = _normaliser_input_keys.empty()
+            ? default_input_keys(*_normaliser, helicity_reweighting)
+            : _normaliser_input_keys;
+        auto output_keys = _normaliser_output_keys.empty() ? default_output_keys(*_normaliser)
+                                                             : _normaliser_output_keys;
+        const ReloadSpec* reload =
+            (reload_per_batch && _normaliser_reload) ? &*_normaliser_reload : nullptr;
+        auto norm_weightor = reload
+            ? make_weightor(*_normaliser, input_keys, output_keys, result_keys_for(), *reload, _incoming_swapped)
+            : make_weightor(*_normaliser, input_keys, output_keys, result_keys_for(), _incoming_swapped);
+        bool helicity_summed = std::find(input_keys.begin(), input_keys.end(),
+                                         UMAMI_IN_HELICITY_INDEX) == input_keys.end();
+        if (!_stored_normalisation.empty() && helicity_summed) {
+            proc_rwgt->set_normaliser(stored_normaliser(_stored_normalisation, *norm_weightor));
+        } else {
+            proc_rwgt->set_normaliser(*norm_weightor);
+        }
+    }
+
+    proc_rwgt->max_deferred_iterations = max_deferred_iterations;
+
+    return proc_rwgt;
+}
+
+TupperWare& WareHouse::add_process() {
+    _processes.emplace_back();
+    return _processes.back();
+}
+
+WareHouse& WareHouse::set_iterators(std::vector<REX::tea::iterator> iters) {
+    _iterators = std::move(iters);
+    return *this;
+}
+
+WareHouse& WareHouse::add_iterator(REX::tea::iterator iter) {
+    _iterators.push_back(std::move(iter));
+    return *this;
+}
+
+WareHouse& WareHouse::set_initialise(REX::tea::iterator init) {
+    _initialise = std::move(init);
+    return *this;
+}
+
+WareHouse& WareHouse::set_finalise(REX::tea::iterator fin) {
+    _finalise = std::move(fin);
+    return *this;
+}
+
+WareHouse& WareHouse::set_launch_names(std::vector<std::string> names) {
+    _launch_names = std::move(names);
+    return *this;
+}
+
+WareHouse& WareHouse::add_launch_name(const std::string& name) {
+    _launch_names.push_back(name);
+    return *this;
+}
+
+const std::vector<std::string>& WareHouse::launch_names() const { return _launch_names; }
+
+WareHouse& WareHouse::set_max_batch(std::size_t max_batch) {
+    _max_batch = max_batch;
+    return *this;
+}
+
+std::size_t WareHouse::max_batch() const { return _max_batch; }
+
+WareHouse& WareHouse::set_threads(std::size_t threads) {
+    _threads = threads;
+    return *this;
+}
+
+std::size_t WareHouse::threads() const { return _threads; }
+
+WareHouse& WareHouse::set_min_batch(std::size_t min_batch) {
+    _min_batch = min_batch;
+    return *this;
+}
+
+std::size_t WareHouse::min_batch() const { return _min_batch; }
+
+std::size_t WareHouse::batch_size_for(std::size_t n_events) const {
+    if (_max_batch != REX::npos) {
+        return _max_batch;
+    }
+    if (_min_batch == REX::npos || _min_batch == 0) {
+        return REX::npos;
+    }
+    std::size_t threads = _threads ? _threads : std::thread::hardware_concurrency();
+    if (threads <= 1 || n_events < 2 * _min_batch) {
+        return REX::npos;
+    }
+    return std::max(_min_batch, (n_events + threads - 1) / threads);
+}
+
+WareHouse& WareHouse::set_max_deferred_iterations(std::size_t max_deferred_iterations) {
+    _max_deferred_iterations = max_deferred_iterations;
+    return *this;
+}
+
+std::size_t WareHouse::max_deferred_iterations() const { return _max_deferred_iterations; }
+
+REX::verbosity& WareHouse::verbose() { return _verbose; }
+
+const REX::verbosity& WareHouse::verbose() const { return _verbose; }
+
+WareHouse& WareHouse::set_verbose(bool v) {
+    _verbose.set_verbose(v);
+    return *this;
+}
+
+bool WareHouse::is_verbose() const { return _verbose.is_verbose(); }
+
+namespace {
+
+std::vector<std::shared_ptr<REX::tea::procReweightor>> build_proc_reweightors(
+    std::deque<TupperWare>& processes,
+    std::size_t n_iterations,
+    bool reload_per_batch,
+    std::size_t max_deferred_iterations
+) {
+    if (processes.empty()) {
+        throw std::runtime_error("WareHouse::build: no processes registered");
+    }
+    std::vector<std::shared_ptr<REX::tea::procReweightor>> result;
+    result.reserve(processes.size());
+    for (auto& process : processes) {
+        result.push_back(process.build(n_iterations, reload_per_batch, max_deferred_iterations));
+    }
+    return result;
+}
+
+} // namespace
+
+REX::tea::reweightor& WareHouse::build(REX::lhe&& mother) {
+    // REX::tea::reweightor defaults to a single no-op iteration
+    // ({true_function}) when set_iterators() is never called -- mirror
+    // that here so a deferred/cuda-hip api's device output buffer is sized
+    // for the same number of calls it will actually receive.
+    std::size_t n_iterations = _iterators.empty() ? 1 : _iterators.size();
+    // An explicit max_batch, or the split across threads of set_min_batch()
+    // (a streaming lhe has no events yet: its batches are its own)
+    mother.max_batch = batch_size_for(mother.events.size());
+    // Once max_batch is actually set, every registered api with a
+    // ReloadSpec reloads its own instance per batch instead of every
+    // overflow clone sharing (and serializing calls through) the one
+    // instance it was registered with; see TupperWare::build().
+    bool reload_per_batch = mother.max_batch != REX::npos;
+    auto proc_rwgts = build_proc_reweightors(
+        _processes, n_iterations, reload_per_batch, _max_deferred_iterations
+    );
+    _reweightor =
+        std::make_unique<REX::tea::reweightor>(std::move(mother), std::move(proc_rwgts));
+    _reweightor->pool_threads = _threads;
+    if (!_iterators.empty()) {
+        _reweightor->set_iterators(_iterators);
+    }
+    if (_initialise) {
+        _reweightor->set_initialise(_initialise);
+    }
+    if (_finalise) {
+        _reweightor->set_finalise(_finalise);
+    }
+    if (!_launch_names.empty()) {
+        _reweightor->set_launch_names(_launch_names);
+    }
+    // Muted by default (see _verbose's member initializer); linking here
+    // shares the flag both ways, so verbose().set_verbose(true) unmutes the
+    // reweightor (and, via reweightor::extract_sorter(), every
+    // procReweightor/process batch it runs) after the fact too.
+    _reweightor->verbose.link(_verbose);
+    return *_reweightor;
+}
+
+REX::tea::reweightor& WareHouse::build(const REX::lhe& mother) {
+    return build(REX::lhe(mother));
+}
+
+REX::tea::reweightor& WareHouse::get() {
+    if (!_reweightor) {
+        throw std::runtime_error("WareHouse::get: reweightor has not been built yet");
+    }
+    return *_reweightor;
+}
+
+const REX::tea::reweightor& WareHouse::get() const {
+    if (!_reweightor) {
+        throw std::runtime_error("WareHouse::get: reweightor has not been built yet");
+    }
+    return *_reweightor;
+}
+
+bool WareHouse::built() const noexcept { return static_cast<bool>(_reweightor); }
+
+TupperWare& WareHouse::process(std::size_t index) {
+    if (index >= _processes.size()) {
+        throw std::out_of_range("WareHouse::process: index out of range");
+    }
+    return _processes[index];
+}
+
+const TupperWare& WareHouse::process(std::size_t index) const {
+    if (index >= _processes.size()) {
+        throw std::out_of_range("WareHouse::process: index out of range");
+    }
+    return _processes[index];
+}
+
+std::deque<TupperWare>& WareHouse::processes() {
+    return _processes;
+}
+
+const std::deque<TupperWare>& WareHouse::processes() const {
+    return _processes;
+}
+
+WareHouse& WareHouse::set_processes(std::vector<TupperWare> processes) {
+    _processes.assign(std::make_move_iterator(processes.begin()),
+                      std::make_move_iterator(processes.end()));
+    return *this;
+}
+
+WareHouse& WareHouse::add_process(TupperWare process) {
+    _processes.push_back(std::move(process));
+    return *this;
+}
+
+WareHouse& WareHouse::add_process(std::vector<TupperWare> processes) {
+    _processes.insert(_processes.end(),
+                      std::make_move_iterator(processes.begin()),
+                      std::make_move_iterator(processes.end()));
+    return *this;
+}
+
+} // namespace madtrex
diff --git a/madtrex/tools/check_helicity_map.py b/madtrex/tools/check_helicity_map.py
new file mode 100755
index 0000000000..1a0c2f822d
--- /dev/null
+++ b/madtrex/tools/check_helicity_map.py
@@ -0,0 +1,136 @@
+#!/usr/bin/env python3
+"""Checks MadtRex's per-event helicity map against a matrix element library.
+
+Independent of MadtRex: talks to the MadMatrix libraries of an MG7 process
+directory directly through the UMAMI C ABI (ctypes). For each LHE event it
+does what MadtRex does to build the helicity index -- match the event to a
+flavor channel positionally, put a beam-swapped event of a mirrored channel
+back into the library's leg order, look its SPINUP values up in the
+subprocesses.json helicity table -- and then checks, per subprocess, that
+
+- the per-helicity |M_h|^2 (UMAMI_IN_HELICITY_INDEX) of all helicities sum to
+  the helicity-summed |M|^2;
+- no event's own helicity has |M_h|^2 = 0;
+- the events' helicities are distributed as sum over events of |M_h|^2/|M|^2
+  (chi2), as they are when MG7 drew them -- a wrong map fails this.
+
+With --set NAME VALUE --ratios FILE it also writes, per event (in file
+order), the expected reweighting ratios after that parameter change, with
+the event's own helicity and helicity-summed, to compare with MadtRex's
+weights.
+
+Usage: check_helicity_map.py PROC_DIR EVENTS.lhe[.gz] [--n N]
+"""
+import argparse, ctypes, glob, gzip, json, math, os, re, sys
+
+p = argparse.ArgumentParser()
+p.add_argument("proc"); p.add_argument("events"); p.add_argument("--n", type=int, default=400)
+p.add_argument("--set", nargs=2, metavar=("NAME", "VALUE")); p.add_argument("--ratios")
+p.add_argument("--no-swap", action="store_true", help="emulate the old behaviour: momenta not swapped back")
+a = p.parse_args()
+
+IN_MOMENTA, IN_ALPHA_S, IN_FLAVOR, IN_HEL = 0, 1, 2, 6
+OUT_ME = 0
+subs = json.load(open(os.path.join(a.proc, "SubProcesses", "subprocesses.json")))
+
+def open_lhe(path):
+    return gzip.open(path, "rt") if path.endswith(".gz") else open(path)
+events = []
+for ev in re.findall(r"<event>(.*?)<", open_lhe(a.events).read(), re.S):
+    lines = [l.split() for l in ev.strip().splitlines()]
+    head, parts = lines[0], [l for l in lines[1:] if len(l) == 13]
+    inc = [q for q in parts if q[1] == "-1"]; out = [q for q in parts if q[1] == "1"]
+    legs = inc + out
+    events.append(dict(
+        pdgs=[int(q[0]) for q in legs],
+        mom=[(float(q[9]), float(q[6]), float(q[7]), float(q[8])) for q in legs],  # E px py pz
+        spin=[float(q[12]) for q in legs], alphas=float(head[5])))
+    if len(events) >= a.n: break
+
+libs = {}
+def library(sub):
+    if sub["me_path"] in libs: return libs[sub["me_path"]]
+    pattern = os.path.join(a.proc, sub["me_path"].replace("{device}", "*").replace("{fptype}", "*"))
+    path = sorted(glob.glob(pattern))[0]
+    for common in glob.glob(os.path.join(os.path.dirname(path), "libmadmatrix_common*")):
+        ctypes.CDLL(common, mode=ctypes.RTLD_GLOBAL)
+    lib = ctypes.CDLL(path, mode=ctypes.RTLD_GLOBAL)
+    h = ctypes.c_void_p()
+    assert lib.umami_initialize(ctypes.byref(h), os.path.join(a.proc, "Cards", "param_card.dat").encode()) == 0
+    libs[sub["me_path"]] = (lib, h, path)
+    return libs[sub["me_path"]]
+
+def me(lib, h, evs, flav, hel=None):
+    n = len(evs); npar = len(evs[0]["mom"])
+    mom = (ctypes.c_double * (n * npar * 4))()
+    for i, e in enumerate(evs):
+        for ip, pv in enumerate(e["mom"]):
+            for mu in range(4): mom[n * (npar * mu + ip) + i] = pv[mu]
+    als = (ctypes.c_double * n)(*[e["alphas"] for e in evs])
+    fl = (ctypes.c_uint * n)(*flav)
+    keys = [IN_MOMENTA, IN_ALPHA_S, IN_FLAVOR]; ins = [ctypes.cast(mom, ctypes.c_void_p), ctypes.cast(als, ctypes.c_void_p), ctypes.cast(fl, ctypes.c_void_p)]
+    if hel is not None:
+        hl = (ctypes.c_int * n)(*hel); keys.append(IN_HEL); ins.append(ctypes.cast(hl, ctypes.c_void_p))
+    out = (ctypes.c_double * n)()
+    k = (ctypes.c_int * len(keys))(*keys); i_ = (ctypes.c_void_p * len(ins))(*ins)
+    ok = (ctypes.c_int * 1)(OUT_ME); o_ = (ctypes.c_void_p * 1)(ctypes.cast(out, ctypes.c_void_p))
+    st = lib.umami_matrix_element(h, ctypes.c_size_t(n), ctypes.c_size_t(n), ctypes.c_size_t(0), ctypes.c_size_t(len(keys)), k, i_, ctypes.c_size_t(1), ok, o_)
+    assert st == 0, st
+    return list(out)
+
+# match events to subprocesses / flavor channels, positionally as MadtRex does
+groups = {}
+for idx, e in enumerate(events):
+    for si, sub in enumerate(subs):
+        for f in sub["flavors"]:
+            for opt in f["options"]:
+                swapped = [opt[1], opt[0]] + opt[2:]
+                if e["pdgs"] == opt: m = False
+                elif f.get("mirror") and e["pdgs"] == swapped: m = True
+                else: continue
+                ee = dict(e)
+                if m and not a.no_swap:
+                    ee["mom"] = [e["mom"][1], e["mom"][0]] + e["mom"][2:]
+                spins = [e["spin"][1], e["spin"][0]] + e["spin"][2:] if m else list(e["spin"])
+                ee["hel"] = sub["helicities"].index([int(round(s)) for s in spins])
+                ee["mirrored"] = m; ee["flav"] = f["index"]; ee["idx"] = idx
+                groups.setdefault(si, []).append(ee); break
+            else: continue
+            break
+        else: continue
+        break
+
+worst = 0.0; zero = 0; ratios = {}
+for si, evs in groups.items():
+    sub = subs[si]; lib, h, path = library(sub); ncomb = len(sub["helicities"])
+    flav = [e["flav"] for e in evs]
+    summed = me(lib, h, evs, flav)
+    per = [me(lib, h, evs, flav, [ih] * len(evs)) for ih in range(ncomb)]
+    own = me(lib, h, evs, flav, [e["hel"] for e in evs])
+    for i, e in enumerate(evs):
+        s = sum(per[ih][i] for ih in range(ncomb))
+        worst = max(worst, abs(s - summed[i]) / summed[i])
+        if own[i] <= 0: zero += 1
+    expected = [sum(per[ih][i] / summed[i] for i in range(len(evs))) for ih in range(ncomb)]
+    observed = [sum(1 for e in evs if e["hel"] == ih) for ih in range(ncomb)]
+    chi2 = sum((o - x) ** 2 / x for o, x in zip(observed, expected) if x > 0.5)
+    ndf = sum(1 for x in expected if x > 0.5) - 1
+    print(f"  helicity frequencies: chi2/ndf = {chi2:.1f}/{ndf}   (observed vs sum of |M_h|^2/|M|^2 over events)")
+    print("  observed:", observed); print("  expected:", [round(x, 1) for x in expected])
+    if any(e["mirrored"] for e in evs):
+        sw = [e for e in evs if e["mirrored"]]
+        unswapped = [dict(e, mom=[e["mom"][1], e["mom"][0]] + e["mom"][2:]) for e in sw]
+        s1 = me(lib, h, sw, [e["flav"] for e in sw]); s0 = me(lib, h, unswapped, [e["flav"] for e in sw])
+        print(f"  beam-swapped events: max rel. diff of summed |M|^2 with/without the swap back = {max(abs(x - y) / x for x, y in zip(s1, s0)):.2e}")
+    if a.set:
+        lib.umami_set_parameter.argtypes = [ctypes.c_void_p, ctypes.c_char_p, ctypes.c_double, ctypes.c_double]
+        assert lib.umami_set_parameter(h, a.set[0].encode(), float(a.set[1]), 0.0) == 0
+        own2 = me(lib, h, evs, flav, [e["hel"] for e in evs]); summed2 = me(lib, h, evs, flav)
+        for i, e in enumerate(evs):
+            ratios[e["idx"]] = (own2[i] / own[i], summed2[i] / summed[i])
+    print(f"{os.path.basename(path)}: {len(evs)} events ({sum(e['mirrored'] for e in evs)} beam-swapped), {ncomb} helicities")
+
+print(f"max |sum_h |M_h|^2 - summed| / summed = {worst:.2e}")
+print(f"events whose own (mapped) helicity has |M_h|^2 = 0: {zero}")
+if a.ratios:
+    json.dump(ratios, open(a.ratios, "w"))
diff --git a/tests/acceptance_tests/test_cmd.py b/tests/acceptance_tests/test_cmd.py
index c9cde7b309..e0147705fa 100755
--- a/tests/acceptance_tests/test_cmd.py
+++ b/tests/acceptance_tests/test_cmd.py
@@ -1472,7 +1472,9 @@ def test_standalone_interference_helicity_choice(self):
         # process's symbols into every library loaded later in this (shared)
         # test process.
         libdir = pjoin(self.out_dir, 'lib')
-        lib = ctypes.CDLL(glob.glob(pjoin(libdir, 'libmadmatrix_P*.so'))[0])
+        # libmadmatrix_<model>_<Pdir>_<backend>_<fptype>.so
+        lib = ctypes.CDLL([path for path in glob.glob(pjoin(libdir, 'libmadmatrix_*_P*.so'))
+                           if 'libmadmatrix_common' not in path][0])
         handle = ctypes.c_void_p()
         card = pjoin(self.out_dir, 'Cards', 'param_card.dat').encode()
         self.assertEqual(0, lib.umami_initialize(ctypes.byref(handle), card))
@@ -1697,8 +1699,8 @@ def evaluate(ks, flavors):
                                   cwd=proc_dir)
             self.assertEqual(0, make.returncode, 'the %s library did not build:\n%s'
                              % (build, make.stdout.decode(errors='replace')[-3000:]))
-            lib = glob.glob(pjoin(self.out_dir, 'lib',
-                                  'libmadmatrix_P*_%s.so' % backend))[0]
+            lib = glob.glob(pjoin(self.out_dir, 'lib', 'libmadmatrix_*_P*_%s_%s.so'
+                                  % (backend, fptype)))[0]
             runs = {}
             for first in ('single', 'batch', 'forward'):
                 out = subprocess.check_output(
diff --git a/tests/acceptance_tests/test_cmd_reweight.py b/tests/acceptance_tests/test_cmd_reweight.py
index 991b24d954..a8b9891ad9 100755
--- a/tests/acceptance_tests/test_cmd_reweight.py
+++ b/tests/acceptance_tests/test_cmd_reweight.py
@@ -451,7 +451,10 @@ def test_nlo_output2(self):
             me_cmd.run_cmd('reweight run_01 --from_cards')
         #solutions = [41.511565, 41.930505, 41.511565, 41.511565, 41.586169, 41.511565, 41.511565, 41.511565, 41.511565, 42.046806, 41.511565, 44.164503, 41.511565, -41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.778368, 41.511565, 42.05448, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 42.115494, 41.511854, 41.511567, 41.511565, 42.00028, 42.120605, 41.514867, -41.511565, 41.511565, 45.125706, 41.511565, 42.180208, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.997509, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.513926, 41.511565, 41.882499, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 47.532736, 41.511565, 41.809063, 41.511565, 41.511565, 41.511565, 41.927695, 41.511565, 41.511565, 41.555692, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, -41.511565, 41.511565, 42.029675, 41.725129, 41.511565, 41.511565, 41.511778, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 42.050439, 41.511565, 41.511565, 41.511565, -41.511565, 42.071105, 41.511565, 41.511565]
         #solutions =  [41.511565, 41.75164, 41.511565, 41.511565, 41.557009, 41.511565, 41.511565, 41.511565, 41.511565, 41.759881, 41.511565, 44.717244, 41.511565, -41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 42.041827, 41.511565, 41.663092, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.920917, 41.511736, 41.511566, 41.511565, 205.29791, 42.538389, 41.513573, -41.511565, 41.511565, 44.650974, 41.511565, 42.549187, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.431553, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.513265, 41.511565, 43.902643, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 8853.0397, 41.511565, 333.38491, 41.511565, 41.511565, 41.511565, 75.701362, 41.511565, 41.511565, 41.54695, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, -41.511565, 41.511565, 41.454923, 41.53984, 41.511565, 41.511565, 41.511743, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.732627, 41.511565, 41.511565, 41.511565, -41.511565, 41.395865, 41.511565, 41.511565]
-        solutions = [41.511565, 42.03619, 41.511565, 41.511565, 41.60499, 41.511565, 41.511565, 41.511565, 41.511565, 42.18183, 41.511565, 44.833753, 41.511565, -41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.845674, 41.511565, 42.19144, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 42.267845, 41.511927, 41.511567, 41.511565, 42.123567, 42.274246, 41.515701, -41.511565, 41.511565, 46.037437, 41.511565, 42.348885, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 42.120097, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.514522, 41.511565, 41.976074, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 49.051683, 41.511565, 41.884112, 41.511565, 41.511565, 41.511565, 42.032671, 41.511565, 41.511565, 41.566824, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, -41.511565, 41.511565, 42.160378, 41.779005, 41.511565, 41.511565, 41.511832, 41.511565, 41.511565, 41.511565, 41.511565, 41.511565, 42.186379, 41.511565, 41.511565, 41.511565, -41.511565, 42.212259, 41.511565, 41.511565]
+        # hj_nlo.lhe.gz regenerated in the current NLO event format (2026):
+        # the born-like events reweight by heft's finite top-mass factor
+        # for g g > h (1.065), which the earlier sample did not have
+        solutions = [41.08596, 41.08596, -41.08596, 51.938959, 41.08596, 41.099953, 41.08596, 41.08596, 41.08596, -41.08596, 41.08596, 41.08596, 41.08596, 41.08596, 41.08596, 41.08596, -41.08596, 48.82417, 41.08596, 41.08596, 41.08596, 41.08596, 41.08596, 41.08596, 41.164686, 41.08596, 41.08596, 41.08596, 41.08596, -40.780961, 41.08596, 41.08596, 41.08596, 41.08596, -41.08596, 41.08596, 41.08596, 41.08596, 41.08596, -40.957947, 41.08596, 41.08596, 41.08596, 41.087589, 63.333702, 41.08596, 41.08596, 41.094059, 41.08596, 41.08596, 41.08596, 41.08596, 41.08596, 41.08596, 41.08596, 41.08596, 66.933958, -41.08596, 41.08596, 41.08596, 41.08596, 41.08596, 41.08596, 41.08596, 41.08596, 41.323998, 41.08596, 41.08596, 41.092332, 41.387532, 41.08596, 41.08596, 41.08596, 41.08596, 41.08596, 52.261103, 41.08596, 70.033996, 41.08596, 41.923138, 41.08596, 41.085967, 41.624517, 41.08596, 41.08596, 41.08596, 41.08596, 41.113697, 41.524823, 46.308714, 41.08596, 41.08596, 41.08596, 41.08596, 41.692762, 41.08596, 41.08596, 41.08596, 41.08596, 41.808881]
         lhe = lhe_parser.EventFile(pjoin(self.run_dir,'Events','run_01', 'rwgt_events_tree_rwgt_1.lhe.gz'))
         lhe_orig = lhe_parser.EventFile(pjoin(self.run_dir,'Events','run_01', 'events.lhe.gz'))
         
@@ -516,7 +519,7 @@ def test_loop_improved_reweighting(self):
         #misc.sprint(solutions)
             
     def test_mg7_reweighting_at_generation(self):
-        """reweight=ON on an mg7 output. Its weights average to sigma: the log
+        """reweight=classic on an mg7 output. Its weights average to sigma: the log
         printed sum(weights), N_events times sigma, for the original and the
         new cross sections. The launch runs as an agent drives it, with stdin
         a pipe left open, where the param_card question of each launch block
@@ -557,7 +560,8 @@ def run(name, script, stdin):
             fsock.write("launch --rwgt_name=wz26\nset decay 23 2.6\n"
                         "launch --rwgt_name=wz23\nset decay 23 2.3\n")
         # on the Z pole, where the cross section goes as 1/width^2
-        log = run('launch', "launch %(dir)s\nreweight=ON\nset nevents 1000\n"
+        # the classic module: with madtrex installed, ON (auto) picks MadtRex
+        log = run('launch', "launch %(dir)s\nreweight=classic\nset nevents 1000\n"
                             "set iseed 7\nset beam.e_cm 91.188\n"
                             "set output_format lhe\ndone\n",
                   subprocess.PIPE)
diff --git a/tests/acceptance_tests/test_reweight_backends.py b/tests/acceptance_tests/test_reweight_backends.py
new file mode 100644
index 0000000000..bcd2ebb8ea
--- /dev/null
+++ b/tests/acceptance_tests/test_reweight_backends.py
@@ -0,0 +1,385 @@
+################################################################################
+#
+# Copyright (c) 2026 The MadGraph7 Development team and Contributors
+#
+# This file is a part of the MadGraph7 project, an application which
+# automatically generates Feynman diagrams and matrix elements for arbitrary
+# high-energy processes in the Standard Model and beyond.
+#
+# It is subject to the MadGraph7 license which should accompany this
+# distribution.
+#
+# For more information, visit madgraph.phys.ucl.ac.be and amcatnlo.web.cern.ch
+#
+################################################################################
+"""The two reweighting backends of an MG7 output give the same weights, in
+the same place.
+
+A small MG7 process (u u~ > t t~, 500 events, LHE output, systematics on)
+is generated once; the run is copied, and one copy is reweighted with
+MadtRex, the other with the classic module (launch.run_reweighting, as the
+launch runs them). The card uses the shared vocabulary: 'change helicity
+False' and a parameter given by its param_card name ('set MT 175'). Both
+add their weights to the run's own event file (Events/<run>/events.lhe.gz),
+after the systematics weights, and agree event by event. A second process
+directory with output_format = lhe_npy checks that MadtRex reweights
+events.npy in place (generate_events --reweight=CARD) and that npy_to_lhe
+carries the new weight.
+
+Skipped unless madspace and madtrex are installed and numpy.f2py runs with
+this Python (the classic module compiles its matrix elements with f2py, by
+default the one installed with this Python: misc.default_f2py()).
+"""
+
+import json
+import os
+import shutil
+import subprocess
+import sys
+import tempfile
+import textwrap
+import unittest
+from pathlib import Path
+
+MG_ROOT = Path(__file__).resolve().parents[2]
+
+
+def both_backends_available():
+    if not (MG_ROOT / 'madspace' / 'install' / 'madspace').is_dir():
+        return False
+    if not (MG_ROOT / 'madtrex' / 'install' / 'madtrex').is_dir():
+        return False
+    if shutil.which('gfortran') is None:
+        return False
+    result = subprocess.run([sys.executable, '-m', 'numpy.f2py', '-v'],
+                            capture_output=True, text=True)
+    return result.returncode == 0
+
+
+CARD = "change helicity False\nlaunch --rwgt_name=mt175\nset MT 175\n"
+
+# reweight run <run> of the process directory (the cwd) with <backend>
+REWEIGHT = textwrap.dedent("""
+    import os, sys
+    sys.path.insert(0, {root!r})
+    from madgraph.iolibs.template_files.mg7 import bootstrap
+    bootstrap.ensure_madspace(interactive=False)
+    from madgraph.iolibs.template_files.mg7 import launch
+    class Process:
+        run_path = os.path.join("Events", {run!r})
+        events_path = os.path.join(run_path, "events.lhe.gz")
+        run_card = {{"run": {{"output_format": "lhe"}}}}
+    launch.run_reweighting(Process(), launch.REWEIGHT_CARD, requested={backend!r})
+""")
+
+
+@unittest.skipUnless(both_backends_available(),
+                     'needs madspace, madtrex, gfortran and numpy.f2py')
+class ReweightBackendsAgreeTest(unittest.TestCase):
+
+    def setUp(self):
+        self.tmp = tempfile.mkdtemp()
+
+    def tearDown(self):
+        shutil.rmtree(self.tmp)
+
+    def run_checked(self, args, **kwargs):
+        result = subprocess.run(args, capture_output=True, text=True, **kwargs)
+        self.assertEqual(result.returncode, 0, result.stdout[-3000:] + result.stderr[-3000:])
+        return result
+
+    def make_process(self, name, output_format):
+        proc = os.path.join(self.tmp, name)
+        script = os.path.join(self.tmp, name + '.mg5')
+        with open(script, 'w') as f:
+            f.write('import model sm\ngenerate u u~ > t t~\noutput mg7 %s -f\n' % proc)
+        self.run_checked([sys.executable, str(MG_ROOT / 'bin' / 'madgraph'), script])
+        run_card = os.path.join(proc, 'Cards', 'run_card.toml')
+        with open(run_card) as f:
+            text = f.read()
+        text = text.replace('events = 100000', 'events = 500', 1)
+        text = text.replace('output_format = "lhe_npy"', 'output_format = "%s"' % output_format, 1)
+        with open(run_card, 'w') as f:
+            f.write(text)
+        with open(os.path.join(proc, 'Cards', 'reweight_card.dat'), 'w') as f:
+            f.write(CARD)
+        return proc
+
+    def weights(self, path):
+        sys.path.insert(0, str(MG_ROOT))
+        import madgraph.various.lhe_parser as lhe_parser
+        return [(event.wgt, event.parse_reweight())
+                for event in lhe_parser.EventFile(path)]
+
+    def test_classic_and_madtrex_write_the_same_file(self):
+        proc = self.make_process('PROC', 'lhe')
+        self.run_checked([sys.executable, 'bin/generate_events', '-f'], cwd=proc)
+        events = os.path.join(proc, 'Events')
+        shutil.copytree(os.path.join(events, 'run_01'), os.path.join(events, 'run_02'))
+
+        for run, backend in (('run_01', 'madtrex'), ('run_02', 'classic')):
+            self.run_checked([sys.executable, '-c', REWEIGHT.format(
+                root=str(MG_ROOT), run=run, backend=backend)], cwd=proc)
+            with open(os.path.join(events, run, 'reweighting.json')) as f:
+                record = json.load(f)
+            self.assertEqual(record['backend'], backend)
+            self.assertEqual(record['events'], os.path.join('Events', run, 'events.lhe.gz'))
+            # no separate output file, no stale decompressed copy
+            self.assertNotIn('events_reweighted.lhe', os.listdir(os.path.join(events, run)))
+            self.assertNotIn('events.lhe', os.listdir(os.path.join(events, run)))
+
+        madtrex = self.weights(os.path.join(events, 'run_01', 'events.lhe.gz'))
+        classic = self.weights(os.path.join(events, 'run_02', 'events.lhe.gz'))
+        self.assertEqual(len(madtrex), 500)
+        self.assertEqual(len(classic), 500)
+        for (orig_m, rw_m), (orig_c, rw_c) in zip(madtrex, classic):
+            self.assertAlmostEqual(orig_m, orig_c, delta=1e-6 * abs(orig_c))
+            # the systematics weights are still there, the new one follows them
+            self.assertEqual(list(rw_m)[:109], [str(i) for i in range(1, 110)])
+            self.assertEqual(list(rw_m)[-1], 'mt175')
+            self.assertIn('1', rw_c)
+            self.assertEqual(rw_m['1'], rw_c['1'])
+            # the LHE files carry about 7 significant digits
+            self.assertAlmostEqual(rw_m['mt175'], rw_c['mt175'], delta=1e-6 * abs(rw_c['mt175']))
+        # MT 173 -> 175 changes the weights
+        total = sum(w for w, _ in madtrex)
+        self.assertNotAlmostEqual(sum(r['mt175'] for _, r in madtrex), total,
+                                  delta=1e-3 * abs(total))
+        # the shower & co. read unweighted_events.lhe.gz: the same events
+        tool_copy = os.path.join(events, 'run_02', 'unweighted_events.lhe.gz')
+        self.assertEqual([r['mt175'] for _, r in self.weights(tool_copy)],
+                         [r['mt175'] for _, r in classic])
+
+    def test_madtrex_reweights_npy_events_in_place(self):
+        import numpy as np
+        proc = self.make_process('NPY', 'lhe_npy')
+        card = os.path.join(proc, 'Cards', 'reweight_card.dat')
+        result = self.run_checked([sys.executable, 'bin/generate_events', '-f',
+                                   '--reweight=%s' % card, '--reweight-backend=madtrex'],
+                                  cwd=proc)
+        self.assertIn('Reweighting with madtrex (--reweight-backend)', result.stdout)
+        run = os.path.join(proc, 'Events', 'run_01')
+        events = np.load(os.path.join(run, 'events.npy'), mmap_mode='r')
+        names = [n for n in events.dtype.names if n.startswith('rwgt_')]
+        self.assertEqual(names, ['rwgt_%d' % i for i in range(1, 110)] + ['rwgt_mt175'])
+        # MadtRex's stored generation-time matrix elements are not in the output
+        self.assertNotIn('matrix_element', events.dtype.names)
+        self.assertFalse(any(n.startswith('.madtrex') for n in os.listdir(run)))
+        lhe = self.run_checked([sys.executable, '-c', textwrap.dedent(f"""
+            import sys
+            sys.path.insert(0, {str(MG_ROOT)!r})
+            from madgraph.iolibs.template_files.mg7 import npy_to_lhe
+            print(npy_to_lhe.convert({run!r}, compress=False))
+        """)], cwd=proc).stdout.split()[-1]
+        converted = self.weights(lhe)
+        self.assertEqual(len(converted), 500)
+        for (_, weights), value in zip(converted, events['rwgt_mt175']):
+            self.assertEqual(list(weights)[-1], 'mt175')
+            self.assertAlmostEqual(weights['mt175'], value, delta=1e-6 * abs(value))
+
+    def test_madtrex_lhe_run_is_reweighted_as_npy(self):
+        """An LHE run MadtRex reweights is generated as npy, reweighted, and
+        written as the LHE file it asked for -- also without the weights,
+        when the reweighting fails."""
+        proc = self.make_process('LHE', 'lhe')
+        card = os.path.join(proc, 'Cards', 'reweight_card.dat')
+        result = self.run_checked([sys.executable, 'bin/generate_events', '-f',
+                                   '--reweight=%s' % card, '--reweight-backend=madtrex'],
+                                  cwd=proc)
+        self.assertIn('the events are written as npy', result.stdout)
+        run = os.path.join(proc, 'Events', 'run_01')
+        self.assertIn('events.lhe.gz', os.listdir(run))
+        self.assertNotIn('events.npy', os.listdir(run))
+        self.assertNotIn('header.lhe', os.listdir(run))
+        with open(os.path.join(run, 'reweighting.json')) as f:
+            self.assertEqual(json.load(f)['events'], os.path.join(run, 'events.lhe.gz'))
+        events = self.weights(os.path.join(run, 'events.lhe.gz'))
+        self.assertEqual(len(events), 500)
+        for _, weights in events:
+            self.assertEqual(list(weights), [str(i) for i in range(1, 110)] + ['mt175'])
+        total = sum(w for w, _ in events)
+        self.assertNotAlmostEqual(sum(r['mt175'] for _, r in events), total,
+                                  delta=1e-3 * abs(total))
+
+        bad = os.path.join(self.tmp, 'bad_card.dat')
+        with open(bad, 'w') as f:
+            f.write("launch --rwgt_name=bad\nset MASS 99999 175\n")
+        result = self.run_checked([sys.executable, 'bin/generate_events', '-f',
+                                   '--reweight=%s' % bad, '--reweight-backend=madtrex'],
+                                  cwd=proc)
+        self.assertIn('reweighting failed', result.stdout + result.stderr)
+        run = os.path.join(proc, 'Events', 'run_02')
+        self.assertNotIn('events.npy', os.listdir(run))
+        events = self.weights(os.path.join(run, 'events.lhe.gz'))
+        self.assertEqual(len(events), 500)
+        self.assertEqual(list(events[0][1]), [str(i) for i in range(1, 110)])
+
+
+    def test_madtrex_change_model_divides_by_the_original_matrix_element(self):
+        """A "change model" launch is normalised by the matrix element the
+        events were generated with: to the same model, it reproduces the
+        event weights (no parameter change) and the original model's own
+        reweighting (MT 175). The generation stores its matrix elements for
+        the first, helicity-summed reweighting; the later passes, on the
+        events without them, evaluate the original matrix element (summed
+        and at the event's helicity) and agree."""
+        import numpy as np
+        proc = self.make_process('CM', 'lhe_npy')
+        change = 'change model sm\nchange process u u~ > t t~\n'
+        card = os.path.join(proc, 'Cards', 'reweight_card.dat')
+        with open(card, 'w') as f:
+            f.write('change helicity False\nlaunch --rwgt_name=mt175\nset MT 175\n' + change +
+                    'launch --rwgt_name=same\nlaunch --rwgt_name=cm175\nset MT 175\n')
+        self.run_checked([sys.executable, 'bin/generate_events', '-f',
+                          '--reweight=%s' % card, '--reweight-backend=madtrex'], cwd=proc)
+        events_path = os.path.join(proc, 'Events', 'run_01', 'events.npy')
+        for helicity, suffix in (('False', 'eval'), ('True', 'hel')):
+            again = os.path.join(self.tmp, 'again_%s.dat' % suffix)
+            with open(again, 'w') as f:
+                f.write('change helicity %s\nlaunch --rwgt_name=mt175%s\nset MT 175\n%s'
+                        'launch --rwgt_name=same%s\nlaunch --rwgt_name=cm175%s\nset MT 175\n'
+                        % (helicity, suffix, change, suffix, suffix))
+            self.run_checked([sys.executable, '-c', textwrap.dedent(f"""
+                import sys
+                sys.path.insert(0, {str(MG_ROOT)!r})
+                from madgraph.iolibs.template_files.mg7 import bootstrap
+                bootstrap.ensure_madspace(interactive=False)
+                from madgraph.iolibs.template_files.mg7 import madtrex
+                madtrex.run_parameter_reweighting({again!r}, events_path={events_path!r},
+                                                  process_directory={proc!r})
+            """)], cwd=proc)
+        events = np.load(events_path)
+        self.assertNotIn('matrix_element', events.dtype.names)
+        weight = events['weight']
+
+        def close(a, b):
+            np.testing.assert_allclose(a, b, rtol=1e-12, atol=0)
+        for suffix in ('', 'eval', 'hel'):
+            close(events['rwgt_same' + suffix], weight)
+            close(events['rwgt_cm175' + suffix], events['rwgt_mt175' + suffix])
+        # stored and evaluated normalisation (both helicity-summed) agree
+        close(events['rwgt_mt175eval'], events['rwgt_mt175'])
+        close(events['rwgt_cm175eval'], events['rwgt_cm175'])
+        # and MT 175 does change the weights
+        self.assertGreater(np.max(np.abs(events['rwgt_mt175'] / weight - 1)), 1e-3)
+
+    def test_madtrex_normalises_with_the_generation_param_card(self):
+        """MadtRex reweights from the param_card in the events' header (the
+        one they were generated with), whatever Cards/param_card.dat says
+        by then, and warns about the difference; events without one in
+        their header are normalised with Cards/param_card.dat, with a
+        warning that this cannot be checked."""
+        import numpy as np
+        import re
+        proc = self.make_process('PC', 'lhe_npy')
+        self.run_checked([sys.executable, 'bin/generate_events', '-f'], cwd=proc)
+        run = os.path.join(proc, 'Events', 'run_01')
+        card = os.path.join(self.tmp, 'mt175.dat')
+        with open(card, 'w') as f:
+            f.write(CARD)
+        copies = {}
+        for name, header in (('as_generated', True), ('edited', True), ('no_header', False)):
+            copy = os.path.join(self.tmp, name)
+            os.makedirs(copy)
+            shutil.copy(os.path.join(run, 'events.npy'), copy)
+            if header:
+                shutil.copy(os.path.join(run, 'header.lhe'), copy)
+            copies[name] = os.path.join(copy, 'events.npy')
+
+        def reweight(events):
+            return self.run_checked([sys.executable, '-c', textwrap.dedent(f"""
+                import logging, sys
+                sys.path.insert(0, {str(MG_ROOT)!r})
+                logging.basicConfig(level=logging.WARNING)
+                from madgraph.iolibs.template_files.mg7 import bootstrap
+                bootstrap.ensure_madspace(interactive=False)
+                from madgraph.iolibs.template_files.mg7 import madtrex
+                madtrex.run_parameter_reweighting({card!r}, events_path={events!r},
+                                                  process_directory={proc!r})
+            """)], cwd=proc).stderr
+
+        log = reweight(copies['as_generated'])
+        self.assertNotIn('differs from the param_card', log)
+        param_card = os.path.join(proc, 'Cards', 'param_card.dat')
+        with open(param_card) as f:
+            text = f.read()
+        edited, count = re.subn(r'(?m)^(\s*6\s+)\S+(\s*#\s*MT\b)', r'\g<1>1.800000e+02\2', text)
+        self.assertEqual(count, 1)
+        with open(param_card, 'w') as f:
+            f.write(edited)
+        log = reweight(copies['edited'])
+        self.assertIn('differs from the param_card', log)
+        self.assertIn('mass 6: 173.0 -> 180.0', log)
+        log = reweight(copies['no_header'])
+        self.assertIn('carry no param_card', log)
+
+        weights = {name: np.load(path)['rwgt_mt175'] for name, path in copies.items()}
+        np.testing.assert_allclose(weights['edited'], weights['as_generated'], rtol=1e-12, atol=0)
+        # without the header, normalised with the edited card: MT 180 -> 175
+        self.assertGreater(np.max(np.abs(weights['no_header'] / weights['as_generated'] - 1)), 1e-3)
+
+    def test_madtrex_later_group_is_normalised_at_the_generation_parameters(self):
+        """A group after a launch that changed parameters (here an fptype
+        group, whose normaliser is the first group's library) is normalised
+        at the generation's parameters, also when the sample is split into
+        batches that reload the libraries."""
+        import numpy as np
+        proc = self.make_process('GROUPS', 'lhe_npy')
+        self.run_checked([sys.executable, 'bin/generate_events', '-f'], cwd=proc)
+        events_path = os.path.join(proc, 'Events', 'run_01', 'events.npy')
+        card = os.path.join(self.tmp, 'groups.dat')
+        with open(card, 'w') as f:
+            f.write("change nb_core 4\nchange min_batch 100\n"
+                    "launch --rwgt_name=mt175\nset MT 175\n"
+                    "launch --rwgt_name=same_d\nchange fptype d\n")
+        self.run_checked([sys.executable, '-c', textwrap.dedent(f"""
+            import sys
+            sys.path.insert(0, {str(MG_ROOT)!r})
+            from madgraph.iolibs.template_files.mg7 import bootstrap
+            bootstrap.ensure_madspace(interactive=False)
+            from madgraph.iolibs.template_files.mg7 import madtrex
+            madtrex.run_parameter_reweighting({card!r}, events_path={events_path!r},
+                                              process_directory={proc!r})
+        """)], cwd=proc)
+        events = np.load(events_path)
+        self.assertGreater(np.max(np.abs(events['rwgt_mt175'] / events['weight'] - 1)), 1e-3)
+        np.testing.assert_allclose(events['rwgt_same_d'], events['weight'], rtol=1e-5, atol=0)
+
+
+    def test_madtrex_reweights_mlm_events_with_their_stored_matrix_elements(self):
+        """An MLM run's npy events (with per-particle cluster scales) are
+        read by MadtRex, and their stored matrix elements are at the
+        event's own alpha_s: a no-op helicity-summed launch gives the event
+        weight back."""
+        import numpy as np
+        proc = os.path.join(self.tmp, 'MLM')
+        script = os.path.join(self.tmp, 'MLM.mg5')
+        with open(script, 'w') as f:
+            f.write('import model sm\ngenerate g g > t t~ g\noutput mg7 %s -f\n' % proc)
+        self.run_checked([sys.executable, str(MG_ROOT / 'bin' / 'madgraph'), script])
+        run_card = os.path.join(proc, 'Cards', 'run_card.toml')
+        with open(run_card) as f:
+            text = f.read()
+        for old, new in (('events = 100000', 'events = 2000'),
+                         ('dynamical_scale_choice = "half_transverse_mass"',
+                          'dynamical_scale_choice = "mlm"'),
+                         ('xqcut = 0.0', 'xqcut = 20.0'),
+                         ('[beam]\n', '[beam]\nclustering_history = "diagram"\n')):
+            self.assertIn(old, text)
+            text = text.replace(old, new, 1)
+        with open(run_card, 'w') as f:
+            f.write(text)
+        card = os.path.join(self.tmp, 'mlm.dat')
+        with open(card, 'w') as f:
+            f.write("change helicity False\nlaunch --rwgt_name=same\n")
+        result = self.run_checked([sys.executable, 'bin/generate_events', '-f',
+                                   '--reweight=%s' % card, '--reweight-backend=madtrex'],
+                                  cwd=proc)
+        self.assertNotIn('differ from the ones', result.stdout + result.stderr)
+        events = np.load(os.path.join(proc, 'Events', 'run_01', 'events.npy'))
+        self.assertIn('part1_cluster_scale', events.dtype.names)
+        np.testing.assert_allclose(events['rwgt_same'], events['weight'], rtol=1e-5, atol=0)
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/tests/input_files/IOTestsComparison/IOTestFDGauge/FDgauge_madmatrix/%FD_madmatrix%src%HelAmps_sm.h b/tests/input_files/IOTestsComparison/IOTestFDGauge/FDgauge_madmatrix/%FD_madmatrix%src%HelAmps_sm.h
index 63f0eac4b4..436e64a059 100644
--- a/tests/input_files/IOTestsComparison/IOTestFDGauge/FDgauge_madmatrix/%FD_madmatrix%src%HelAmps_sm.h
+++ b/tests/input_files/IOTestsComparison/IOTestFDGauge/FDgauge_madmatrix/%FD_madmatrix%src%HelAmps_sm.h
@@ -141,7 +141,7 @@ namespace madmatrix
   __host__ __device__ INLINE void
   ixxxxx( const fptype_momenta momenta[], // input: momenta
           const fptype_amp fmass,     // input: fermion mass
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavour
           ALOHAOBJ & fi,          // output: aloha objects
@@ -156,7 +156,7 @@ namespace madmatrix
   __host__ __device__ INLINE void
   ipzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavor index
           ALOHAOBJ & fi,          // output: wavefunctions
@@ -171,7 +171,7 @@ namespace madmatrix
   __host__ __device__ INLINE void
   imzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavor index
           ALOHAOBJ & fi,          // output: wavefunctions
@@ -186,7 +186,7 @@ namespace madmatrix
   __host__ __device__ INLINE void
   ixzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavor index
           ALOHAOBJ & fi,          // output: wavefunctions
@@ -200,7 +200,7 @@ namespace madmatrix
   __host__ __device__ INLINE void
   vxxxxx( const fptype_momenta momenta[], // input: momenta
           const fptype_amp vmass,     // input: vector boson mass
-          const int nhel,         // input: -1, 0 (only if vmass!=0) or +1 (helicity of vector boson)
+          const fptype_amp_sv nhel, // input: -1, 0 (only if vmass!=0) or +1 (helicity of vector boson, per SIMD lane/GPU thread)
           const int nsv,          // input: +1 (final) or -1 (initial)
           const int flv,          // input: flavor index
           ALOHAOBJ & vc,          // output: wavefunctions
@@ -228,7 +228,7 @@ namespace madmatrix
   __host__ __device__ INLINE void
   oxxxxx( const fptype_momenta momenta[], // input: momenta
           const fptype_amp fmass,     // input: fermion mass
-          const int nhel,         // input: -1, 0 (only if vmass!=0) or +1 (helicity of vector boson)
+          const fptype_amp_sv nhel, // input: -1, 0 (only if vmass!=0) or +1 (helicity of vector boson, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavor index
           ALOHAOBJ & fo,          // output: wavefunctions
@@ -243,7 +243,7 @@ namespace madmatrix
   __host__ __device__ INLINE void
   opzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavor index
           ALOHAOBJ & fo,          // output: wavefunctions
@@ -258,7 +258,7 @@ namespace madmatrix
   __host__ __device__ INLINE void
   omzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavor index
           ALOHAOBJ & fo,          // output: wavefunctions
@@ -272,7 +272,7 @@ namespace madmatrix
   __host__ __device__ INLINE void
   oxzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavor index
           ALOHAOBJ & fo,          // output: wavefunctions
@@ -313,7 +313,7 @@ namespace madmatrix
   __host__ __device__ void
   ixxxxx( const fptype_momenta momenta[], // input: momenta
           const fptype_amp fmass,     // input: fermion mass
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavour
           ALOHAOBJ & fi,          // output: wavefunctions
@@ -338,7 +338,12 @@ namespace madmatrix
     fi.pvec[2] = -pvec2_ * static_cast<fptype_momenta>(nsf);
     fi.pvec[3] = -pvec3_ * static_cast<fptype_momenta>(nsf);
     fi.flv_index = flv;
-    const int nh = nhel * nsf;
+    const fptype_amp_sv nh = nhel * static_cast<fptype_amp>( nsf );
+    // BRANCHLESS HELICITY: the helicity is not a branch condition but enters as the truth values
+    // tp (nh == +1) and tm (nh == -1), 1 or 0, which multiply the two alternatives. Different
+    // SIMD lanes or GPU threads can then carry different helicities and still run in lock-step.
+    const fptype_amp_sv tp = ( 1 + nh ) * (fptype_amp)0.5;
+    const fptype_amp_sv tm = 1 - tp;
     if( fmass != 0. )
     {
 #ifndef MGONGPU_CPPSIMD
@@ -347,11 +352,9 @@ namespace madmatrix
       volatile fptype_amp_sv p2 = pvec1 * pvec1 + pvec2 * pvec2 + pvec3 * pvec3; // volatile fixes #736
       const fptype_amp_sv pp = fpmin( pvec0, fpsqrt( p2 ) );
 #endif
-      // In C++ ixxxxx, use a single ip/im numbering that is valid both for pp==0 and pp>0, which have two numbering schemes in Fortran ixxxxx:
-      // for pp==0, Fortran sqm(0:1) has indexes 0,1 as in C++; but for Fortran pp>0, omega(2) has indexes 1,2 and not 0,1
-      // NB: this is only possible in ixxxx, but in oxxxxx two different numbering schemes must be used
-      const int ip = ( 1 + nh ) / 2; // NB: same as in Fortran pp==0, differs from Fortran pp>0, which is (3+nh)/2 because omega(2) has indexes 1,2
-      const int im = ( 1 - nh ) / 2; // NB: same as in Fortran pp==0, differs from Fortran pp>0, which is (3-nh)/2 because omega(2) has indexes 1,2
+      // The Fortran ip/im indices ( 1 + nh ) / 2 and ( 1 - nh ) / 2 (for pp > 0, ( 3 +- nh ) / 2 as omega(2) has
+      // indexes 1,2) are the truth values tp and tm: an array element x[ip] is tm * x[0] + tp * x[1], and
+      // ip * x[ip] is just tp * x[1] (and likewise for im)
 #ifndef MGONGPU_CPPSIMD
       if( pp == 0. )
       {
@@ -359,10 +362,10 @@ namespace madmatrix
         fptype_amp sqm[2] = { fpsqrt( std::abs( fmass ) ), 0. }; // possibility of negative fermion masses
         //sqm[1] = ( fmass < 0. ? -abs( sqm[0] ) : abs( sqm[0] ) ); // AV: why abs here?
         sqm[1] = ( fmass < 0. ? -sqm[0] : sqm[0] ); // AV: removed an abs here
-        w[0] = cxmake( ip * sqm[ip], 0 );
-        w[1] = cxmake( im * nsf * sqm[ip], 0 );
-        w[2] = cxmake( ip * nsf * sqm[im], 0 );
-        w[3] = cxmake( im * sqm[im], 0 );
+        w[0] = cxmake( tp * sqm[1], 0 );
+        w[1] = cxmake( tm * nsf * sqm[0], 0 );
+        w[2] = cxmake( tp * nsf * sqm[0], 0 );
+        w[3] = cxmake( tm * sqm[1], 0 );
       }
       else
       {
@@ -370,30 +373,32 @@ namespace madmatrix
                                fptype_amp( 1 + nsf - ( 1 - nsf ) * nh ) * (fptype_amp)0.5 };
         fptype_amp omega[2] = { fpsqrt( pvec0 + pp ), 0. };
         omega[1] = fmass / omega[0];
-        const fptype_amp sfomega[2] = { sf[0] * omega[ip], sf[1] * omega[im] };
+        const fptype_amp sfomega[2] = { sf[0] * ( tm * omega[0] + tp * omega[1] ), sf[1] * ( tp * omega[0] + tm * omega[1] ) };
         const fptype_amp pp3 = fpmax( fpternary( fpsignbit( pvec3 ), ( pvec1 * pvec1 + pvec2 * pvec2 ) / ( pp - pvec3 ), pp + pvec3 ), 0. );
         const cxtype_amp chi[2] = { cxmake( fpsqrt( pp3 * (fptype_amp)0.5 / pp ), 0. ),
                                 ( pp3 == 0. ? cxmake( -nh, 0. ) : cxmake( nh * pvec1, pvec2 ) / fpsqrt( (fptype_amp)2. * pp * pp3 ) ) };
-        w[0] = sfomega[0] * chi[im];
-        w[1] = sfomega[0] * chi[ip];
-        w[2] = sfomega[1] * chi[im];
-        w[3] = sfomega[1] * chi[ip];
+        const cxtype_amp chi_im = tp * chi[0] + tm * chi[1]; // chi[im]
+        const cxtype_amp chi_ip = tm * chi[0] + tp * chi[1]; // chi[ip]
+        w[0] = sfomega[0] * chi_im;
+        w[1] = sfomega[0] * chi_ip;
+        w[2] = sfomega[1] * chi_im;
+        w[3] = sfomega[1] * chi_ip;
       }
 #else
       // Branch A: pp == 0.
       // NB: Do not use "abs" for floats! It returns an integer with no build warning! Use std::abs!
       fptype_amp sqm[2] = { fpsqrt( std::abs( fmass ) ), 0 }; // possibility of negative fermion masses (NB: SCALAR!)
       sqm[1] = ( fmass < 0 ? -sqm[0] : sqm[0] );          // AV: removed an abs here (as above)
-      const cxtype_amp fiA_2 = ip * sqm[ip];                  // scalar cxtype: real part initialised from fptype, imag part = 0
-      const cxtype_amp fiA_3 = im * nsf * sqm[ip];            // scalar cxtype: real part initialised from fptype, imag part = 0
-      const cxtype_amp fiA_4 = ip * nsf * sqm[im];            // scalar cxtype: real part initialised from fptype, imag part = 0
-      const cxtype_amp fiA_5 = im * sqm[im];                  // scalar cxtype: real part initialised from fptype, imag part = 0
+      const cxtype_amp_v fiA_2 = cxmake( tp * sqm[1], 0 );
+      const cxtype_amp_v fiA_3 = cxmake( tm * (fptype_amp)nsf * sqm[0], 0 );
+      const cxtype_amp_v fiA_4 = cxmake( tp * (fptype_amp)nsf * sqm[0], 0 );
+      const cxtype_amp_v fiA_5 = cxmake( tm * sqm[1], 0 );
       // Branch B: pp != 0.
-      const fptype_amp sf[2] = { fptype_amp( 1 + nsf + ( 1 - nsf ) * nh ) * (fptype_amp)0.5,
-                             fptype_amp( 1 + nsf - ( 1 - nsf ) * nh ) * (fptype_amp)0.5 };
+      const fptype_amp_v sf[2] = { ( (fptype_amp)( 1 + nsf ) + (fptype_amp)( 1 - nsf ) * nh ) * (fptype_amp)0.5,
+                                   ( (fptype_amp)( 1 + nsf ) - (fptype_amp)( 1 - nsf ) * nh ) * (fptype_amp)0.5 };
       fptype_v omega[2] = { fpsqrt( pvec0 + pp ), 0 };
       omega[1] = fmass / omega[0];
-      const fptype_amp_v sfomega[2] = { sf[0] * omega[ip], sf[1] * omega[im] };
+      const fptype_amp_v sfomega[2] = { sf[0] * ( tm * omega[0] + tp * omega[1] ), sf[1] * ( tp * omega[0] + tm * omega[1] ) };
       const fptype_amp_v pp3 = fpmax( fpternary( fpsignbit( pvec3 ), ( pvec1 * pvec1 + pvec2 * pvec2 ) / ( pp - pvec3 ), pp + pvec3 ), 0 );
       volatile fptype_amp_v ppDENOM = fpternary( pp != 0, pp, 1. );    // hack: ppDENOM[ieppV]=1 if pp[ieppV]==0
       volatile fptype_amp_v pp3DENOM = fpternary( pp3 != 0, pp3, 1. ); // hack: pp3DENOM[ieppV]=1 if pp3[ieppV]==0
@@ -401,11 +406,13 @@ namespace madmatrix
       const cxtype_amp_v chi[2] = { cxmake( fpsqrt( chi0r2 ), 0 ),     // hack: dummy[ieppV] is not used if pp[ieppV]==0
                                 cxternary( ( pp3 == 0. ),
                                            cxmake( -nh, 0 ),
-                                           cxmake( (fptype_amp)nh * pvec1, pvec2 ) / fpsqrt( (fptype_amp)2. * ppDENOM * pp3DENOM ) ) }; // hack: dummy[ieppV] is not used if pp[ieppV]==0
-      const cxtype_amp_v fiB_2 = sfomega[0] * chi[im];
-      const cxtype_amp_v fiB_3 = sfomega[0] * chi[ip];
-      const cxtype_amp_v fiB_4 = sfomega[1] * chi[im];
-      const cxtype_amp_v fiB_5 = sfomega[1] * chi[ip];
+                                           cxmake( nh * pvec1, pvec2 ) / fpsqrt( (fptype_amp)2. * ppDENOM * pp3DENOM ) ) }; // hack: dummy[ieppV] is not used if pp[ieppV]==0
+      const cxtype_amp_v chi_im = tp * chi[0] + tm * chi[1]; // chi[im]
+      const cxtype_amp_v chi_ip = tm * chi[0] + tp * chi[1]; // chi[ip]
+      const cxtype_amp_v fiB_2 = sfomega[0] * chi_im;
+      const cxtype_amp_v fiB_3 = sfomega[0] * chi_ip;
+      const cxtype_amp_v fiB_4 = sfomega[1] * chi_im;
+      const cxtype_amp_v fiB_5 = sfomega[1] * chi_ip;
       // Choose between the results from branch A and branch B
       const bool_v mask = ( pp == 0. );
       w[0] = cxternary( mask, fiA_2, fiB_2 );
@@ -431,8 +438,8 @@ namespace madmatrix
       volatile fptype_amp_sv sqp0p3DENOM = fpternary( sqp0p3 != 0, (fptype_sv)sqp0p3, 1. ); // hack: dummy sqp0p3DENOM[ieppV]=1 if sqp0p3[ieppV]==0
       cxtype_sv chi[2] = { cxmake( (fptype_v)sqp0p3, 0. ),
                            cxternary( sqp0p3 == 0,
-                                      cxmake( -(fptype_amp)nhel * fpsqrt( (fptype_amp)2. * pvec0 ), 0. ),
-                                      cxmake( (fptype_amp)nh * pvec1, pvec2 ) / (const fptype_v)sqp0p3DENOM ) }; // hack: dummy[ieppV] is not used if sqp0p3[ieppV]==0
+                                      cxmake( -nhel * fpsqrt( (fptype_amp)2. * pvec0 ), 0. ),
+                                      cxmake( nh * pvec1, pvec2 ) / (const fptype_v)sqp0p3DENOM ) }; // hack: dummy[ieppV] is not used if sqp0p3[ieppV]==0
 #else
       // pvec0+pvec3 is a cancelling difference of two ~|p| numbers for a
       // backward-moving massless fermion, and everything below divides by it.
@@ -445,22 +452,12 @@ namespace madmatrix
                                           fptype_sv{ 0 },
                                           fpsqrt( p0p3stable ) * (fptype_amp)nsf );
       const cxtype_amp_sv chi[2] = { cxmake( sqp0p3, 0. ),
-                                 ( sqp0p3 == 0. ? cxmake( -(fptype_amp)nhel * fpsqrt( (fptype_amp)2. * pvec0 ), 0. ) : cxmake( (fptype_amp)nh * pvec1, pvec2 ) / sqp0p3 ) };
+                                 ( sqp0p3 == 0. ? cxmake( -nhel * fpsqrt( (fptype_amp)2. * pvec0 ), 0. ) : cxmake( nh * pvec1, pvec2 ) / sqp0p3 ) };
 #endif
-      if( nh == 1 )
-      {
-        w[0] = cxzero_sv();
-        w[1] = cxzero_sv();
-        w[2] = chi[0];
-        w[3] = chi[1];
-      }
-      else
-      {
-        w[0] = chi[1];
-        w[1] = chi[0];
-        w[2] = cxzero_sv();
-        w[3] = cxzero_sv();
-      }
+      w[0] = tm * chi[1];
+      w[1] = tm * chi[0];
+      w[2] = tp * chi[0];
+      w[3] = tp * chi[1];
     }
     mgDebug( 1, __FUNCTION__ );
     return;
@@ -474,7 +471,7 @@ namespace madmatrix
   __host__ __device__ void
   ipzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavour
           ALOHAOBJ & fi,          // output: wavefunctions
@@ -489,20 +486,17 @@ namespace madmatrix
     fi.pvec[2] = fptype_momenta_sv{ 0 };
     fi.pvec[3] = -pvec3_ * static_cast<fptype_momenta>(nsf);
     fi.flv_index = flv;
-    const int nh = nhel * nsf;
+    const fptype_amp_sv nh = nhel * static_cast<fptype_amp>( nsf );
+    // BRANCHLESS HELICITY: the helicity is not a branch condition but enters as the truth values
+    // tp (nh == +1) and tm (nh == -1), 1 or 0, which multiply the two alternatives. Different
+    // SIMD lanes or GPU threads can then carry different helicities and still run in lock-step.
+    const fptype_amp_sv tp = ( 1 + nh ) * (fptype_amp)0.5;
+    const fptype_amp_sv tm = 1 - tp;
     const cxtype_amp_sv sqp0p3 = cxmake( fpsqrt( (fptype_amp)2. * pvec3 ) * (fptype_amp)nsf, 0. );
-    w[0] = cxmake( fpamp_of_mom( fi.pvec[1] ), fpamp_of_mom( fi.pvec[2] ) );
-    if( nh == 1 )
-    {
-      w[1] = cxmake( fpamp_of_mom( fi.pvec[1] ), fpamp_of_mom( fi.pvec[2] ) );
-      w[2] = sqp0p3;
-    }
-    else
-    {
-      w[1] = sqp0p3;
-      w[2] = cxmake( fpamp_of_mom( fi.pvec[1] ), fpamp_of_mom( fi.pvec[2] ) );
-    }
-    w[3] = cxmake( fpamp_of_mom( fi.pvec[1] ), fpamp_of_mom( fi.pvec[2] ) );
+    w[0] = cxzero_sv();
+    w[1] = tm * sqp0p3;
+    w[2] = tp * sqp0p3;
+    w[3] = cxzero_sv();
     mgDebug( 1, __FUNCTION__ );
     return;
   }
@@ -515,7 +509,7 @@ namespace madmatrix
   __host__ __device__ void
   imzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavour
           ALOHAOBJ & fi,          // output: wavefunctions
@@ -530,20 +524,17 @@ namespace madmatrix
     fi.pvec[2] = fptype_momenta_sv{ 0 };
     fi.pvec[3] = -pvec3_ * static_cast<fptype_momenta>(nsf);
     fi.flv_index = flv;
-    const int nh = nhel * nsf;
-    const cxtype_amp_sv chi = cxmake( -(fptype_amp)nhel * fpsqrt( -(fptype_amp)2. * pvec3 ), 0. );
+    const fptype_amp_sv nh = nhel * static_cast<fptype_amp>( nsf );
+    // BRANCHLESS HELICITY: the helicity is not a branch condition but enters as the truth values
+    // tp (nh == +1) and tm (nh == -1), 1 or 0, which multiply the two alternatives. Different
+    // SIMD lanes or GPU threads can then carry different helicities and still run in lock-step.
+    const fptype_amp_sv tp = ( 1 + nh ) * (fptype_amp)0.5;
+    const fptype_amp_sv tm = 1 - tp;
+    const cxtype_amp_sv chi = cxmake( -nhel * fpsqrt( -(fptype_amp)2. * pvec3 ), 0. );
+    w[0] = tm * chi;
     w[1] = cxzero_sv();
     w[2] = cxzero_sv();
-    if( nh == 1 )
-    {
-      w[0] = cxzero_sv();
-      w[3] = chi;
-    }
-    else
-    {
-      w[0] = chi;
-      w[3] = cxzero_sv();
-    }
+    w[3] = tp * chi;
     mgDebug( 1, __FUNCTION__ );
     return;
   }
@@ -556,7 +547,7 @@ namespace madmatrix
   __host__ __device__ void
   ixzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavour
           ALOHAOBJ & fi,          // output: wavefunctions
@@ -578,7 +569,12 @@ namespace madmatrix
     fi.pvec[3] = -pvec3_ * static_cast<fptype_momenta>(nsf);
     fi.flv_index = flv;
 
-    const int nh = nhel * nsf;
+    const fptype_amp_sv nh = nhel * static_cast<fptype_amp>( nsf );
+    // BRANCHLESS HELICITY: the helicity is not a branch condition but enters as the truth values
+    // tp (nh == +1) and tm (nh == -1), 1 or 0, which multiply the two alternatives. Different
+    // SIMD lanes or GPU threads can then carry different helicities and still run in lock-step.
+    const fptype_amp_sv tp = ( 1 + nh ) * (fptype_amp)0.5;
+    const fptype_amp_sv tm = 1 - tp;
     //const float sqp0p3 = sqrtf( pvec0 + pvec3 ) * nsf; // AV: why force a float here?
     // pvec0+pvec3 is a cancelling difference of two ~|p| numbers for a
     // backward-moving massless fermion, and everything below divides by it.
@@ -590,21 +586,11 @@ namespace madmatrix
                                             pvec0 + pvec3 );
     const fptype_amp_sv sqp0p3 = fpsqrt( p0p3stable ) * static_cast<fptype_amp>(nsf);
     const cxtype_amp_sv chi0 = cxmake( sqp0p3, 0. );
-    const cxtype_amp_sv chi1 = cxmake( (fptype_amp)nh * pvec1 / sqp0p3, pvec2 / sqp0p3 );
-    if( nh == 1 )
-    {
-      w[0] = cxzero_sv();
-      w[1] = cxzero_sv();
-      w[2] = chi0;
-      w[3] = chi1;
-    }
-    else
-    {
-      w[0] = chi1;
-      w[1] = chi0;
-      w[2] = cxzero_sv();
-      w[3] = cxzero_sv();
-    }
+    const cxtype_amp_sv chi1 = cxmake( nh * pvec1 / sqp0p3, pvec2 / sqp0p3 );
+    w[0] = tm * chi1;
+    w[1] = tm * chi0;
+    w[2] = tp * chi0;
+    w[3] = tp * chi1;
     mgDebug( 1, __FUNCTION__ );
     return;
   }
@@ -616,7 +602,7 @@ namespace madmatrix
   __host__ __device__ void
   vxxxxx( const fptype_momenta momenta[], // input: momenta
           const fptype_amp vmass,     // input: vector boson mass
-          const int nhel,         // input: -1, 0 (only if vmass!=0) or +1 (helicity of vector boson)
+          const fptype_amp_sv nhel, // input: -1, 0 (only if vmass!=0) or +1 (helicity of vector boson, per SIMD lane/GPU thread)
           const int nsv,          // input: +1 (final) or -1 (initial)
           const int flv,          // input: flavour
           ALOHAOBJ & vc,          // output: wavefunctions
@@ -642,7 +628,10 @@ namespace madmatrix
     vc.pvec[3] = pvec3_ * static_cast<fptype_momenta>(nsv);
     vc.flv_index = flv;
     const fptype_amp sqh = fpsqrt( 0.5 ); // AV this is > 0!
-    const fptype_amp hel = nhel;
+    const fptype_amp_sv hel = nhel;
+    // BRANCHLESS HELICITY: |hel| is hel * hel for hel in { -1, 0, +1 }, so that the helicity can
+    // differ from one SIMD lane or GPU thread to the next with no branch and no std::abs
+    const fptype_amp_sv abshel = hel * hel;
 
     // FD gauge
      const cxtype_amp_sv cI = cxmake( 0. + fptype_amp_sv{ 0. },  1. + fptype_amp_sv{ 0. }  );
@@ -653,8 +642,8 @@ namespace madmatrix
 
     if( vmass != 0. )
     {
-      const int nsvahl = nsv * std::abs( hel );
-      const fptype_amp hel0 = 1. - std::abs( hel );
+      const fptype_amp_sv nsvahl = static_cast<fptype_amp>( nsv ) * abshel;
+      const fptype_amp_sv hel0 = 1 - abshel; // truth value of hel == 0 (longitudinal polarisation)
 #ifndef MGONGPU_CPPSIMD
       const fptype_amp_sv pt2 = ( pvec1 * pvec1 ) + ( pvec2 * pvec2 );
       const fptype_amp_sv pp = fpmin( pvec0, fpsqrt( pt2 + ( pvec3 * pvec3 ) ) );
@@ -709,17 +698,14 @@ namespace madmatrix
 
       nk = n[0]*pvec0 - n[1]*pvec1 - n[2]*pvec2 - n[3]*pvec3;
 
-      if ( abs(nhel) == 1)
-      {
-        w[4] = cxzero_sv<cxtype_amp_sv>();
-      }
-      else{
-        w[0] = cxmake( -vmass/nk * n[0], zero );
-        w[1] = cxmake( -vmass/nk * n[1], zero );
-        w[2] = cxmake( -vmass/nk * n[2], zero );
-        w[3] = cxmake( -vmass/nk * n[3], zero );
-        w[4] = -static_cast<fptype_amp>(nsv)*cI;
-      }
+      // BRANCHLESS HELICITY: |hel| = 1 keeps w[0..3] with w[4] = 0, hel = 0 (hel0 = 1) takes the
+      // FD-gauge longitudinal polarisation instead (nk > 0 for a massive momentum, so the unused
+      // alternative is finite and vanishes when multiplied by its zero truth value)
+      w[0] = abshel * w[0] + hel0 * cxmake( -vmass / nk * n[0], zero );
+      w[1] = abshel * w[1] + hel0 * cxmake( -vmass / nk * n[1], zero );
+      w[2] = abshel * w[2] + hel0 * cxmake( -vmass / nk * n[2], zero );
+      w[3] = abshel * w[3] + hel0 * cxmake( -vmass / nk * n[3], zero );
+      w[4] = hel0 * ( -static_cast<fptype_amp>( nsv ) * cI );
 
 #else
       volatile fptype_amp_sv pt2 = ( pvec1 * pvec1 ) + ( pvec2 * pvec2 );
@@ -728,9 +714,9 @@ namespace madmatrix
       const fptype_amp_sv pt = fpmin( pp, fpsqrt( pt2 ) );
       // Branch A: pp == 0.
       const cxtype_amp vcA_2 = cxmake( 0, 0 );
-      const cxtype_amp vcA_3 = cxmake( -hel * sqh, 0 );
-      const cxtype_amp vcA_4 = cxmake( 0, nsvahl * sqh );
-      const cxtype_amp vcA_5 = cxmake( hel0, 0 );
+      const cxtype_amp_v vcA_3 = cxmake( -hel * sqh, 0 );
+      const cxtype_amp_v vcA_4 = cxmake( 0., nsvahl * sqh );
+      const cxtype_amp_v vcA_5 = cxmake( hel0, 0 );
       // Branch B: pp != 0.
       volatile fptype_amp_v ppDENOM = fpternary( pp != 0, pp, 1. ); // hack: ppDENOM[ieppV]=1 if pp[ieppV]==0
       const fptype_amp_v emp = pvec0 / ( vmass * ppDENOM );         // hack: dummy[ieppV] is not used if pp[ieppV]==0
@@ -739,11 +725,11 @@ namespace madmatrix
       // Branch B1: pp != 0. and pt != 0.
       volatile fptype_amp_v ptDENOM = fpternary( pt != 0, pt, 1. );                                                     // hack: ptDENOM[ieppV]=1 if pt[ieppV]==0
       const fptype_amp_v pzpt = pvec3 / ( ppDENOM * ptDENOM ) * sqh * hel;                                              // hack: dummy[ieppV] is not used if pp[ieppV]==0
-      const cxtype_amp_v vcB1_3 = cxmake( hel0 * pvec1 * emp - pvec1 * pzpt, -fpamp_scalar(nsvahl) * pvec2 / ptDENOM * sqh ); // hack: dummy[ieppV] is not used if pt[ieppV]==0
-      const cxtype_amp_v vcB1_4 = cxmake( hel0 * pvec2 * emp - pvec2 * pzpt, fpamp_scalar(nsvahl) * pvec1 / ptDENOM * sqh );  // hack: dummy[ieppV] is not used if pt[ieppV]==0
+      const cxtype_amp_v vcB1_3 = cxmake( hel0 * pvec1 * emp - pvec1 * pzpt, -nsvahl * pvec2 / ptDENOM * sqh ); // hack: dummy[ieppV] is not used if pt[ieppV]==0
+      const cxtype_amp_v vcB1_4 = cxmake( hel0 * pvec2 * emp - pvec2 * pzpt, nsvahl * pvec1 / ptDENOM * sqh );  // hack: dummy[ieppV] is not used if pt[ieppV]==0
       // Branch B2: pp != 0. and pt == 0.
-      const cxtype_amp vcB2_3 = cxmake( -hel * sqh, 0. );
-      const cxtype_amp_v vcB2_4 = cxmake( 0., fpamp_scalar(nsvahl) * fpternary( ( pvec3 < 0 ), -sqh, sqh ) ); // AV: removed an abs here
+      const cxtype_amp_v vcB2_3 = cxmake( -hel * sqh, 0. );
+      const cxtype_amp_v vcB2_4 = cxmake( 0., nsvahl * fpternary( ( pvec3 < 0 ), -sqh, sqh ) ); // AV: removed an abs here
       // Choose between the results from branch A and branch B (and from branch B1 and branch B2)
       const bool_v mask = ( pp == 0. );
       const bool_v maskB = ( pt != 0. );
@@ -765,20 +751,14 @@ namespace madmatrix
 
       nk = n[0]*pvec0 - n[1]*pvec1 - n[2]*pvec2 - n[3]*pvec3;
 
-      // nhel is a scalar: no need for a per-lane mask here (and a bool_v built
-      // from a single value would only set the first lane)
-      if ( abs(nhel) == 1 )
-      {
-        w[4] = zero;
-      }
-      else
-      {
-        w[0] = cxmake( -vmass/nk * n[0], zero );
-        w[1] = cxmake( -vmass/nk * n[1], zero );
-        w[2] = cxmake( -vmass/nk * n[2], zero );
-        w[3] = cxmake( -vmass/nk * n[3], zero );
-        w[4] = -static_cast<fptype_amp>(nsv)*cI;
-      }
+      // BRANCHLESS HELICITY: |hel| = 1 keeps w[0..3] with w[4] = 0, hel = 0 (hel0 = 1) takes the
+      // FD-gauge longitudinal polarisation instead (nk > 0 for a massive momentum, so the unused
+      // alternative is finite and vanishes when multiplied by its zero truth value)
+      w[0] = abshel * w[0] + hel0 * cxmake( -vmass / nk * n[0], zero );
+      w[1] = abshel * w[1] + hel0 * cxmake( -vmass / nk * n[1], zero );
+      w[2] = abshel * w[2] + hel0 * cxmake( -vmass / nk * n[2], zero );
+      w[3] = abshel * w[3] + hel0 * cxmake( -vmass / nk * n[3], zero );
+      w[4] = hel0 * ( -static_cast<fptype_amp>( nsv ) * cI );
 #endif
     }
     else
@@ -810,11 +790,11 @@ namespace madmatrix
       // Branch A: pt != 0.
       volatile fptype_amp_v ptDENOM = fpternary( pt != 0, pt, 1. );                             // hack: ptDENOM[ieppV]=1 if pt[ieppV]==0
       const fptype_amp_v pzpt = pvec3 / ( pp * ptDENOM ) * sqh * hel;                           // hack: dummy[ieppV] is not used if pt[ieppV]==0
-      const cxtype_amp_v vcA_3 = cxmake( -pvec1 * pzpt, -fpamp_scalar(nsv) * pvec2 / ptDENOM * sqh ); // hack: dummy[ieppV] is not used if pt[ieppV]==0
-      const cxtype_amp_v vcA_4 = cxmake( -pvec2 * pzpt, fpamp_scalar(nsv) * pvec1 / ptDENOM * sqh );  // hack: dummy[ieppV] is not used if pt[ieppV]==0
+      const cxtype_amp_v vcA_3 = cxmake( -pvec1 * pzpt, -(fptype_amp)nsv * pvec2 / ptDENOM * sqh ); // hack: dummy[ieppV] is not used if pt[ieppV]==0
+      const cxtype_amp_v vcA_4 = cxmake( -pvec2 * pzpt, (fptype_amp)nsv * pvec1 / ptDENOM * sqh );  // hack: dummy[ieppV] is not used if pt[ieppV]==0
       // Branch B: pt == 0.
-      const cxtype_amp vcB_3 = cxmake( -(fptype_amp)hel * sqh, 0 );
-      const cxtype_amp_v vcB_4 = cxmake( 0, fpamp_scalar(nsv) * fpternary( ( pvec3 < 0 ), -sqh, sqh ) ); // AV: removed an abs here
+      const cxtype_amp_v vcB_3 = cxmake( -hel * sqh, 0 );
+      const cxtype_amp_v vcB_4 = cxmake( 0, (fptype_amp)nsv * fpternary( ( pvec3 < 0 ), -sqh, sqh ) ); // AV: removed an abs here
       // Choose between the results from branch A and branch B
       const bool_v mask = ( pt != 0. );
       w[1] = cxternary( mask, vcA_3, vcB_3 );
@@ -873,7 +853,7 @@ namespace madmatrix
   __host__ __device__ void
   oxxxxx( const fptype_momenta momenta[], // input: momenta
           const fptype_amp fmass,     // input: fermion mass
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           int flv,                // input: flavour
           ALOHAOBJ & fo,          // output: wavefunctions
@@ -899,7 +879,12 @@ namespace madmatrix
     fo.pvec[2] = pvec2_ * static_cast<fptype_momenta>(nsf);
     fo.pvec[3] = pvec3_ * static_cast<fptype_momenta>(nsf);
     fo.flv_index = flv;
-    const int nh = nhel * nsf;
+    const fptype_amp_sv nh = nhel * static_cast<fptype_amp>( nsf );
+    // BRANCHLESS HELICITY: the helicity is not a branch condition but enters as the truth values
+    // tp (nh == +1) and tm (nh == -1), 1 or 0, which multiply the two alternatives. Different
+    // SIMD lanes or GPU threads can then carry different helicities and still run in lock-step.
+    const fptype_amp_sv tp = ( 1 + nh ) * (fptype_amp)0.5;
+    const fptype_amp_sv tm = 1 - tp;
     if( fmass != 0. )
     {
 #ifndef MGONGPU_CPPSIMD
@@ -910,12 +895,12 @@ namespace madmatrix
         fptype_amp sqm[2] = { fpsqrt( std::abs( fmass ) ), 0. }; // possibility of negative fermion masses
         //sqm[1] = ( fmass < 0. ? -abs( sqm[0] ) : abs( sqm[0] ) ); // AV: why abs here?
         sqm[1] = ( fmass < 0. ? -sqm[0] : sqm[0] ); // AV: removed an abs here
-        const int ip = -( ( 1 - nh ) / 2 ) * nhel;  // NB: Fortran sqm(0:1) also has indexes 0,1 as in C++
-        const int im = ( 1 + nh ) / 2 * nhel;       // NB: Fortran sqm(0:1) also has indexes 0,1 as in C++
-        w[0] = cxmake( im * sqm[std::abs( ip )], 0 );
-        w[1] = cxmake( ip * nsf * sqm[std::abs( ip )], 0 );
-        w[2] = cxmake( im * nsf * sqm[std::abs( im )], 0 );
-        w[3] = cxmake( ip * sqm[std::abs( im )], 0 );
+        // The Fortran ip = -( ( 1 - nh ) / 2 ) * nhel and im = ( 1 + nh ) / 2 * nhel are -tm * nhel and
+        // tp * nhel, so that |ip| = tm and |im| = tp: sqm[|ip|] is sqm[1] wherever ip != 0, and so on
+        w[0] = cxmake( tp * nhel * sqm[0], 0 );
+        w[1] = cxmake( -tm * nhel * nsf * sqm[1], 0 );
+        w[2] = cxmake( tp * nhel * nsf * sqm[1], 0 );
+        w[3] = cxmake( -tm * nhel * sqm[0], 0 );
       }
       else
       {
@@ -923,17 +908,18 @@ namespace madmatrix
                                fptype_amp( 1 + nsf - ( 1 - nsf ) * nh ) * (fptype_amp)0.5 };
         fptype_amp omega[2] = { fpsqrt( pvec0 + pp ), 0. };
         omega[1] = fmass / omega[0];
-        const int ip = ( 1 + nh ) / 2; // NB: Fortran is (3+nh)/2 because omega(2) has indexes 1,2 and not 0,1
-        const int im = ( 1 - nh ) / 2; // NB: Fortran is (3-nh)/2 because omega(2) has indexes 1,2 and not 0,1
-        const fptype_amp sfomeg[2] = { sf[0] * omega[ip], sf[1] * omega[im] };
+        // The Fortran ip/im indices ( 3 +- nh ) / 2 (omega(2) has indexes 1,2) are the truth values tp/tm
+        const fptype_amp sfomeg[2] = { sf[0] * ( tm * omega[0] + tp * omega[1] ), sf[1] * ( tp * omega[0] + tm * omega[1] ) };
         const fptype_amp pp3 = fpmax( fpternary( fpsignbit( pvec3 ), ( pvec1 * pvec1 + pvec2 * pvec2 ) / ( pp - pvec3 ), pp + pvec3 ), 0. );
         const cxtype_amp chi[2] = { cxmake( fpsqrt( pp3 * (fptype_amp)0.5 / pp ), 0. ),
                                 ( ( pp3 == 0. ) ? cxmake( -nh, 0. )
                                                 : cxmake( nh * pvec1, -pvec2 ) / fpsqrt( (fptype_amp)2. * pp * pp3 ) ) };
-        w[0] = sfomeg[1] * chi[im];
-        w[1] = sfomeg[1] * chi[ip];
-        w[2] = sfomeg[0] * chi[im];
-        w[3] = sfomeg[0] * chi[ip];
+        const cxtype_amp chi_im = tp * chi[0] + tm * chi[1]; // chi[im]
+        const cxtype_amp chi_ip = tm * chi[0] + tp * chi[1]; // chi[ip]
+        w[0] = sfomeg[1] * chi_im;
+        w[1] = sfomeg[1] * chi_ip;
+        w[2] = sfomeg[0] * chi_im;
+        w[3] = sfomeg[0] * chi_ip;
       }
 #else
       volatile fptype_amp_sv p2 = pvec1 * pvec1 + pvec2 * pvec2 + pvec3 * pvec3; // volatile fixes #736
@@ -942,20 +928,17 @@ namespace madmatrix
       // NB: Do not use "abs" for floats! It returns an integer with no build warning! Use std::abs!
       fptype_amp sqm[2] = { fpsqrt( std::abs( fmass ) ), 0 }; // possibility of negative fermion masses
       sqm[1] = ( fmass < 0 ? -sqm[0] : sqm[0] );          // AV: removed an abs here (as above)
-      const int ipA = -( ( 1 - nh ) / 2 ) * nhel;
-      const int imA = ( 1 + nh ) / 2 * nhel;
-      const cxtype_amp foA_2 = imA * sqm[std::abs( ipA )];
-      const cxtype_amp foA_3 = ipA * nsf * sqm[std::abs( ipA )];
-      const cxtype_amp foA_4 = imA * nsf * sqm[std::abs( imA )];
-      const cxtype_amp foA_5 = ipA * sqm[std::abs( imA )];
+      // (see the scalar code above for the truth values tp/tm in place of the Fortran ip/im indices)
+      const cxtype_amp_v foA_2 = cxmake( tp * nhel * sqm[0], 0 );
+      const cxtype_amp_v foA_3 = cxmake( -tm * nhel * (fptype_amp)nsf * sqm[1], 0 );
+      const cxtype_amp_v foA_4 = cxmake( tp * nhel * (fptype_amp)nsf * sqm[1], 0 );
+      const cxtype_amp_v foA_5 = cxmake( -tm * nhel * sqm[0], 0 );
       // Branch B: pp != 0.
-      const fptype_amp sf[2] = { fptype_amp( 1 + nsf + ( 1 - nsf ) * nh ) * (fptype_amp)0.5,
-                             fptype_amp( 1 + nsf - ( 1 - nsf ) * nh ) * (fptype_amp)0.5 };
+      const fptype_amp_v sf[2] = { ( (fptype_amp)( 1 + nsf ) + (fptype_amp)( 1 - nsf ) * nh ) * (fptype_amp)0.5,
+                                   ( (fptype_amp)( 1 + nsf ) - (fptype_amp)( 1 - nsf ) * nh ) * (fptype_amp)0.5 };
       fptype_v omega[2] = { fpsqrt( pvec0 + pp ), 0 };
       omega[1] = fmass / omega[0];
-      const int ipB = ( 1 + nh ) / 2;
-      const int imB = ( 1 - nh ) / 2;
-      const fptype_amp_v sfomeg[2] = { sf[0] * omega[ipB], sf[1] * omega[imB] };
+      const fptype_amp_v sfomeg[2] = { sf[0] * ( tm * omega[0] + tp * omega[1] ), sf[1] * ( tp * omega[0] + tm * omega[1] ) };
       const fptype_amp_v pp3 = fpmax( fpternary( fpsignbit( pvec3 ), ( pvec1 * pvec1 + pvec2 * pvec2 ) / ( pp - pvec3 ), pp + pvec3 ), 0. );
       volatile fptype_amp_v ppDENOM = fpternary( pp != 0, pp, 1. );    // hack: ppDENOM[ieppV]=1 if pp[ieppV]==0
       volatile fptype_amp_v pp3DENOM = fpternary( pp3 != 0, pp3, 1. ); // hack: pp3DENOM[ieppV]=1 if pp3[ieppV]==0
@@ -963,11 +946,13 @@ namespace madmatrix
       const cxtype_amp_v chi[2] = { cxmake( fpsqrt( chi0r2 ), 0. ),    // hack: dummy[ieppV] is not used if pp[ieppV]==0
                                 ( cxternary( ( pp3 == 0. ),
                                              cxmake( -nh, 0. ),
-                                             cxmake( (fptype_amp)nh * pvec1, -pvec2 ) / fpsqrt( (fptype_amp)2. * ppDENOM * pp3DENOM ) ) ) }; // hack: dummy[ieppV] is not used if pp[ieppV]==0
-      const cxtype_amp_v foB_2 = sfomeg[1] * chi[imB];
-      const cxtype_amp_v foB_3 = sfomeg[1] * chi[ipB];
-      const cxtype_amp_v foB_4 = sfomeg[0] * chi[imB];
-      const cxtype_amp_v foB_5 = sfomeg[0] * chi[ipB];
+                                             cxmake( nh * pvec1, -pvec2 ) / fpsqrt( (fptype_amp)2. * ppDENOM * pp3DENOM ) ) ) }; // hack: dummy[ieppV] is not used if pp[ieppV]==0
+      const cxtype_amp_v chi_im = tp * chi[0] + tm * chi[1]; // chi[im]
+      const cxtype_amp_v chi_ip = tm * chi[0] + tp * chi[1]; // chi[ip]
+      const cxtype_amp_v foB_2 = sfomeg[1] * chi_im;
+      const cxtype_amp_v foB_3 = sfomeg[1] * chi_ip;
+      const cxtype_amp_v foB_4 = sfomeg[0] * chi_im;
+      const cxtype_amp_v foB_5 = sfomeg[0] * chi_ip;
       // Choose between the results from branch A and branch B
       const bool_v mask = ( pp == 0. );
       w[0] = cxternary( mask, foA_2, foB_2 );
@@ -994,7 +979,7 @@ namespace madmatrix
       const cxtype_amp_v chi[2] = { cxmake( (fptype_v)sqp0p3, 0. ),
                                 cxternary( ( sqp0p3 == 0. ),
                                            cxmake( -nhel, 0. ) * fpsqrt( (fptype_amp)2. * pvec0 ),
-                                           cxmake( (fptype_amp)nh * pvec1, -pvec2 ) / (const fptype_amp_sv)sqp0p3DENOM ) }; // hack: dummy[ieppV] is not used if sqp0p3[ieppV]==0
+                                           cxmake( nh * pvec1, -pvec2 ) / (const fptype_amp_sv)sqp0p3DENOM ) }; // hack: dummy[ieppV] is not used if sqp0p3[ieppV]==0
 #else
       // pvec0+pvec3 is a cancelling difference of two ~|p| numbers for a
       // backward-moving massless fermion, and everything below divides by it.
@@ -1007,22 +992,12 @@ namespace madmatrix
                                           fptype_sv{ 0 },
                                           fpsqrt( p0p3stable ) * (fptype_amp)nsf );
       const cxtype_amp_sv chi[2] = { cxmake( sqp0p3, 0. ),
-                                 ( sqp0p3 == 0. ? cxmake( -nhel, 0. ) * fpsqrt( (fptype_amp)2. * pvec0 ) : cxmake( (fptype_amp)nh * pvec1, -pvec2 ) / sqp0p3 ) };
+                                 ( sqp0p3 == 0. ? cxmake( -nhel, 0. ) * fpsqrt( (fptype_amp)2. * pvec0 ) : cxmake( nh * pvec1, -pvec2 ) / sqp0p3 ) };
 #endif
-      if( nh == 1 )
-      {
-        w[0] = chi[0];
-        w[1] = chi[1];
-        w[2] = cxzero_sv();
-        w[3] = cxzero_sv();
-      }
-      else
-      {
-        w[0] = cxzero_sv();
-        w[1] = cxzero_sv();
-        w[2] = chi[1];
-        w[3] = chi[0];
-      }
+      w[0] = tp * chi[0];
+      w[1] = tp * chi[1];
+      w[2] = tm * chi[1];
+      w[3] = tm * chi[0];
     }
     mgDebug( 1, __FUNCTION__ );
     return;
@@ -1036,7 +1011,7 @@ namespace madmatrix
   __host__ __device__ void
   opzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavour
           ALOHAOBJ & fo,          // output: wavefunctions
@@ -1051,20 +1026,17 @@ namespace madmatrix
     fo.pvec[2] = fptype_momenta_sv{ 0 };
     fo.pvec[3] = pvec3_ * static_cast<fptype_momenta>(nsf);
     fo.flv_index = flv;
-    const int nh = nhel * nsf;
+    const fptype_amp_sv nh = nhel * static_cast<fptype_amp>( nsf );
+    // BRANCHLESS HELICITY: the helicity is not a branch condition but enters as the truth values
+    // tp (nh == +1) and tm (nh == -1), 1 or 0, which multiply the two alternatives. Different
+    // SIMD lanes or GPU threads can then carry different helicities and still run in lock-step.
+    const fptype_amp_sv tp = ( 1 + nh ) * (fptype_amp)0.5;
+    const fptype_amp_sv tm = 1 - tp;
     const cxtype_amp_sv csqp0p3 = cxmake( fpsqrt( (fptype_amp)2. * pvec3 ) * (fptype_amp)nsf, 0. );
+    w[0] = tp * csqp0p3;
     w[1] = cxzero_sv();
     w[2] = cxzero_sv();
-    if( nh == 1 )
-    {
-      w[0] = csqp0p3;
-      w[3] = cxzero_sv();
-    }
-    else
-    {
-      w[0] = cxzero_sv();
-      w[3] = csqp0p3;
-    }
+    w[3] = tm * csqp0p3;
     mgDebug( 1, __FUNCTION__ );
     return;
   }
@@ -1077,7 +1049,7 @@ namespace madmatrix
   __host__ __device__ void
   omzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavour
           ALOHAOBJ & fo,          // output: wavefunctions
@@ -1092,23 +1064,17 @@ namespace madmatrix
     fo.pvec[2] = fptype_momenta_sv{ 0 };
     fo.pvec[3] = pvec3_ * static_cast<fptype_momenta>(nsf);
     fo.flv_index = flv;
-    const int nh = nhel * nsf;
+    const fptype_amp_sv nh = nhel * static_cast<fptype_amp>( nsf );
+    // BRANCHLESS HELICITY: the helicity is not a branch condition but enters as the truth values
+    // tp (nh == +1) and tm (nh == -1), 1 or 0, which multiply the two alternatives. Different
+    // SIMD lanes or GPU threads can then carry different helicities and still run in lock-step.
+    const fptype_amp_sv tp = ( 1 + nh ) * (fptype_amp)0.5;
+    const fptype_amp_sv tm = 1 - tp;
     const cxtype_amp_sv chi1 = cxmake( -nhel, 0. ) * fpsqrt( -(fptype_amp)2. * pvec3 );
-    if( nh == 1 )
-    {
-      w[0] = cxzero_sv();
-      w[1] = chi1;
-      w[2] = cxzero_sv();
-      w[3] = cxzero_sv();
-    }
-    else
-    {
-      w[0] = cxzero_sv();
-      w[1] = cxzero_sv();
-      w[2] = chi1;
-      //w[3] = chi1; // AV: BUG!
-      w[3] = cxzero_sv(); // AV: BUG FIX
-    }
+    w[0] = cxzero_sv();
+    w[1] = tp * chi1;
+    w[2] = tm * chi1;
+    w[3] = cxzero_sv();
     mgDebug( 1, __FUNCTION__ );
     return;
   }
@@ -1121,7 +1087,7 @@ namespace madmatrix
   __host__ __device__ void
   oxzxxx( const fptype_momenta momenta[], // input: momenta
           //const fptype_amp fmass,   // [skip: ASSUME fermion mass==0]
-          const int nhel,         // input: -1 or +1 (helicity of fermion)
+          const fptype_amp_sv nhel, // input: -1 or +1 (helicity of fermion, per SIMD lane/GPU thread)
           const int nsf,          // input: +1 (particle) or -1 (antiparticle)
           const int flv,          // input: flavour
           ALOHAOBJ & fo,          // output: wavefunctions
@@ -1143,7 +1109,12 @@ namespace madmatrix
     fo.pvec[2] = pvec2_ * static_cast<fptype_momenta>(nsf);
     fo.pvec[3] = pvec3_ * static_cast<fptype_momenta>(nsf);
     fo.flv_index = flv;
-    const int nh = nhel * nsf;
+    const fptype_amp_sv nh = nhel * static_cast<fptype_amp>( nsf );
+    // BRANCHLESS HELICITY: the helicity is not a branch condition but enters as the truth values
+    // tp (nh == +1) and tm (nh == -1), 1 or 0, which multiply the two alternatives. Different
+    // SIMD lanes or GPU threads can then carry different helicities and still run in lock-step.
+    const fptype_amp_sv tp = ( 1 + nh ) * (fptype_amp)0.5;
+    const fptype_amp_sv tm = 1 - tp;
     //const float sqp0p3 = sqrtf( pvec0 + pvec3 ) * nsf; // AV: why force a float here?
     // pvec0+pvec3 is a cancelling difference of two ~|p| numbers for a
     // backward-moving massless fermion, and everything below divides by it.
@@ -1155,21 +1126,11 @@ namespace madmatrix
                                             pvec0 + pvec3 );
     const fptype_amp_sv sqp0p3 = fpsqrt( p0p3stable ) * static_cast<fptype_amp>(nsf);
     const cxtype_amp_sv chi0 = cxmake( sqp0p3, 0. );
-    const cxtype_amp_sv chi1 = cxmake( (fptype_amp)nh * pvec1 / sqp0p3, -pvec2 / sqp0p3 );
-    if( nh == 1 )
-    {
-      w[0] = chi0;
-      w[1] = chi1;
-      w[2] = cxzero_sv();
-      w[3] = cxzero_sv();
-    }
-    else
-    {
-      w[0] = cxzero_sv();
-      w[1] = cxzero_sv();
-      w[2] = chi1;
-      w[3] = chi0;
-    }
+    const cxtype_amp_sv chi1 = cxmake( nh * pvec1 / sqp0p3, -pvec2 / sqp0p3 );
+    w[0] = tp * chi0;
+    w[1] = tp * chi1;
+    w[2] = tm * chi1;
+    w[3] = tm * chi0;
     mgDebug( 1, __FUNCTION__ );
     return;
   }
diff --git a/tests/input_files/hj_nlo.lhe.gz b/tests/input_files/hj_nlo.lhe.gz
index e7c4e3f8ff..4105e6fb19 100644
Binary files a/tests/input_files/hj_nlo.lhe.gz and b/tests/input_files/hj_nlo.lhe.gz differ
diff --git a/tests/unit_tests/interface/test_mg7_launch.py b/tests/unit_tests/interface/test_mg7_launch.py
index 6e4af33ae2..93f1c37e7f 100644
--- a/tests/unit_tests/interface/test_mg7_launch.py
+++ b/tests/unit_tests/interface/test_mg7_launch.py
@@ -30,6 +30,7 @@
 
 import logging
 import multiprocessing
+import json
 import os
 import shutil
 import subprocess
@@ -39,6 +40,7 @@
 import unittest
 from unittest import mock
 
+import madgraph
 import madgraph.interface.extended_cmd as ext_cmd
 import madgraph.interface.master_interface as mgcmd
 from madgraph.iolibs.template_files.mg7 import bootstrap as mg7_bootstrap
@@ -442,9 +444,19 @@ def setUp(self):
 
     def installer_args(self, line):
         with mock.patch('subprocess.run') as run:
+            run.return_value.returncode = 0
             self.cmd.do_install(line)
         return [str(a) for a in run.call_args[0][0]]
 
+    def test_a_failing_installer_is_an_error(self):
+        """A failed madspace/madtrex install.py (e.g. madtrex without
+        madspace) is reported at the prompt, not passed over silently."""
+        with mock.patch('subprocess.run') as run:
+            run.return_value.returncode = 1
+            with self.assertRaises(madgraph.InvalidCmd) as ctx:
+                self.cmd.do_install('madtrex --yes')
+        self.assertIn('install madtrex failed', str(ctx.exception))
+
     def test_nb_core_is_forwarded_as_the_job_count(self):
         """Without it the installer's cmake build is serial whenever ninja is
         missing, whatever nb_core says."""
@@ -467,6 +479,15 @@ def test_an_explicit_job_count_is_left_alone(self):
                                   or a.startswith('--jobs')]), 1, args)
             self.assertNotIn('5', args)
 
+    def test_madtrex_runs_its_own_installer(self):
+        """`install madtrex` runs madtrex/install.py (not madspace's), with
+        the same job-count forwarding."""
+        self.cmd.do_set('nb_core 5 --no_save')
+        args = self.installer_args('madtrex --yes --clean')
+        self.assertEqual(os.path.basename(os.path.dirname(args[1])), 'madtrex')
+        self.assertTrue(args[1].endswith('install.py'))
+        self.assertEqual(args[2:], ['--yes', '--clean', '-j', '5'])
+
 
 @unittest.skipUnless(mg7_bootstrap.madspace_is_installed(),
                      'madspace is not installed')
@@ -487,17 +508,24 @@ def make_cmd(self):
 
     def test_option_parsing(self):
         cmd = self.make_cmd()
+        none = {'reweight_card': None, 'reweight_backend': None}
         self.assertEqual(cmd._parse_run_options('-f'),
-                         {'force': True, 'name': '', 'laststep': ''})
+                         dict(none, force=True, name='', laststep=''))
         self.assertEqual(cmd._parse_run_options('--name=abc')['name'], 'abc')
         self.assertEqual(cmd._parse_run_options('--laststep=parton')['laststep'],
                          'parton')
         self.assertEqual(
             cmd._parse_run_options('-f --name=abc --laststep=parton'),
-            {'force': True, 'name': 'abc', 'laststep': 'parton'})
+            dict(none, force=True, name='abc', laststep='parton'))
         # space-separated forms, as typed at the MG7> prompt
         self.assertEqual(cmd._parse_run_options('-n abc'),
-                         {'force': False, 'name': 'abc', 'laststep': ''})
+                         dict(none, force=False, name='abc', laststep=''))
+        opts = cmd._parse_run_options('--reweight=rw.dat --reweight-backend madtrex')
+        self.assertEqual((opts['reweight_card'], opts['reweight_backend']),
+                         ('rw.dat', 'madtrex'))
+        opts = cmd._parse_run_options('--reweight rw.dat --reweight-backend=classic')
+        self.assertEqual((opts['reweight_card'], opts['reweight_backend']),
+                         ('rw.dat', 'classic'))
         self.assertEqual(cmd._parse_run_options('-s parton')['laststep'],
                          'parton')
         # a bare word is the run name
@@ -1059,6 +1087,488 @@ def test_it_restores_the_level_after_a_failure(self):
         self.assertIn('back to normal', self.captured.getvalue())
 
 
+@unittest.skipUnless(mg7_bootstrap.madspace_is_installed(),
+                     'madspace is not installed')
+class MG7ReweightingTest(unittest.TestCase):
+    """Reweighting runs once per generation, after the systematics, by the
+    backend prepare_reweighting() chose."""
+
+    CARD = "change helicity False\nlaunch --rwgt_name=mt\nset MT 175\n"
+
+    def setUp(self):
+        from madgraph.iolibs.template_files.mg7 import launch
+        self.launch = launch
+        self.tmpdir = tempfile.TemporaryDirectory()
+        self.cwd = os.getcwd()
+        self.me_dir = make_mg7_dir(os.path.join(self.tmpdir.name, 'PROC'))
+        os.chdir(self.me_dir)
+
+    def tearDown(self):
+        os.chdir(self.cwd)
+        self.tmpdir.cleanup()
+
+    def write_card(self, text, path=os.path.join('Cards', 'reweight_card.dat')):
+        with open(path, 'w') as f:
+            f.write(text)
+        return path
+
+    def prepare(self, switch, cli=None, f2py=True, madtrex=True):
+        from madgraph.various import reweight_backend as rb
+        probe = rb.probe_environment
+
+        def fake_probe(me_dir, options=None, **kwargs):
+            env = probe(me_dir, options, has_f2py=f2py,
+                        madtrex_install_dir=self.tmpdir.name)
+            env.madtrex_installed = madtrex
+            return env
+        with mock.patch.object(rb, 'probe_environment', fake_probe):
+            return self.launch.prepare_reweighting(switch, cli, {'mg5_path': '/mg5'})
+
+    def test_off_or_without_launch_does_nothing(self):
+        self.assertIsNone(self.prepare({'reweight': 'OFF'}))
+        self.assertIsNone(self.prepare({'reweight': 'ON'}))  # no card
+        self.write_card("# launch\n")
+        self.assertIsNone(self.prepare({'reweight': 'ON'}))
+
+    def test_auto_prefers_madtrex(self):
+        self.write_card(self.CARD)
+        decision = self.prepare({'reweight': 'ON'})
+        self.assertEqual((decision.backend, decision.card_kind), ('madtrex', 'param'))
+        self.assertEqual(self.prepare({'reweight': 'ON'}, madtrex=False).backend, 'classic')
+
+    def test_switch_and_command_line(self):
+        self.write_card("change backend madtrex\n" + self.CARD)
+        self.assertEqual(self.prepare({'reweight': 'classic'}).backend, 'classic')
+        self.assertEqual(self.prepare({'reweight': 'classic'}, cli='madtrex').backend,
+                         'madtrex')
+        self.assertEqual(self.prepare({'reweight': 'density'}).card_kind, 'density')
+        decision = self.prepare({'reweight': 'my_plugin'})
+        self.assertEqual((decision.backend, decision.plugin), ('classic', 'my_plugin'))
+
+    def test_impossible_request_fails_before_generating(self):
+        from madgraph.various import reweight_backend as rb
+        self.write_card(self.CARD)
+        with self.assertRaises(rb.ReweightBackendError):
+            self.prepare({'reweight': 'madtrex'}, madtrex=False)
+
+    def test_install_reweight_card_keeps_the_old_one(self):
+        self.write_card("old\n")
+        new = self.write_card(self.CARD, os.path.join(self.tmpdir.name, 'rw.dat'))
+        self.launch.install_reweight_card(new)
+        with open(os.path.join('Cards', 'reweight_card.dat')) as f:
+            self.assertEqual(f.read(), self.CARD)
+        with open(os.path.join('Cards', 'reweight_card.dat.bak')) as f:
+            self.assertEqual(f.read(), "old\n")
+
+    def test_reweight_option_with_switch_on_reweights_once(self):
+        """--reweight=CARD with the reweight switch already on: one run,
+        after the systematics, and no 'reweight' among the tools."""
+        card = self.write_card(self.CARD, os.path.join(self.tmpdir.name, 'rw.dat'))
+        calls = []
+
+        class Process(object):
+            gridpack_run_mode = 'regular'
+            run_path = 'Events/run_01'
+
+        launch = self.launch
+
+        def fake_run_generation(switch, reweight):
+            launch.run_single(switch, reweight)
+
+        def fake_tools(switch, process):
+            calls.append(('tools', tuple(launch.selected_tools(switch))))
+
+        def fake_prepare(switch, cli, options):
+            calls.append(('prepare', switch.get('reweight'), cli))
+            return 'DECISION'
+        with mock.patch.object(launch, 'ask_edit_cards', return_value={'reweight': 'ON'}), \
+                mock.patch.object(launch, 'prepare_reweighting', fake_prepare), \
+                mock.patch.object(launch, 'force_lhe_output_if_needed'), \
+                mock.patch.object(launch, 'run_generation', fake_run_generation), \
+                mock.patch.object(launch, 'compute_auto_widths'), \
+                mock.patch.object(launch, 'MadgraphProcess', lambda: mock.Mock(**{
+                    'gridpack_run_mode': 'regular'})), \
+                mock.patch.object(launch, 'run_lhe_postprocessing',
+                                  lambda p: calls.append(('systematics',))), \
+                mock.patch.object(launch, 'run_reweighting',
+                                  lambda p, c, decision: calls.append(('reweight', decision))), \
+                mock.patch.object(launch, 'run_selected_tools', fake_tools):
+            launch.MG7Cmd(me_dir=self.me_dir).do_generate_events('--reweight=%s' % card)
+        self.assertEqual(calls, [('prepare', 'ON', None), ('systematics',),
+                                 ('reweight', 'DECISION'), ('tools', ())])
+        with open(os.path.join(self.me_dir, 'Cards', 'reweight_card.dat')) as f:
+            self.assertEqual(f.read(), self.CARD)
+
+    def test_reweight_option_switches_reweighting_on(self):
+        card = self.write_card(self.CARD, os.path.join(self.tmpdir.name, 'rw.dat'))
+        seen = []
+        launch = self.launch
+        with mock.patch.object(launch, 'prepare_reweighting',
+                               lambda switch, cli, options: seen.append((switch, cli))), \
+                mock.patch.object(launch, 'force_lhe_output_if_needed'), \
+                mock.patch.object(launch, 'run_generation'):
+            launch.MG7Cmd(me_dir=self.me_dir).do_generate_events(
+                '-f --reweight=%s --reweight-backend=classic' % card)
+        self.assertEqual(seen, [({'reweight': 'ON'}, 'classic')])
+
+    def run_card(self, output_format='lhe', time_of_flight=-1.0, legacy=False,
+                 native=True):
+        return {'run': {'output_format': output_format},
+                'postprocessing': {'time_of_flight': time_of_flight,
+                                   'systematics': legacy},
+                'systematics': {'enable': native}}
+
+    def test_madtrex_lhe_run_is_generated_as_npy(self):
+        """An "lhe" run MadtRex reweights is generated as npy, unless an
+        LHE-only step runs before the reweighting."""
+        from madgraph.various import reweight_backend as rb
+        madtrex, classic = rb.Decision('madtrex', 'param'), rb.Decision('classic', 'param')
+        process = mock.Mock(gridpack_run_mode='regular', run_card=self.run_card())
+        via_binary = self.launch.lhe_via_binary
+        self.assertTrue(via_binary(process, madtrex))
+        self.assertFalse(via_binary(process, classic))
+        self.assertFalse(via_binary(process, None))
+        for card in (self.run_card('lhe_npy'), self.run_card('compact_npy'),
+                     self.run_card(time_of_flight=1.0),
+                     self.run_card(legacy=True, native=False)):
+            process.run_card = card
+            self.assertFalse(via_binary(process, madtrex), card)
+        # the legacy systematics.py only runs without madspace's systematics
+        process.run_card = self.run_card(legacy=True, native=True)
+        self.assertTrue(via_binary(process, madtrex))
+        process.run_card, process.gridpack_run_mode = self.run_card(), 'minimal'
+        self.assertFalse(via_binary(process, madtrex))
+
+    def run_single_as_npy(self, reweighting):
+        """run_single() of an "lhe" run MadtRex reweights, with the
+        generation, the reweighting (reweighting(process)) and the npy->LHE
+        conversion faked; returns (process, calls)."""
+        from madgraph.various import reweight_backend as rb
+        from madgraph.iolibs.template_files.mg7 import npy_to_lhe
+        run_path = os.path.abspath(os.path.join('Events', 'run_01'))
+        os.makedirs(run_path)
+        calls = []
+        process = mock.Mock(gridpack_run_mode='regular', run_card=self.run_card(),
+                            run_path=run_path)
+
+        def generate_events(output_format=None):
+            calls.append(('generate', output_format))
+            process.events_path = os.path.join(run_path, 'events.npy')
+            for name in ('events.npy', 'header.lhe'):
+                open(os.path.join(run_path, name), 'w').close()
+        process.generate_events = generate_events
+
+        def convert(path):
+            calls.append(('convert', os.path.basename(path)))
+            lhe = os.path.join(run_path, 'events.lhe.gz')
+            open(lhe, 'w').close()
+            return lhe
+
+        def run_reweighting(p, card, decision):
+            calls.append(('reweight', p.events_path))
+            reweighting(p)
+        launch = self.launch
+        with mock.patch.object(launch, 'compute_auto_widths'), \
+                mock.patch.object(launch, 'MadgraphProcess', lambda: process), \
+                mock.patch.object(launch, 'run_lhe_postprocessing'), \
+                mock.patch.object(launch, 'run_reweighting', run_reweighting), \
+                mock.patch.object(npy_to_lhe, 'convert', convert):
+            launch.run_single(None, rb.Decision('madtrex', 'param'))
+        return process, calls
+
+    def test_madtrex_writes_the_requested_lhe(self):
+        def reweighting(process):
+            with open(os.path.join(process.run_path, 'reweighting.json'), 'w') as f:
+                json.dump({'events': process.events_path}, f)
+        process, calls = self.run_single_as_npy(reweighting)
+        npy = os.path.join(process.run_path, 'events.npy')
+        self.assertEqual(calls, [('generate', 'lhe_npy'), ('reweight', npy),
+                                 ('convert', 'events.npy')])
+        self.assertEqual(sorted(os.listdir(process.run_path)),
+                         ['events.lhe.gz', 'reweighting.json'])
+        lhe = os.path.join(process.run_path, 'events.lhe.gz')
+        self.assertEqual(process.events_path, lhe)
+        with open(os.path.join(process.run_path, 'reweighting.json')) as f:
+            self.assertEqual(json.load(f)['events'], lhe)
+
+    def test_failed_madtrex_still_writes_the_requested_lhe(self):
+        def reweighting(process):
+            raise RuntimeError('madtrex failed')
+        with self.assertLogs('madevent', 'WARNING') as logs:
+            process, calls = self.run_single_as_npy(reweighting)
+        self.assertEqual([c[0] for c in calls], ['generate', 'reweight', 'convert'])
+        self.assertTrue(any('reweighting failed: madtrex failed' in line
+                            for line in logs.output))
+        self.assertIn('events.lhe.gz', os.listdir(process.run_path))
+        self.assertNotIn('events.npy', os.listdir(process.run_path))
+
+    def test_the_decision_is_logged_and_recorded(self):
+        from madgraph.various import reweight_backend as rb
+        card = self.write_card("change backend auto\n" + self.CARD)
+        run_path = os.path.join('Events', 'run_01')
+        os.makedirs(run_path)
+        decision = rb.Decision('madtrex', 'param',
+                               ['MG7 output', 'madtrex installed',
+                                'no classic-only directives in {card}'])
+
+        class Process(object):
+            events_path = os.path.join(run_path, 'events.npy')
+            run_card = {'run': {'output_format': 'lhe_npy'}}
+        Process.run_path = run_path
+        seen = {}
+
+        def fake_madtrex(card, events_path, card_kind, process_directory,
+                         drop_stored_matrix_elements=False):
+            with open(card) as f:
+                seen['card'] = f.read()
+            seen['kind'] = card_kind
+        from madgraph.iolibs.template_files.mg7 import madtrex
+        with mock.patch.object(madtrex, 'run_reweighting', fake_madtrex), \
+                self.assertLogs('madevent', 'INFO') as logs:
+            self.launch.run_reweighting(Process(), card, decision=decision)
+        self.assertIn('Reweighting with madtrex (auto: MG7 output, madtrex installed, '
+                      'no classic-only directives in reweight_card.dat)', logs.output[0])
+        self.assertNotIn('\nchange backend', '\n' + seen['card'])
+        self.assertEqual(seen['kind'], 'param')
+        with open(os.path.join(run_path, 'reweighting.json')) as f:
+            record = json.load(f)
+        self.assertEqual((record['backend'], record['source'], record['weight_group']),
+                         ('madtrex', 'auto', 'tearex_reweighting'))
+        self.assertIn('change backend auto', record['card'])
+        # the weights went into the run's own event file
+        self.assertEqual(record['events'], Process.events_path)
+
+    def test_madtrex_refreshes_the_tools_copy_of_lhe_events(self):
+        import gzip
+        from madgraph.various import reweight_backend as rb
+        card = self.write_card(self.CARD)
+        run_path = os.path.join('Events', 'run_01')
+        os.makedirs(run_path)
+        events = os.path.join(run_path, 'events.lhe.gz')
+        copy = os.path.join(run_path, 'unweighted_events.lhe.gz')
+        for path in (events, copy):
+            with gzip.open(path, 'wt') as f:
+                f.write('before\n')
+
+        class Process(object):
+            events_path = events
+            run_card = {'run': {'output_format': 'lhe'}}
+        Process.run_path = run_path
+
+        def fake_madtrex(card, events_path, card_kind, process_directory,
+                         drop_stored_matrix_elements=False):
+            with gzip.open(events_path, 'wt') as f:
+                f.write('reweighted\n')
+        from madgraph.iolibs.template_files.mg7 import madtrex
+        with mock.patch.object(madtrex, 'run_reweighting', fake_madtrex):
+            self.launch.run_reweighting(Process(), card,
+                                        decision=rb.Decision('madtrex', 'param'))
+        with gzip.open(copy, 'rt') as f:
+            self.assertEqual(f.read(), 'reweighted\n')
+
+    def test_sync_lhe_copy(self):
+        import gzip
+        from madgraph.iolibs.template_files.mg7.run_interface import sync_lhe_copy
+        with gzip.open('a.lhe.gz', 'wt') as f:
+            f.write('events\n')
+        sync_lhe_copy('a.lhe.gz', 'b.lhe')
+        with open('b.lhe') as f:
+            self.assertEqual(f.read(), 'events\n')
+        sync_lhe_copy('b.lhe', 'c.lhe.gz')
+        with gzip.open('c.lhe.gz', 'rt') as f:
+            self.assertEqual(f.read(), 'events\n')
+        sync_lhe_copy('c.lhe.gz', 'a.lhe.gz')
+        self.assertFalse(os.path.exists('a.lhe.gz.tmp'))
+
+    def test_compact_npy_is_refused_for_madtrex(self):
+        from madgraph.various import reweight_backend as rb
+        card = self.write_card(self.CARD)
+        os.makedirs(os.path.join('Events', 'run_01'))
+
+        class Process(object):
+            run_path = os.path.join('Events', 'run_01')
+            run_card = {'run': {'output_format': 'compact_npy'}}
+        with self.assertRaisesRegex(ValueError, 'compact_npy'):
+            self.launch.run_reweighting(Process(), card,
+                                        decision=rb.Decision('madtrex', 'param'))
+
+    def test_reweighting_is_not_a_post_processing_tool(self):
+        self.assertEqual(self.launch.selected_tools({'reweight': 'ON'}), [])
+
+    def test_a_failing_reweighting_does_not_stop_the_run(self):
+        """The events are kept and MadSpin/shower/analyses still run; the
+        failure is logged with how to reweight again."""
+        os.makedirs(os.path.join('Events', 'run_01'))
+        launch = self.launch
+        process = mock.Mock(gridpack_run_mode='regular',
+                            run_path=os.path.abspath(os.path.join('Events', 'run_01')))
+        calls = []
+
+        def failing(p, c, decision):
+            raise RuntimeError('SPINUP 9 in event 3')
+        with mock.patch.object(launch, 'compute_auto_widths'), \
+                mock.patch.object(launch, 'MadgraphProcess', lambda: process), \
+                mock.patch.object(launch, 'run_lhe_postprocessing'), \
+                mock.patch.object(launch, 'run_reweighting', failing), \
+                mock.patch.object(launch, 'run_selected_tools',
+                                  lambda switch, p: calls.append('tools')), \
+                self.assertLogs('madevent', 'WARNING') as logs:
+            launch.run_single({'reweight': 'ON', 'shower': 'Pythia8'}, 'DECISION')
+        self.assertEqual(calls, ['tools'])
+        text = '\n'.join(logs.output)
+        self.assertIn('reweighting failed: SPINUP 9 in event 3', text)
+        self.assertIn("'reweight run_01'", text)
+        self.assertTrue(os.path.exists(os.path.join('Events', 'run_01',
+                                                    'reweighting_crash.log')))
+
+    def test_existing_run_finds_the_event_file_and_its_format(self):
+        import gzip
+        for run, files in (('lhe', ['events.lhe.gz']),
+                           ('lhe_npy', ['events.npy', 'header.lhe', 'events.lhe.gz']),
+                           ('compact_npy', ['events.npy', 'lhe_completer.json'])):
+            os.makedirs(os.path.join('Events', run))
+            for name in files:
+                open(os.path.join('Events', run, name), 'w').close()
+        os.makedirs(os.path.join('Events', 'empty'))
+        for run, events in (('lhe', 'events.lhe.gz'), ('lhe_npy', 'events.npy'),
+                            ('compact_npy', 'events.npy')):
+            existing = self.launch.ExistingRun(run)
+            self.assertEqual(existing.run_card['run']['output_format'], run)
+            self.assertEqual(os.path.basename(existing.events_path), events)
+        self.assertEqual(self.launch.list_runs(), ['compact_npy', 'lhe', 'lhe_npy'])
+        with self.assertRaisesRegex(ValueError, 'no event file'):
+            self.launch.ExistingRun('empty')
+        with self.assertRaisesRegex(ValueError, "no run 'run_09'.*runs: compact_npy"):
+            self.launch.ExistingRun('run_09')
+
+    def test_reweight_command_reweights_an_existing_run(self):
+        os.makedirs(os.path.join('Events', 'run_01'))
+        open(os.path.join('Events', 'run_01', 'events.lhe.gz'), 'w').close()
+        card = self.write_card(self.CARD, os.path.join(self.tmpdir.name, 'rw.dat'))
+        launch = self.launch
+        cmd = launch.MG7Cmd(me_dir=self.me_dir)
+        seen = []
+
+        def fake_prepare(switch, cli, options):
+            seen.append(('prepare', switch, cli))
+            return 'DECISION'
+
+        def fake_run(process, card_path, decision):
+            seen.append(('run', os.path.basename(process.run_path),
+                         process.run_card['run']['output_format'], decision))
+        with mock.patch.object(launch, 'prepare_reweighting', fake_prepare), \
+                mock.patch.object(launch, 'run_reweighting', fake_run):
+            cmd.do_reweight('run_01 --reweight=%s --reweight-backend=madtrex' % card)
+        self.assertEqual(seen, [('prepare', {'reweight': 'ON'}, 'madtrex'),
+                                ('run', 'run_01', 'lhe', 'DECISION')])
+        with open(os.path.join('Cards', 'reweight_card.dat')) as f:
+            self.assertEqual(f.read(), self.CARD)
+        with self.assertRaisesRegex(cmd.InvalidCmd, r'give the run to reweight \(run_01\)'):
+            cmd.do_reweight('')
+        with self.assertRaisesRegex(cmd.InvalidCmd, "no run 'run_02'"):
+            cmd.do_reweight('run_02')
+
+
+class ReweightSwitchTest(unittest.TestCase):
+    """The reweight switch of the launch question: an MG7 output offers the
+    backends that can run (ON is auto); a madevent output keeps its list."""
+
+    OPTIONS = {'pythia-pgs_path': None, 'pythia8_path': None,
+               'madanalysis5_path': None, 'exrootanalysis_path': None,
+               'delphes_path': None, 'rivet_path': None, 'mg5_path': '/mg5',
+               'f2py_compiler': None}
+
+    def setUp(self):
+        self.tmpdir = tempfile.TemporaryDirectory()
+        self.mg7 = make_mg7_dir(os.path.join(self.tmpdir.name, 'MG7'))
+        self.madevent = os.path.join(self.tmpdir.name, 'ME')
+        os.makedirs(os.path.join(self.madevent, 'Cards'))
+        os.makedirs(os.path.join(self.madevent, 'bin'))
+        open(os.path.join(self.madevent, 'bin', 'madevent'), 'w').close()
+
+    def tearDown(self):
+        self.tmpdir.cleanup()
+
+    def question(self, me_dir, f2py, madtrex, card=None):
+        from madgraph.interface import madevent_interface as me
+        from madgraph.various import reweight_backend as rb
+        if card is not None:
+            with open(os.path.join(me_dir, 'Cards', 'reweight_card.dat'), 'w') as f:
+                f.write(card)
+        question = me.AskRun.__new__(me.AskRun)
+        question.me_dir = me_dir
+        question.mother_interface = mock.Mock(plugin_path=[])
+        question.switch = {}
+        with mock.patch.object(me.misc, 'has_f2py', return_value=f2py), \
+                mock.patch.object(rb, 'madtrex_is_installed', return_value=madtrex):
+            question.check_available_module(dict(self.OPTIONS))
+            question.set_default_reweight()
+        return question
+
+    def test_mg7_with_both_backends(self):
+        q = self.question(self.mg7, f2py=True, madtrex=True)
+        self.assertEqual(q.get_allowed_reweight(),
+                         ['OFF', 'ON', 'classic', 'madtrex', 'density'])
+        self.assertEqual(q.switch['reweight'], 'OFF')
+
+    def test_mg7_without_f2py(self):
+        q = self.question(self.mg7, f2py=False, madtrex=True)
+        self.assertEqual(q.get_allowed_reweight(), ['OFF', 'ON', 'madtrex'])
+
+    def test_mg7_without_madtrex(self):
+        q = self.question(self.mg7, f2py=True, madtrex=False)
+        self.assertEqual(q.get_allowed_reweight(), ['OFF', 'ON', 'classic', 'density'])
+
+    def test_mg7_with_neither(self):
+        q = self.question(self.mg7, f2py=False, madtrex=False)
+        self.assertEqual(q.get_allowed_reweight(), [])
+        self.assertEqual(q.switch['reweight'], 'Not Avail.')
+
+    def test_madevent_is_unchanged(self):
+        for madtrex in (True, False):
+            q = self.question(self.madevent, f2py=True, madtrex=madtrex)
+            self.assertEqual(q.get_allowed_reweight(), ['OFF', 'ON', 'density'])
+        q = self.question(self.madevent, f2py=False, madtrex=True)
+        self.assertEqual(q.get_allowed_reweight(), [])
+
+    def test_switch_value_from_the_card(self):
+        q = self.question(self.mg7, True, True, card="change backend madtrex\nlaunch\n")
+        self.assertEqual(q.switch['reweight'], 'madtrex')
+        q = self.question(self.mg7, True, True, card="change backend auto\nlaunch\n")
+        self.assertEqual(q.switch['reweight'], 'ON')
+        q = self.question(self.mg7, True, True,
+                          card="change particle_in_density_matrix 3\nlaunch\n")
+        self.assertEqual(q.switch['reweight'], 'density')
+        # a backend that cannot run here is just ON
+        q = self.question(self.mg7, True, False, card="change backend madtrex\nlaunch\n")
+        self.assertEqual(q.switch['reweight'], 'ON')
+        # a madevent output does not read the backend line
+        q = self.question(self.madevent, True, True, card="change backend classic\nlaunch\n")
+        self.assertEqual(q.switch['reweight'], 'ON')
+
+    def test_a_backend_picked_in_the_question_survives_the_card_sync(self):
+        q = self.question(self.mg7, True, True, card="launch\nset MT 175\n")
+        self.assertEqual(q.switch['reweight'], 'ON')
+        for value in ('classic', 'madtrex'):
+            q.switch['reweight'] = value
+            self.assertEqual(q.switch_value_from_card_reweight(), value)
+        # also over the card's own backend line, which only replaces ON
+        with open(os.path.join(self.mg7, 'Cards', 'reweight_card.dat'), 'w') as f:
+            f.write("change backend classic\nlaunch\n")
+        q.switch['reweight'] = 'madtrex'
+        self.assertEqual(q.switch_value_from_card_reweight(), 'madtrex')
+        q.switch['reweight'] = 'ON'
+        self.assertEqual(q.switch_value_from_card_reweight(), 'classic')
+
+    def test_backend_values_enable_the_card(self):
+        from madgraph.interface.madevent_interface import AskRunEditCard
+        spec = [c for c in AskRunEditCard.switch_cards if c['key'] == 'reweight'][0]
+        for value in ('ON', 'classic', 'madtrex', 'density'):
+            self.assertTrue(spec['on']({'reweight': value}), value)
+        q = self.question(self.mg7, True, True, card="launch\n")
+        self.assertEqual(q.get_cardcmd_for_reweight('madtrex'), [])
+
+
 class TestPostProcessingIsSkippedWhenThereIsNothingToDo(unittest.TestCase):
     """A plain generate/output/launch announced a post-processing step with
     nothing after the colon, and built the run interface to do nothing.
@@ -1345,7 +1855,19 @@ def test_scan_card_survives(self):
     def test_missing_output_format_line(self):
         with open(self.card, 'w') as f:
             f.write('[run]\nseed = 3\n')
-        self.launch.force_lhe_output_if_needed({'reweight': 'ON'})
+        self.launch.force_lhe_output_if_needed({'madspin': 'ON'})
+        self.assertEqual(self.output_format(), 'lhe')
+
+    def test_reweighting_forces_lhe_for_the_classic_module_only(self):
+        from madgraph.various import reweight_backend as rb
+        madtrex = rb.Decision('madtrex', 'param')
+        self.launch.force_lhe_output_if_needed({'reweight': 'ON'}, madtrex)
+        self.assertEqual(self.output_format(), 'lhe_npy')
+        # without a decision (no card, no launch) nothing is reweighted
+        self.launch.force_lhe_output_if_needed({'reweight': 'ON'}, None)
+        self.assertEqual(self.output_format(), 'lhe_npy')
+        self.launch.force_lhe_output_if_needed({'reweight': 'ON'},
+                                               rb.Decision('classic', 'classic'))
         self.assertEqual(self.output_format(), 'lhe')
 
 
diff --git a/tests/unit_tests/iolibs/test_export_cpp.py b/tests/unit_tests/iolibs/test_export_cpp.py
index 1f5369b615..f6961727b3 100755
--- a/tests/unit_tests/iolibs/test_export_cpp.py
+++ b/tests/unit_tests/iolibs/test_export_cpp.py
@@ -911,6 +911,7 @@ def test_cpp_export_decay_chain_broken_symmetry_metadata(self):
             'process_lines': '',
             'model_name': 'sm',
             'initProc_lines': '',
+            'refreshMasses_lines': '',
             'reset_jamp_lines': '',
             'sigmaKin_lines': '',
             'sigmaHat_lines': 'return 0.;',
diff --git a/tests/unit_tests/iolibs/test_madmatrix_density.py b/tests/unit_tests/iolibs/test_madmatrix_density.py
index b7fa245559..625519a59a 100644
--- a/tests/unit_tests/iolibs/test_madmatrix_density.py
+++ b/tests/unit_tests/iolibs/test_madmatrix_density.py
@@ -102,10 +102,10 @@ def touch(self, name):
 
     def test_make_variables(self):
         self.assertEqual(density.make_variables(),
-                         ['PGOAL=lib', 'BACKEND=scalar', 'DENSITY=1'])
+                         ['PGOAL=lib', 'BACKEND=scalar', 'FPTYPE=m', 'DENSITY=1'])
         self.assertEqual(density.build_command(4, 'simd_128', '_ms2d'),
-                         ['make', '-j4', 'PGOAL=lib', 'BACKEND=simd_128', 'DENSITY=1',
-                          'LIBTAG=_ms2d'])
+                         ['make', '-j4', 'PGOAL=lib', 'BACKEND=simd_128', 'FPTYPE=m',
+                          'DENSITY=1', 'LIBTAG=_ms2d'])
 
     def test_find_libtag(self):
         self.assertRaises(density.MadDensityError, density.find_libtag, self.tmp)
@@ -116,6 +116,21 @@ def test_find_libtag(self):
         self.touch('libmadmatrix_common_scalar.so')
         self.assertRaises(density.MadDensityError, density.find_libtag, self.tmp)
 
+    def test_find_libtag_model_fptype(self):
+        """An output whose process table names the common library (with the
+        model and FPTYPE): the tag is read off that name."""
+        with open(os.path.join(self.tmp, 'SubProcesses', density.PROCESSES_FILE), 'w') as f:
+            json.dump({'common_library': 'libmadmatrix_common{libtag}_sm_{backend}_{fptype}.so',
+                       'subprocesses': []}, f)
+        self.assertRaises(density.MadDensityError, density.find_libtag, self.tmp)
+        self.touch('libmadmatrix_common_ms1p_sm_scalar_m.so')
+        self.touch('libmadmatrix_common_ms1p_sm_scalar_d.so')
+        self.assertEqual(density.find_libtag(self.tmp), '_ms1p')
+        self.assertEqual(density.common_library_path(self.tmp, '_ms1p'),
+                         os.path.join(self.tmp, 'lib', 'libmadmatrix_common_ms1p_sm_scalar_m.so'))
+        self.touch('libmadmatrix_common_sm_scalar_m.so')
+        self.assertRaises(density.MadDensityError, density.find_libtag, self.tmp)
+
     def test_is_output(self):
         self.assertFalse(density.is_output(self.tmp))
         with open(os.path.join(self.tmp, 'SubProcesses', density.PROCESSES_FILE), 'w') as f:
diff --git a/tests/unit_tests/iolibs/test_madmatrix_writer.py b/tests/unit_tests/iolibs/test_madmatrix_writer.py
index a9756dea07..1fa72f32b9 100644
--- a/tests/unit_tests/iolibs/test_madmatrix_writer.py
+++ b/tests/unit_tests/iolibs/test_madmatrix_writer.py
@@ -168,3 +168,12 @@ def test_no_optimised_call_anywhere(self):
             for call in self.external_calls(ids):
                 for name in self.OPTIMISED:
                     self.assertNotIn(name, call, ids)
+
+    def test_helicity_is_per_lane(self):
+        """The external wavefunctions take the helicity of each SIMD lane/GPU
+        thread (hel_sv, built in calculate_jamps), not one helicity for all
+        events, so that events with different helicities run in lock-step."""
+        calls = self.external_calls([81, -81, 23, 21])   # q q~ > z g
+        for ipar, call in enumerate(calls):
+            self.assertIn('hel_sv[%d]' % ipar, call)
+            self.assertNotIn('cHel[', call)
diff --git a/tests/unit_tests/various/test_madtrex_standalone.py b/tests/unit_tests/various/test_madtrex_standalone.py
new file mode 100644
index 0000000000..04ef7aaecc
--- /dev/null
+++ b/tests/unit_tests/various/test_madtrex_standalone.py
@@ -0,0 +1,814 @@
+################################################################################
+#
+# Copyright (c) 2009 The MadGraph7 Development team and Contributors
+#
+# This file is a part of the MadGraph7 project, an application which
+# automatically generates Feynman diagrams and matrix elements for arbitrary
+# high-energy processes in the Standard Model and beyond.
+#
+# It is subject to the MadGraph7 license which should accompany this
+# distribution.
+#
+# For more information, visit madgraph.phys.ucl.ac.be and amcatnlo.web.cern.ch
+#
+################################################################################
+"""mg7/madtrex_driver.py, the standalone MadtRex reweighting driver that
+gridpacks ship as bin/reweight, runs without the rest of MadGraph.
+
+It is run as a plain script, in a subprocess where the madgraph package
+cannot be imported. Skipped unless madspace and madtrex are installed in
+madspace/install and madtrex/install.
+"""
+
+import os
+import shutil
+import subprocess
+import sys
+import tempfile
+import textwrap
+import unittest
+from pathlib import Path
+
+MG_ROOT = Path(__file__).resolve().parents[3]
+MG7_DIR = MG_ROOT / "madgraph" / "iolibs" / "template_files" / "mg7"
+SCRIPT = MG7_DIR / "madtrex_driver.py"
+
+
+def madtrex_is_installed():
+    return (MG_ROOT / "madspace" / "install" / "madspace").is_dir() and (
+        MG_ROOT / "madtrex" / "install" / "madtrex"
+    ).is_dir()
+
+
+@unittest.skipUnless(madtrex_is_installed(), "madspace/madtrex are not installed")
+class MadtrexStandaloneTest(unittest.TestCase):
+
+    def run_isolated(self, code):
+        """Runs code with -I (no PYTHONPATH, no cwd on sys.path), from a
+        directory outside MadGraph, so madgraph cannot be imported."""
+        with tempfile.TemporaryDirectory() as tmp:
+            return subprocess.run(
+                [sys.executable, "-I", "-c", code],
+                cwd=tmp, capture_output=True, text=True,
+            )
+
+    def test_script_runs_without_madgraph(self):
+        result = self.run_isolated(textwrap.dedent(f"""
+            import runpy, sys
+            sys.argv = [{str(SCRIPT)!r}, "--help"]
+            try:
+                runpy.run_path({str(SCRIPT)!r}, run_name="__main__")
+            except SystemExit as exc:
+                assert exc.code == 0, exc.code
+            assert not any(m == "madgraph" or m.startswith("madgraph.")
+                           for m in sys.modules), "madgraph was imported"
+            import madtrex
+            assert madtrex.__file__.startswith({str(MG_ROOT / "madtrex")!r}), madtrex.__file__
+        """))
+        self.assertEqual(result.returncode, 0, result.stderr)
+        self.assertIn("--param-rwgt-card", result.stdout)
+
+    def test_run_card_is_read_without_madgraph(self):
+        with tempfile.TemporaryDirectory() as proc_dir:
+            os.mkdir(os.path.join(proc_dir, "Cards"))
+            with open(os.path.join(proc_dir, "Cards", "run_card.toml"), "w") as f:
+                f.write('[run]\nprecision = "all64"\n\n[beam]\npdf1 = "X"\npdf2 = "X"\n')
+            result = self.run_isolated(textwrap.dedent(f"""
+                import runpy
+                ns = runpy.run_path({str(SCRIPT)!r})
+                print(ns["_read_run_fptype"]({proc_dir!r}),
+                      ns["_read_default_pdf_set"]({proc_dir!r})[0],
+                      ns["_read_run_card"]({proc_dir!r} + "/Cards/run_card.toml")["run"]["device"])
+            """))
+        self.assertEqual(result.returncode, 0, result.stderr)
+        self.assertEqual(result.stdout.split(), ["d", "X", "['cpu']"])
+
+    def test_reweighting_threads(self):
+        """MadtRex uses as many threads as the generation: the run card's
+        cpu_thread_pool_size, all available cores for -1."""
+        with tempfile.TemporaryDirectory() as proc_dir:
+            os.mkdir(os.path.join(proc_dir, "Cards"))
+            card = os.path.join(proc_dir, "Cards", "run_card.toml")
+            result = []
+            for threads in (3, -1):
+                with open(card, "w") as f:
+                    f.write("[run]\ncpu_thread_pool_size = %d\n" % threads)
+                result.append(self.run_isolated(textwrap.dedent(f"""
+                    import os, runpy
+                    ns = runpy.run_path({str(SCRIPT)!r})
+                    print(ns["reweighting_threads"]({proc_dir!r}) ==
+                          {{3: 3, -1: len(os.sched_getaffinity(0))}}[{threads}])
+                """)))
+        for each in result:
+            self.assertEqual(each.returncode, 0, each.stderr)
+            self.assertEqual(each.stdout.strip(), "True")
+
+    def test_generation_run_card_comes_from_the_events(self):
+        """The baseline PDF set and the normaliser fptype are those the
+        events were generated with (the run card in their LHE header), not
+        those of the current Cards/run_card.toml, edited since."""
+        import gzip
+        with tempfile.TemporaryDirectory() as proc_dir:
+            os.mkdir(os.path.join(proc_dir, "Cards"))
+            with open(os.path.join(proc_dir, "Cards", "run_card.toml"), "w") as f:
+                f.write('[run]\nprecision = "all32"\n\n[beam]\npdf1 = "NEW"\npdf2 = "NEW"\n')
+            header = ('<LesHouchesEvents version="3.0">\n<header>\n<MG7RunCard>\n'
+                      '[run]\nprecision = &quot;all64&quot;\n\n[beam]\n'
+                      'pdf1 = &quot;OLD&quot;\npdf2 = &quot;OLD&quot;\n</MG7RunCard>\n'
+                      '</header>\n<init>\n2212 2212 6.5e+03 6.5e+03 0 0 1 1 3 1\n</init>\n')
+            npy_run = os.path.join(proc_dir, "Events", "run_01")
+            lhe_run = os.path.join(proc_dir, "Events", "run_02")
+            for run in (npy_run, lhe_run):
+                os.makedirs(run)
+            with open(os.path.join(npy_run, "header.lhe"), "w") as f:
+                f.write(header + "</LesHouchesEvents>\n")
+            with gzip.open(os.path.join(lhe_run, "events.lhe.gz"), "wt") as f:
+                f.write(header + "<event>\n</event>\n</LesHouchesEvents>\n")
+            result = self.run_isolated(textwrap.dedent(f"""
+                import runpy
+                ns = runpy.run_path({str(SCRIPT)!r})
+                for events in ({npy_run + "/events.npy"!r}, {lhe_run + "/events.lhe.gz"!r}, None):
+                    print(ns["_read_run_fptype"]({proc_dir!r}, events),
+                          ns["_read_default_pdf_set"]({proc_dir!r}, events)[0])
+            """))
+        self.assertEqual(result.returncode, 0, result.stderr)
+        self.assertEqual(result.stdout.split(), ["d", "OLD", "d", "OLD", "f", "NEW"])
+
+    def test_classic_cards_are_refused(self):
+        """bin/reweight is MadtRex only: a card for the classic module says
+        that classic reweighting needs MadGraph."""
+        with tempfile.TemporaryDirectory() as tmp:
+            cards = {"classic": "change backend classic\nlaunch\nset MT 175\n",
+                     "output": "change output 2.0\nlaunch\nset MT 175\n",
+                     "madtrex": "change backend madtrex\nlaunch\nset MT 175\n",
+                     "auto": "change backend auto\nlaunch\nset MT 175\n"}
+            for name, text in cards.items():
+                with open(os.path.join(tmp, name), "w") as f:
+                    f.write(text)
+            result = self.run_isolated(textwrap.dedent(f"""
+                import runpy
+                ns = runpy.run_path({str(SCRIPT)!r})
+                for name in ("classic", "output", "madtrex", "auto"):
+                    try:
+                        ns["check_card_for_standalone"]({tmp!r} + "/" + name)
+                        print(name, "ok")
+                    except SystemExit as exc:
+                        print(name, "MadGraph" in str(exc))
+            """))
+        self.assertEqual(result.returncode, 0, result.stderr)
+        self.assertEqual(result.stdout.split(), ["classic", "True", "output", "True",
+                                                 "madtrex", "ok", "auto", "ok"])
+
+    def test_gridpack_layout(self):
+        """As bin/reweight in a gridpack: madspace/madtrex are found next to
+        bin/, the gridpack is the default process directory, and the
+        resolved cpu backend comes from grid_run_card.toml."""
+        with tempfile.TemporaryDirectory() as gridpack:
+            # resolved: the driver locates itself through its resolved path
+            # (macOS's /var/folders is a symlink into /private)
+            gridpack = Path(gridpack).resolve()
+            for pkg in ("madspace", "madtrex"):
+                (gridpack / pkg / "install").mkdir(parents=True)
+                os.symlink(MG_ROOT / pkg / "install" / pkg,
+                           gridpack / pkg / "install" / pkg)
+            os.symlink(MG_ROOT / "madtrex" / "install" / "pyrex",
+                       gridpack / "madtrex" / "install" / "pyrex")
+            (gridpack / "bin").mkdir()
+            for name in ("param_reweight_card.py", "pdf_reweight_card.py",
+                         "generic_reweight_card.py"):
+                shutil.copy(MG7_DIR / name, gridpack / "bin" / name)
+            shutil.copy(SCRIPT, gridpack / "bin" / "reweight")
+            (gridpack / "SubProcesses").mkdir()
+            (gridpack / "SubProcesses" / "subprocesses.json").write_text("[]")
+            (gridpack / "Cards").mkdir()
+            (gridpack / "Cards" / "run_card.toml").write_text('[run]\ncpu_mode = "auto"\n')
+            (gridpack / "Cards" / "grid_run_card.toml").write_text(
+                '[run]\ndevice = ["cpu"]\ncpu_mode = "avx512y"\n')
+            script = gridpack / "bin" / "reweight"
+            result = self.run_isolated(textwrap.dedent(f"""
+                import runpy
+                ns = runpy.run_path({str(script)!r})
+                import madtrex
+                print(madtrex.__file__.startswith({str(gridpack)!r}),
+                      ns["_default_process_directory"]() == {str(gridpack)!r},
+                      ns["_read_devices"]({str(gridpack)!r}))
+            """))
+        self.assertEqual(result.returncode, 0, result.stderr)
+        self.assertEqual(result.stdout.split(), ["True", "True", "['avx512y']"])
+
+
+@unittest.skipUnless(madtrex_is_installed(), "madspace/madtrex are not installed")
+class RunReweightingRoutingTest(unittest.TestCase):
+    """madtrex_driver.run_reweighting() hands each card kind to its runner."""
+
+    def test_routing(self):
+        with tempfile.TemporaryDirectory() as tmp:
+            cards = {"param": "launch\nset MT 175\n",
+                     "pdf": "launch\nset PDF_SET X\n",
+                     "generic": "change mode pdf\nlaunch\nset PDF_SET X\n"}
+            for kind, text in cards.items():
+                with open(os.path.join(tmp, kind + ".dat"), "w") as f:
+                    f.write(text)
+            result = subprocess.run(
+                [sys.executable, "-I", "-c", textwrap.dedent(f"""
+                    import importlib.util, sys
+                    spec = importlib.util.spec_from_file_location("drv", {str(SCRIPT)!r})
+                    drv = importlib.util.module_from_spec(spec)
+                    spec.loader.exec_module(drv)
+                    for name in ("run_parameter_reweighting", "run_pdf_reweighting",
+                                 "run_generic_reweighting"):
+                        setattr(drv, name, lambda card, _n=name, **kw: _n)
+                    out = [drv.run_reweighting({tmp!r} + "/" + k + ".dat")
+                           for k in ("param", "pdf", "generic")]
+                    out.append(drv.run_reweighting({tmp!r} + "/param.dat", card_kind="generic"))
+                    try:
+                        drv.run_reweighting({tmp!r} + "/param.dat", card_kind="nlo")
+                    except ValueError:
+                        out.append("error")
+                    print(" ".join(out))
+                """)],
+                cwd=tmp, capture_output=True, text=True,
+            )
+        self.assertEqual(result.returncode, 0, result.stderr)
+        self.assertEqual(result.stdout.split(), [
+            "run_parameter_reweighting", "run_pdf_reweighting", "run_generic_reweighting",
+            "run_generic_reweighting", "error"])
+
+
+def load_driver_module():
+    """mg7/madtrex_driver.py, loaded by path (it imports madspace/madtrex)."""
+    import importlib.util
+    spec = importlib.util.spec_from_file_location("madtrex_driver_under_test", SCRIPT)
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+@unittest.skipUnless(madtrex_is_installed(), "madspace/madtrex are not installed")
+class InPlaceOutputTest(unittest.TestCase):
+    """By default the weights go into the event file that was reweighted,
+    after the weights already there -- the same file the classic module and
+    the other tools use."""
+
+    def setUp(self):
+        self.driver = load_driver_module()
+        self.tmp = tempfile.mkdtemp()
+
+    def tearDown(self):
+        shutil.rmtree(self.tmp)
+
+    class Result:
+        def __init__(self, ids, contexts=()):
+            self.weight_ids = ids
+            self.weight_context = list(contexts)
+
+    def test_lhe_gz_is_replaced_and_stays_gzipped(self):
+        import gzip
+        source = os.path.join(self.tmp, "events.lhe.gz")
+        with gzip.open(source, "wt") as f:
+            f.write("original\n")
+        plain = self.driver.prepare_events(source)
+        self.assertTrue(os.path.isfile(plain))
+
+        class Driver:
+            def write_weights(self, path):
+                with open(path, "w") as f:
+                    f.write("original + new weights\n")
+        out = self.driver.write_weights_in_place(Driver(), self.Result(["mt"]), source, plain)
+        self.assertEqual(out, source)
+        with gzip.open(source, "rt") as f:
+            self.assertEqual(f.read(), "original + new weights\n")
+        # the decompressed copy is gone: it would shadow events.lhe.gz
+        self.assertEqual(sorted(os.listdir(self.tmp)), ["events.lhe.gz"])
+
+    def test_plain_lhe_is_replaced(self):
+        source = os.path.join(self.tmp, "events.lhe")
+        with open(source, "w") as f:
+            f.write("original\n")
+
+        class Driver:
+            def write_weights(self, path):
+                with open(path, "w") as f:
+                    f.write("new\n")
+        self.driver.write_weights_in_place(Driver(), self.Result(["mt"]), source, source)
+        with open(source) as f:
+            self.assertEqual(f.read(), "new\n")
+        self.assertEqual(os.listdir(self.tmp), ["events.lhe"])
+
+    def npy_run(self):
+        import numpy as np
+        dtype = [("process_id", "<i4"), ("weight", "<f8"), ("rwgt_1", "<f8"),
+                 ("rwgt_2", "<f8"), ("part1_pdg_id", "<i4"), ("part1_px", "<f8")]
+        events = np.zeros(3, dtype=dtype)
+        events["weight"] = [1., 2., 3.]
+        events["rwgt_1"] = [1.1, 2.1, 3.1]
+        events["part1_px"] = [7., 8., 9.]
+        source = os.path.join(self.tmp, "events.npy")
+        np.save(source, events)
+        with open(os.path.join(self.tmp, "header.lhe"), "w") as f:
+            f.write("<LesHouchesEvents>\n<header>\n<initrwgt>\n<weightgroup name='scale'>\n"
+                    "<weight id='1'>x</weight>\n</weightgroup>\n</initrwgt>\n</header>\n")
+        return source
+
+    def npy_driver(self, weights):
+        import numpy as np
+
+        class Driver:
+            def write_weights(self, prefix):
+                for weight_id, values in weights.items():
+                    data = np.zeros(len(values), dtype=[("weight", "<f8")])
+                    data["weight"] = values
+                    np.save(f"{prefix}_{weight_id}.npy", data)
+        return Driver()
+
+    def test_npy_gets_named_columns_after_the_systematics(self):
+        import numpy as np
+        source = self.npy_run()
+        driver = self.npy_driver({"mt175": [10., 20., 30.]})
+        self.driver.write_weights_in_place(driver, self.Result(["mt175"], ["set MT 175"]),
+                                           source, source)
+        events = np.load(source)
+        self.assertEqual(events.dtype.names, ("process_id", "weight", "rwgt_1", "rwgt_2",
+                                              "rwgt_mt175", "part1_pdg_id", "part1_px"))
+        self.assertEqual(list(events["rwgt_mt175"]), [10., 20., 30.])
+        self.assertEqual(list(events["rwgt_1"]), [1.1, 2.1, 3.1])
+        self.assertEqual(list(events["part1_px"]), [7., 8., 9.])
+        with open(os.path.join(self.tmp, "header.lhe")) as f:
+            header = f.read()
+        self.assertIn("<weight id='1'>x</weight>", header)
+        self.assertIn("<weight id='mt175'>set MT 175</weight>", header)
+        self.assertLess(header.index("id='1'"), header.index("id='mt175'"))
+        self.assertEqual(sorted(os.listdir(self.tmp)), ["events.npy", "header.lhe"])
+
+        # a second reweighting follows the first; the same name again is an error
+        driver = self.npy_driver({"mt170": [4., 5., 6.]})
+        self.driver.write_weights_in_place(driver, self.Result(["mt170"]), source, source)
+        names = np.load(source).dtype.names
+        self.assertEqual(names[4:6], ("rwgt_mt175", "rwgt_mt170"))
+        driver = self.npy_driver({"mt170": [4., 5., 6.]})
+        with self.assertRaisesRegex(ValueError, "already has a weight"):
+            self.driver._merge_npy_weights(source, os.path.join(self.tmp, ".madtrex_weights"),
+                                           ["mt170"], [""])
+
+    def test_npy_without_systematics(self):
+        import numpy as np
+        source = os.path.join(self.tmp, "events.npy")
+        events = np.zeros(2, dtype=[("weight", "<f8"), ("bias", "<f8"), ("part1_px", "<f8")])
+        np.save(source, events)
+        self.driver.write_weights_in_place(self.npy_driver({"a": [1., 2.]}), self.Result(["a"]),
+                                           source, source)
+        self.assertEqual(np.load(source).dtype.names, ("weight", "rwgt_a", "bias", "part1_px"))
+
+
+class NpyToLheNamedWeightsTest(unittest.TestCase):
+    """npy_to_lhe writes the named (MadtRex) weights after the numbered ones."""
+
+    def test_named_weights(self):
+        sys.path.insert(0, str(MG_ROOT))
+        try:
+            from madgraph.iolibs.template_files.mg7.npy_to_lhe import _add_named_weights
+        finally:
+            sys.path.remove(str(MG_ROOT))
+        event = "<event>\n 1 2 3\n<rwgt>\n<wgt id='1'> +1.0e+00 </wgt>\n</rwgt>\n</event>\n"
+        out = _add_named_weights(event, ["mt175"], [2.5])
+        self.assertIn("<wgt id='1'> +1.0e+00 </wgt>\n<wgt id='mt175'> +2.5000000e+00 </wgt>\n"
+                      "</rwgt>", out)
+        out = _add_named_weights("<event>\n 1 2 3\n</event>\n", ["a"], [1.])
+        self.assertIn("<rwgt>\n<wgt id='a'> +1.0000000e+00 </wgt>\n</rwgt>\n</event>", out)
+        self.assertEqual(_add_named_weights(event, [], []), event)
+
+
+@unittest.skipUnless(madtrex_is_installed(), "madspace/madtrex are not installed")
+class RexHeaderParsingTest(unittest.TestCase):
+    """The LHE header of an MG7 run embeds the run card unescaped, and its
+    comments hold text like "<run>" or "<set>@<member>". Rex's XML parser
+    used to take those for unclosed elements and nest the rest of the header
+    (<initrwgt> included) inside them, so the existing weight ids were not
+    found and MadtRex wrote its own weight names over the first ones."""
+
+    LHE = textwrap.dedent("""\
+        <LesHouchesEvents version="3.0">
+        <header>
+        <MG7RunCard>
+        # plot into Events/<run>/plots; "<set>@<member>"; <group>[-<group>]-<observable>.min
+        events = 10
+        </MG7RunCard>
+        <MG7Seed>
+        42
+        </MG7Seed>
+        <initrwgt>
+        <weightgroup name="scale" combine="envelope">
+        <weight id="1" MUR="0.5"> MUR=0.5 </weight>
+        <weight id="2" MUR="2.0"> MUR=2.0 </weight>
+        </weightgroup>
+        </initrwgt>
+        </header>
+        <init>
+        2212 2212 6.5e+03 6.5e+03 0 0 0 0 3 1
+        1.0e+00 1.0e-02 1.0e+00 1
+        </init>
+        <event>
+        4 1 +1.0e+00 1.0e+02 7.8e-03 1.2e-01
+        21 -1 0 0 501 502 0.0 0.0 5.0e+02 5.0e+02 0.0 0.0 9.0
+        21 -1 0 0 502 503 0.0 0.0 -5.0e+02 5.0e+02 0.0 0.0 9.0
+        6 1 1 2 501 0 1.0e+02 0.0 0.0 4.5e+02 1.73e+02 0.0 9.0
+        -6 1 1 2 0 503 -1.0e+02 0.0 0.0 4.5e+02 1.73e+02 0.0 9.0
+        <rwgt>
+        <wgt id='1'> +1.1e+00 </wgt>
+        <wgt id='2'> +0.9e+00 </wgt>
+        </rwgt>
+        </event>
+        </LesHouchesEvents>
+        """)
+
+    def test_weight_ids_survive_unescaped_card_text(self):
+        with tempfile.TemporaryDirectory() as tmp:
+            path = os.path.join(tmp, "events.lhe")
+            with open(path, "w") as f:
+                f.write(self.LHE)
+            result = subprocess.run(
+                [sys.executable, "-I", "-c", textwrap.dedent(f"""
+                    import sys
+                    sys.path[:0] = [{str(MG_ROOT / "madtrex" / "install")!r},
+                                    {str(MG_ROOT / "madspace" / "install")!r}]
+                    import madtrex
+                    lhe = madtrex.load_lhe_xml({path!r})
+                    lhe.extract_weight_ids()
+                    print(list(lhe.weight_ids))
+                """)],
+                capture_output=True, text=True,
+            )
+        self.assertEqual(result.returncode, 0, result.stderr)
+        self.assertEqual(result.stdout.strip(), "['1', '2']")
+
+
+def load_card_module(name):
+    """One of the mg7 reweight-card modules, which import nothing from
+    MadGraph, loaded by path."""
+    import importlib.util
+    spec = importlib.util.spec_from_file_location(name, MG7_DIR / f"{name}.py")
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+class HelicityReweightingCardTest(unittest.TestCase):
+    """"set helicity_reweighting <bool>" in the reweight cards."""
+
+    LAUNCH = "launch --rwgt_name=x\nset MASS 6 1.75e+02\n"
+
+    def setUp(self):
+        self.param = load_card_module("param_reweight_card")
+
+    def options(self, text):
+        return self.param.parse_reweight_card(text).driver_options
+
+    def test_unset_leaves_the_driver_default(self):
+        self.assertIsNone(self.options(self.LAUNCH).helicity_reweighting)
+
+    def test_set_in_the_preamble(self):
+        for value, expected in (("false", False), ("off", False), ("true", True), ("1", True)):
+            text = f"set helicity_reweighting {value}\n" + self.LAUNCH
+            self.assertIs(self.options(text).helicity_reweighting, expected, value)
+
+    def test_use_helicity_is_an_alias(self):
+        text = "set use_helicity false\n" + self.LAUNCH
+        self.assertIs(self.options(text).helicity_reweighting, False)
+
+    def test_not_allowed_after_a_launch(self):
+        with self.assertRaises(ValueError):
+            self.param.parse_reweight_card(self.LAUNCH + "set helicity_reweighting false\n")
+
+    def test_not_a_boolean(self):
+        with self.assertRaises(ValueError):
+            self.param.parse_reweight_card("set helicity_reweighting maybe\n" + self.LAUNCH)
+
+    def test_applied_to_the_driver(self):
+        calls = []
+
+        class FakeDriver:
+            def set_helicity_reweighting(self, value):
+                calls.append(value)
+
+        self.options(self.LAUNCH).apply(FakeDriver())
+        self.assertEqual(calls, [])
+        self.options("set helicity_reweighting false\n" + self.LAUNCH).apply(FakeDriver())
+        self.assertEqual(calls, [False])
+
+    def test_generic_card_parameter_section(self):
+        generic = load_card_module("generic_reweight_card")
+        sections = generic.split_by_mode(
+            "set mode parameter\nset helicity_reweighting false\n" + self.LAUNCH)
+        self.assertIs(self.options(sections["parameter"]).helicity_reweighting, False)
+
+    def test_rejected_by_the_pdf_card(self):
+        pdf = load_card_module("pdf_reweight_card")
+        for name in ("helicity_reweighting", "use_helicity"):
+            with self.assertRaisesRegex(ValueError, "parameter"):
+                pdf.parse_pdf_reweight_card(
+                    f"set {name} true\nlaunch\nset PDF_SET NNPDF40_lo_as_01180\n")
+
+
+class ClassicVocabularyTest(unittest.TestCase):
+    """The MadtRex card readers share the classic reweighting vocabulary."""
+
+    LAUNCH = "launch --rwgt_name=x\nset MASS 6 1.75e+02\n"
+    PARAM_CARD = ("BLOCK MASS # masses\n  6 1.730000e+02 # MT\n 25 1.25e+02 # MH\n"
+                  "BLOCK SMINPUTS #\n  1 1.3e+02 # aEWM1\nDECAY 6 1.49 # WT\n")
+
+    def setUp(self):
+        self.param = load_card_module("param_reweight_card")
+        self.pdf = load_card_module("pdf_reweight_card")
+        self.generic = load_card_module("generic_reweight_card")
+
+    def test_change_options(self):
+        card = self.param.parse_reweight_card(
+            "change backend auto\nchange mode LO\nchange helicity False\n"
+            "change fptype d\nchange max_batch 10\nchange verbose T\n" + self.LAUNCH)
+        options = card.driver_options
+        self.assertEqual((options.helicity_reweighting, options.max_batch, options.verbose),
+                         (False, 10, True))
+        self.assertEqual(card.baseline_fptype, "d")
+
+    def test_thread_options(self):
+        options = self.param.parse_reweight_card(
+            "change nb_core 4\nchange min_batch 500\n" + self.LAUNCH).driver_options
+        self.assertEqual((options.nb_core, options.min_batch), (4, 500))
+        # auto (or 0) and an unset min_batch leave the generation's defaults
+        options = self.param.parse_reweight_card(
+            "change nb_core auto\n" + self.LAUNCH).driver_options
+        self.assertEqual((options.nb_core, options.min_batch), (None, None))
+        self.assertIsNone(self.param.parse_reweight_card(
+            "change nb_core 0\n" + self.LAUNCH).driver_options.nb_core)
+        with self.assertRaisesRegex(ValueError, "nb_core"):
+            self.param.parse_reweight_card("change nb_core many\n" + self.LAUNCH)
+        with self.assertRaisesRegex(ValueError, "preamble"):
+            self.param.parse_reweight_card(self.LAUNCH + "change min_batch 10\n")
+        # no former "set" spelling, and parameter reweighting only
+        self.assertNotIn("nb_core", self.param.DEPRECATED_SET_NAMES)
+        with self.assertRaisesRegex(ValueError, "parameter"):
+            self.pdf.parse_pdf_reweight_card("change nb_core 2\nlaunch\nset PDF_SET X\n")
+
+    def test_thread_options_are_applied(self):
+        calls = []
+
+        class Driver:
+            def __getattr__(self, name):
+                return lambda *args: calls.append((name,) + args)
+
+        self.param.parse_reweight_card(
+            "change nb_core 3\nchange min_batch 50\n" + self.LAUNCH).driver_options.apply(Driver())
+        self.assertEqual(calls, [("set_threads", 3), ("set_min_batch", 50)])
+
+    def test_old_spellings_warn(self):
+        with self.assertLogs("madgraph7", "WARNING") as logs:
+            card = self.param.parse_reweight_card(
+                "set FPTYPE f\nset helicity_reweighting false\n" + self.LAUNCH)
+        self.assertEqual((card.baseline_fptype, card.driver_options.helicity_reweighting),
+                         ("f", False))
+        self.assertTrue(all("deprecated" in line for line in logs.output))
+
+    def test_classic_only_lines_are_errors(self):
+        for line in ("change output 2.0", "change mode NLO", "change rwgt_dir /tmp",
+                     "change particle_in_density_matrix 3", "change process p p > z --add",
+                     "change backend classic"):
+            with self.assertRaisesRegex(ValueError, "classic", msg=line):
+                self.param.parse_reweight_card(line + "\n" + self.LAUNCH)
+            with self.assertRaisesRegex(ValueError, "classic", msg=line):
+                self.pdf.parse_pdf_reweight_card(line + "\nlaunch\nset PDF_SET X\n")
+
+    def test_pdf_lines_in_a_parameter_card(self):
+        for card in ("change mode pdf\n" + self.LAUNCH, self.LAUNCH + "set PDF_SET X\n",
+                     "change alpha_s_power 2\n" + self.LAUNCH):
+            with self.assertRaisesRegex(ValueError, "pdf", msg=card):
+                self.param.parse_reweight_card(card)
+
+    def test_pdf_card_options(self):
+        with self.assertLogs("madgraph7", "WARNING"):
+            card = self.pdf.parse_pdf_reweight_card(
+                "change mode pdf\nchange alpha_s_power 2\nset VERBOSE true\n"
+                "change scale_variation_output all\nlaunch\nset PDF_SET X\n")
+        self.assertEqual((card.alpha_s_power, card.verbose, card.scale_variation_output),
+                         (2, True, "all"))
+        for line in ("change helicity False", "change fptype d", "set MASS 6 175"):
+            with self.assertRaisesRegex(ValueError, "parameter", msg=line):
+                self.pdf.parse_pdf_reweight_card(line + "\nlaunch\nset PDF_SET X\n")
+
+    def test_generic_sections(self):
+        sections = self.generic.split_by_mode(
+            "change helicity False\n" + self.LAUNCH
+            + "change mode pdf\nlaunch\nset PDF_SET X\nchange mode LO\nlaunch\nset MT 170\n")
+        self.assertEqual(list(sections), ["parameter", "pdf"])
+        self.assertEqual(sections["parameter"].count("launch"), 2)
+        self.assertIn("change helicity False", sections["parameter"])
+        # a preamble without launches is dropped next to another section
+        sections = self.generic.split_by_mode("change helicity False\nchange mode pdf\nlaunch\n")
+        self.assertEqual(list(sections), ["pdf"])
+        # no mode line: a plain parameter card
+        self.assertEqual(list(self.generic.split_by_mode(self.LAUNCH)), ["parameter"])
+        with self.assertRaisesRegex(ValueError, "classic"):
+            self.generic.split_by_mode("change mode NLO\n" + self.LAUNCH)
+
+    def test_named_parameters(self):
+        card = self.param.expand_scans(self.param.parse_reweight_card(
+            "launch --rwgt_name=a\nset MT scan:[170, 175]\nset param_card wt 2\n"
+            "set aewm1 132\nset MASS 25 130\n"))
+        launches = card.groups[0].launches
+        self.assertTrue(self.param.has_parameter_names(launches))
+        with self.assertRaises(ValueError):
+            self.param.render_launches(launches)
+        text = self.param.render_launches(
+            self.param.resolve_parameter_names(launches, self.PARAM_CARD))
+        self.assertEqual(text.splitlines(), [
+            "launch --rwgt_name=a_1", "set MASS 6 170", "set DECAY 6 2",
+            "set SMINPUTS 1 132", "set MASS 25 130",
+            "launch --rwgt_name=a_2", "set MASS 6 175", "set DECAY 6 2",
+            "set SMINPUTS 1 132", "set MASS 25 130"])
+        with self.assertRaisesRegex(ValueError, "no parameter named"):
+            self.param.resolve_parameter_names(
+                self.param.parse_reweight_card("launch\nset MW 80\n").groups[0].launches,
+                self.PARAM_CARD)
+
+    def test_bang_comments(self):
+        """'!' starts a comment, as '#' does (the default card's examples
+        use both), at the start of a line or after whitespace."""
+        card = self.param.expand_scans(self.param.parse_reweight_card(
+            "! a comment line\nlaunch --rwgt_name=a  ! first weight\n"
+            "set MT 173 ! named\nset MASS 25 scan:[120, 130]   ! one weight per value\n"))
+        launches = card.groups[0].launches
+        self.assertEqual([b.name for b in launches], ["a_1", "a_2"])
+        self.assertEqual(launches[0].sets, [("MT", None, "173"), ("MASS", "25", "120")])
+        pdf = self.pdf.parse_pdf_reweight_card(
+            "change mode pdf ! PDF section\nlaunch\nset PDF_SET CT18NNLO ! or BEAM1_PDF_SET\n")
+        self.assertEqual(len(pdf.launches), 1)
+        self.assertEqual(self.generic.card_kind("launch\nset PDF_SET X ! comment\n"), "pdf")
+        self.assertEqual(self.generic.strip_comment("set A 1 # x ! y"), "set A 1 ")
+        self.assertEqual(self.generic.strip_comment("set A b!c"), "set A b!c")
+
+    def test_multi_index_parameters(self):
+        """'set NMIX 1 2 0.5': every index is kept, also when the
+        parameter is given by its param_card name."""
+        param_card = ("BLOCK NMIX #\n  1 1 9.0e-01 # NMIX11\n  1 2 1.0e-01 # NMIX12\n"
+                      "BLOCK MASS #\n  6 1.73e+02 # MT\n")
+        card = self.param.expand_scans(self.param.parse_reweight_card(
+            "launch --rwgt_name=a\nset NMIX 1  2 0.5\nset MASS 6 175\nset nmix11 0.3\n"
+            "launch --rwgt_name=b\nset NMIX 1 2 scan:[0.1, 0.2]\n"))
+        launches = self.param.resolve_parameter_names(card.groups[0].launches, param_card)
+        self.assertEqual(self.param.render_launches(launches).splitlines(), [
+            "launch --rwgt_name=a", "set NMIX 1 2 0.5", "set MASS 6 175", "set NMIX 1 1 0.3",
+            "launch --rwgt_name=b_1", "set NMIX 1 2 0.1",
+            "launch --rwgt_name=b_2", "set NMIX 1 2 0.2"])
+
+    def test_card_kind(self):
+        kind = self.generic.card_kind
+        self.assertEqual(kind(self.LAUNCH), "param")
+        self.assertEqual(kind("change fptype d\n" + self.LAUNCH), "param")
+        self.assertEqual(kind("launch\nset PDF_SET X\n"), "pdf")
+        self.assertEqual(kind("change alpha_s_power 2\nlaunch\n"), "pdf")
+        self.assertEqual(kind("change mode pdf\nlaunch\n"), "generic")
+        self.assertEqual(kind("set mode parameter\n" + self.LAUNCH), "generic")
+
+    def test_classic_options_in_sync_with_reweight_interface(self):
+        sys.path.insert(0, str(MG_ROOT))
+        try:
+            from madgraph.interface.reweight_interface import (
+                REWEIGHT_CHANGE_OPTIONS, DENSITY_CHANGE_OPTIONS)
+            from madgraph.various.reweight_backend import SHARED_CHANGE_OPTIONS
+        finally:
+            sys.path.remove(str(MG_ROOT))
+        self.assertEqual(self.generic.CLASSIC_ONLY_CHANGE_OPTIONS,
+                         (REWEIGHT_CHANGE_OPTIONS | DENSITY_CHANGE_OPTIONS)
+                         - SHARED_CHANGE_OPTIONS)
+
+
+@unittest.skipUnless(madtrex_is_installed(), "madspace/madtrex are not installed")
+class HelicityReweightingDefaultTest(unittest.TestCase):
+    """Driver and TupperWare default to helicity reweighting."""
+
+    def test_defaults_and_alias(self):
+        result = subprocess.run(
+            [sys.executable, "-I", "-c", textwrap.dedent(f"""
+                import sys
+                sys.path[:0] = [{str(MG_ROOT / "madtrex" / "install")!r},
+                                {str(MG_ROOT / "madspace" / "install")!r}]
+                import madtrex
+                d = madtrex.Driver()
+                t = madtrex.TupperWare()
+                out = [d.helicity_reweighting(), t.helicity_reweighting()]
+                d.set_use_helicity(False)
+                t.set_helicity_reweighting(False)
+                out += [d.helicity_reweighting(), d.use_helicity(), t.helicity_reweighting()]
+                print(out)
+            """)],
+            capture_output=True, text=True,
+        )
+        self.assertEqual(result.returncode, 0, result.stderr)
+        self.assertEqual(result.stdout.strip(), "[True, True, False, False, False]")
+
+
+@unittest.skipUnless(madtrex_is_installed(), "madspace/madtrex are not installed")
+class BatchSplittingTest(unittest.TestCase):
+    """WareHouse.batch_size_for(): the events are split across the threads,
+    in batches of at least min_batch events; an explicit max_batch wins."""
+
+    def test_batch_size_for(self):
+        result = subprocess.run(
+            [sys.executable, "-I", "-c", textwrap.dedent(f"""
+                import sys
+                sys.path[:0] = [{str(MG_ROOT / "madtrex" / "install")!r},
+                                {str(MG_ROOT / "madspace" / "install")!r}]
+                import madtrex
+                house = madtrex.WareHouse()
+                npos = house.batch_size_for(10 ** 6)  # nothing set: no splitting
+                house.set_threads(8)
+                out = [house.batch_size_for(10 ** 6) == npos]
+                house.set_min_batch(1000)
+                out += [house.batch_size_for(n) for n in (80000, 4000, 2000)]
+                out.append(house.batch_size_for(1999) == npos)  # under 2 * min_batch
+                house.set_threads(1)
+                out.append(house.batch_size_for(80000) == npos)  # one thread
+                house.set_threads(8).set_max_batch(300)
+                out.append(house.batch_size_for(80000))
+                d = madtrex.Driver().set_threads(4).set_min_batch(7)
+                out += [d.threads(), d.min_batch()]
+                print(out)
+            """)],
+            capture_output=True, text=True,
+        )
+        self.assertEqual(result.returncode, 0, result.stderr)
+        self.assertEqual(result.stdout.strip(),
+                         "[True, 10000, 1000, 1000, True, True, 300, 4, 7]")
+
+
+if __name__ == "__main__":
+    unittest.main()
+
+
+class GridpackSourceBuildTest(unittest.TestCase):
+    """A gridpack made with include_madtrex_source: bin/reweight builds
+    madtrex from the included sources when there is no install, and
+    rebuilds it when the install was compiled against another madspace.
+    madtrex/install.py is a stand-in here, recording how it was called."""
+
+    def setUp(self):
+        self.tmp = tempfile.TemporaryDirectory()
+        self.root = Path(self.tmp.name)
+        (self.root / "bin").mkdir()
+        shutil.copy(SCRIPT, self.root / "bin" / "reweight")
+        (self.root / "madtrex").mkdir()
+        (self.root / "madtrex" / "install.py").write_text(textwrap.dedent("""
+            import sys
+            print("INSTALLER", *sys.argv[1:])
+            sys.exit(3)
+        """))
+        self.include = self.root / "madspace" / "install" / "madspace" / "include"
+        (self.include / "madspace").mkdir(parents=True)
+        (self.include / "madspace" / "umami.h").write_text("// v1\n")
+
+    def tearDown(self):
+        self.tmp.cleanup()
+
+    def run_reweight(self):
+        return subprocess.run(
+            [sys.executable, "-I", str(self.root / "bin" / "reweight"), "--help"],
+            cwd=self.tmp.name, capture_output=True, text=True,
+        )
+
+    def install_madtrex(self, digest_of):
+        """A madtrex install whose recorded madspace headers are those of
+        the directory digest_of."""
+        sys.path.insert(0, str(MG_ROOT / "madtrex" / "madtrex"))
+        try:
+            import _compat
+        finally:
+            sys.path.pop(0)
+        package = self.root / "madtrex" / "install" / "madtrex"
+        package.mkdir(parents=True)
+        shutil.copy(MG_ROOT / "madtrex" / "madtrex" / "_compat.py", package)
+        (package / "_build_info.json").write_text(
+            '{"madspace_headers": "%s"}' % _compat.headers_digest(digest_of))
+
+    def test_no_install_builds_from_the_sources(self):
+        result = self.run_reweight()
+        self.assertIn("this gridpack has no madtrex install: building it", result.stdout)
+        self.assertIn("INSTALLER --yes\n", result.stdout)
+        self.assertIn("madtrex installation failed", result.stderr)
+
+    def test_install_for_another_madspace_is_rebuilt(self):
+        other = self.root / "other"
+        (other / "madspace").mkdir(parents=True)
+        (other / "madspace" / "umami.h").write_text("// v0\n")
+        self.install_madtrex(other)
+        result = self.run_reweight()
+        self.assertIn("built against another madspace", result.stdout)
+        self.assertIn("INSTALLER --yes --clean", result.stdout)
+
+    def test_matching_install_is_not_rebuilt(self):
+        self.install_madtrex(self.include)
+        self.assertNotIn("INSTALLER", self.run_reweight().stdout)
+
+    def test_precompiled_madspace_cannot_build_madtrex(self):
+        shutil.rmtree(self.include)
+        result = self.run_reweight()
+        self.assertNotIn("INSTALLER", result.stdout)
+        self.assertIn("include_madspace_source = true", result.stderr)
diff --git a/tests/unit_tests/various/test_misc.py b/tests/unit_tests/various/test_misc.py
index ccbc0cc2c4..e173a5f958 100755
--- a/tests/unit_tests/various/test_misc.py
+++ b/tests/unit_tests/various/test_misc.py
@@ -253,6 +253,32 @@ def fileno(self):
             self.assertEqual(proc.wait(), 0)
 
 
+_HEPTOOLS_LHAPDF = os.path.join(os.path.dirname(os.path.dirname(os.path.dirname(
+    os.path.dirname(os.path.abspath(__file__))))), 'HEPTools', 'lhapdf6_py3', 'bin', 'lhapdf-config')
+
+
+@unittest.skipUnless(os.path.exists(_HEPTOOLS_LHAPDF), 'no HEPTools lhapdf6_py3')
+class TEST_import_python_lhapdf(unittest.TestCase):
+    """The HEPTools lhapdf python module has no rpath to libLHAPDF: setting
+    LD_LIBRARY_PATH from inside the process is too late for the loader, so
+    import_python_lhapdf loads the library itself first."""
+
+    def test_import_without_ld_library_path(self):
+        import sys
+        root = os.path.dirname(os.path.dirname(os.path.dirname(_HEPTOOLS_LHAPDF)))
+        env = dict(os.environ)
+        env.pop('LD_LIBRARY_PATH', None)
+        env.pop('PYTHONPATH', None)
+        code = ("import sys; sys.path.insert(0, %r)\n"
+                "import madgraph.various.misc as misc\n"
+                "module = misc.import_python_lhapdf(%r)\n"
+                "print(module.version())\n" % (root, _HEPTOOLS_LHAPDF))
+        result = subprocess.run([sys.executable, '-c', code], env=env,
+                                capture_output=True, text=True)
+        self.assertEqual(result.returncode, 0, result.stdout + result.stderr)
+        self.assertRegex(result.stdout.strip().splitlines()[-1], r'^\d+\.\d+')
+
+
 class TEST_pythia8_main164(unittest.TestCase):
     """main164, the Pythia8 driver of the LO shower, is found or compiled with
     the Makefile of the Pythia8 examples directory (a stub here)."""
diff --git a/tests/unit_tests/various/test_reweight_backend.py b/tests/unit_tests/various/test_reweight_backend.py
new file mode 100644
index 0000000000..28a1f6ef67
--- /dev/null
+++ b/tests/unit_tests/various/test_reweight_backend.py
@@ -0,0 +1,471 @@
+################################################################################
+#
+# Copyright (c) 2026 The MadGraph7 Development team and Contributors
+#
+# This file is a part of the MadGraph7 project, an application which
+# automatically generates Feynman diagrams and matrix elements for arbitrary
+# high-energy processes in the Standard Model and beyond.
+#
+# It is subject to the MadGraph7 license which should accompany this
+# distribution.
+#
+# For more information, visit madgraph.phys.ucl.ac.be and amcatnlo.web.cern.ch
+#
+################################################################################
+"""madgraph/various/reweight_backend.py: which backend (the classic module or
+MadtRex) runs a reweight card. The environments are built by hand."""
+
+import os
+import shutil
+import subprocess
+import sys
+import tempfile
+import textwrap
+import unittest
+import unittest.mock
+from pathlib import Path
+
+from madgraph.various import reweight_backend as rb
+
+MG_ROOT = Path(__file__).resolve().parents[3]
+
+LAUNCH = "launch --rwgt_name=mt\nset MASS 6 175\n"
+
+
+def env(**kwargs):
+    """An MG7 lhe_npy output where both backends can run, unless overridden."""
+    values = dict(output_type='mg7', nlo_events=False, output_format='lhe_npy',
+                  classic_available=True, madtrex_installed=True)
+    values.update(kwargs)
+    return rb.Env(**values)
+
+
+def select(card, environment=None, requested=None, plugin=None):
+    return rb.select_backend(requested, rb.classify_card(card),
+                             environment or env(), plugin=plugin)
+
+
+class ClassifyCardTest(unittest.TestCase):
+
+    def test_plain_card_is_shared(self):
+        traits = rb.classify_card("change helicity False\n" + LAUNCH)
+        self.assertEqual(traits.classic_only, [])
+        self.assertEqual(traits.madtrex_only, [])
+        self.assertIsNone(traits.backend)
+        self.assertEqual(traits.madtrex_kind, 'param')
+        self.assertEqual(traits.launches, 1)
+
+    def test_backend_line(self):
+        traits = rb.classify_card("# comment\nchange backend MadtRex # why not\n" + LAUNCH)
+        self.assertEqual((traits.backend, traits.backend_line), ('madtrex', 2))
+
+    def test_unknown_backend(self):
+        with self.assertRaisesRegex(rb.ReweightBackendError, 'unknown reweighting backend'):
+            rb.classify_card("change backend fortran\n" + LAUNCH)
+
+    def test_two_backends(self):
+        with self.assertRaisesRegex(rb.ReweightBackendError, 'two backends'):
+            rb.classify_card("change backend classic\nchange backend madtrex\n" + LAUNCH)
+
+    def test_classic_only_lines_are_numbered(self):
+        card = ("change output 2.0\nchange rwgt_dir /tmp/x\nchange mode NLO\n"
+                "change process p p > z --add\n" + LAUNCH + "/path/to/param_card.dat\n")
+        traits = rb.classify_card(card)
+        self.assertEqual([d.line for d in traits.classic_only], [1, 2, 3, 4, 7])
+        self.assertEqual(traits.madtrex_only, [])
+        self.assertFalse(traits.density)
+
+    def test_density(self):
+        traits = rb.classify_card("change particle_in_density_matrix 3\n" + LAUNCH)
+        self.assertTrue(traits.density)
+        self.assertEqual([d.line for d in traits.classic_only], [1])
+
+    def test_madtrex_only_lines(self):
+        card = ("change fptype d\nset helicity_reweighting false\nchange verbose true\n"
+                + LAUNCH + "set MUF_FACTOR 2\n")
+        traits = rb.classify_card(card)
+        self.assertEqual([d.line for d in traits.madtrex_only], [1, 2, 3, 6])
+        self.assertEqual(traits.classic_only, [])
+        self.assertEqual(traits.madtrex_kind, 'pdf')
+
+    def test_madtrex_kinds(self):
+        self.assertEqual(rb.classify_card("change fptype d\n" + LAUNCH).madtrex_kind, 'param')
+        self.assertEqual(rb.classify_card("launch\nset PDF_SET X\n").madtrex_kind, 'pdf')
+        self.assertEqual(rb.classify_card("change scale_variation_output all\nlaunch\n")
+                         .madtrex_kind, 'pdf')
+        self.assertEqual(rb.classify_card("change mode pdf\nlaunch\nset PDF_SET X\n")
+                         .madtrex_kind, 'generic')
+        self.assertEqual(rb.classify_card("set mode parameter\n" + LAUNCH).madtrex_kind,
+                         'generic')
+        # LO is the classic default: both backends understand it
+        traits = rb.classify_card("change mode LO\n" + LAUNCH)
+        self.assertEqual((traits.classic_only, traits.madtrex_only), ([], []))
+
+    def test_bang_comments(self):
+        """A '!' comment does not hide what a line asks for."""
+        traits = rb.classify_card("launch ! first\nset PDF_SET 331700 ! NNPDF\n")
+        self.assertEqual([d.line for d in traits.madtrex_only], [2])
+        self.assertEqual(traits.madtrex_kind, 'pdf')
+        self.assertEqual(rb.classify_card("! launch\n").launches, 0)
+
+    def test_named_parameters_are_shared(self):
+        traits = rb.classify_card("launch\nset MT 175\nset param_card mass 6 175\n")
+        self.assertEqual((traits.classic_only, traits.madtrex_only), ([], []))
+
+    def test_strip_backend_line(self):
+        card = "change backend auto\nchange helicity False\n  CHANGE BACKEND classic\n" + LAUNCH
+        stripped = rb.strip_backend_line(card)
+        self.assertEqual(len(stripped.splitlines()), len(card.splitlines()))
+        self.assertIsNone(rb.classify_card(stripped).backend)
+        self.assertIn("change helicity False", stripped)
+        self.assertIn("launch --rwgt_name=mt", stripped)
+
+
+class AutoRulesTest(unittest.TestCase):
+    """The auto rules, in order."""
+
+    def test_1_mixed_directives_are_an_error(self):
+        with self.assertRaises(rb.ReweightBackendError) as ctx:
+            select("change output 2.0\nchange fptype d\n" + LAUNCH)
+        text = str(ctx.exception)
+        self.assertIn('line 1', text)
+        self.assertIn('line 2', text)
+
+    def test_2_classic_only_directives(self):
+        decision = select("change output 2.0\n" + LAUNCH)
+        self.assertEqual((decision.backend, decision.card_kind), ('classic', 'classic'))
+        self.assertIn('line 1', decision.describe())
+
+    def test_2_density(self):
+        decision = select("change particle_in_density_matrix 3\n" + LAUNCH)
+        self.assertEqual((decision.backend, decision.card_kind), ('classic', 'density'))
+
+    def test_2_plugin(self):
+        decision = select(LAUNCH, plugin='my_plugin')
+        self.assertEqual(decision.backend, 'classic')
+        self.assertIn('my_plugin', decision.describe())
+
+    def test_2_classic_only_without_classic(self):
+        with self.assertRaisesRegex(rb.ReweightBackendError, 'f2py'):
+            select("change output 2.0\n" + LAUNCH,
+                   env(classic_available=False, classic_reason='needs f2py'))
+
+    def test_3_madtrex_only_directives(self):
+        decision = select("change fptype d\n" + LAUNCH,
+                          env(classic_available=False))
+        self.assertEqual((decision.backend, decision.card_kind), ('madtrex', 'param'))
+
+    def test_3_madtrex_only_without_madtrex(self):
+        with self.assertRaisesRegex(rb.ReweightBackendError, 'not installed'):
+            select("change fptype d\n" + LAUNCH, env(madtrex_installed=False))
+
+    def test_4_madevent_output(self):
+        decision = select(LAUNCH, env(output_type='madevent', output_format=None))
+        self.assertEqual(decision.backend, 'classic')
+        self.assertIn('MG7', decision.describe())
+
+    def test_4_nlo_events(self):
+        decision = select(LAUNCH, env(output_type='aMC@NLO', nlo_events=True))
+        self.assertEqual(decision.backend, 'classic')
+
+    def test_4_madevent_without_classic(self):
+        with self.assertRaises(rb.ReweightBackendError):
+            select(LAUNCH, env(output_type='madevent', classic_available=False,
+                               classic_reason='needs f2py'))
+
+    def test_5_madtrex_preferred(self):
+        for output_format in ('lhe', 'lhe_npy'):
+            decision = select(LAUNCH, env(output_format=output_format))
+            self.assertEqual(decision.backend, 'madtrex', output_format)
+            self.assertEqual(decision.source, 'auto')
+            self.assertEqual(
+                decision.describe(),
+                'Reweighting with madtrex (auto: MG7 output, madtrex installed, '
+                'no classic-only directives in reweight_card.dat)')
+
+    def test_6_classic_when_madtrex_unusable(self):
+        decision = select(LAUNCH, env(madtrex_installed=False))
+        self.assertEqual(decision.backend, 'classic')
+        self.assertIn('not installed', decision.describe())
+        decision = select(LAUNCH, env(output_format='compact_npy'))
+        self.assertEqual(decision.backend, 'classic')
+        self.assertIn('compact_npy', decision.describe())
+
+    def test_7_nothing_available(self):
+        with self.assertRaisesRegex(rb.ReweightBackendError, 'no backend'):
+            select(LAUNCH, env(madtrex_installed=False, classic_available=False,
+                               classic_reason='needs f2py'))
+
+
+class ForcedBackendTest(unittest.TestCase):
+
+    def test_madtrex_on_madevent(self):
+        with self.assertRaisesRegex(rb.ReweightBackendError, 'MG7 output'):
+            select(LAUNCH, env(output_type='madevent'), requested='madtrex')
+
+    def test_madtrex_on_nlo(self):
+        with self.assertRaisesRegex(rb.ReweightBackendError, 'NLO'):
+            select(LAUNCH, env(nlo_events=True), requested='madtrex')
+
+    def test_madtrex_on_compact_npy(self):
+        with self.assertRaisesRegex(rb.ReweightBackendError, 'compact_npy'):
+            select(LAUNCH, env(output_format='compact_npy'), requested='madtrex')
+
+    def test_madtrex_not_installed(self):
+        with self.assertRaisesRegex(rb.ReweightBackendError, 'not installed'):
+            select(LAUNCH, env(madtrex_installed=False), requested='madtrex')
+
+    def test_madtrex_with_classic_only_directive(self):
+        with self.assertRaisesRegex(rb.ReweightBackendError, 'line 1'):
+            select("change output 2.0\n" + LAUNCH, requested='madtrex')
+
+    def test_madtrex_with_plugin(self):
+        with self.assertRaisesRegex(rb.ReweightBackendError, 'plugin'):
+            select(LAUNCH, requested='madtrex', plugin='my_plugin')
+
+    def test_classic_with_madtrex_only_directive(self):
+        with self.assertRaisesRegex(rb.ReweightBackendError, 'line 1'):
+            select("change fptype d\n" + LAUNCH, requested='classic')
+
+    def test_classic_not_available(self):
+        with self.assertRaisesRegex(rb.ReweightBackendError, 'f2py'):
+            select(LAUNCH, env(classic_available=False, classic_reason='needs f2py'),
+                   requested='classic')
+
+    def test_forced_classic_on_mg7(self):
+        decision = select(LAUNCH, requested='classic')
+        self.assertEqual((decision.backend, decision.source), ('classic', 'requested'))
+
+    def test_every_problem_is_listed(self):
+        with self.assertRaises(rb.ReweightBackendError) as ctx:
+            select("change output 2.0\n" + LAUNCH,
+                   env(output_type='madevent', madtrex_installed=False),
+                   requested='madtrex')
+        self.assertEqual(len(ctx.exception.reasons), 3)
+
+
+class PrecedenceTest(unittest.TestCase):
+
+    def test_card_backend(self):
+        decision = select("change backend classic\n" + LAUNCH)
+        self.assertEqual(decision.backend, 'classic')
+        self.assertIn('change backend', decision.source)
+        self.assertEqual(decision.warnings, [])
+
+    def test_card_auto_is_auto(self):
+        self.assertEqual(select("change backend auto\n" + LAUNCH).backend, 'madtrex')
+
+    def test_switch_beats_card(self):
+        decision = select("change backend classic\n" + LAUNCH,
+                          requested=[('reweight switch', 'madtrex')])
+        self.assertEqual(decision.backend, 'madtrex')
+        self.assertEqual(len(decision.warnings), 1)
+        self.assertIn('reweight switch', decision.warnings[0])
+        self.assertIn('change backend', decision.warnings[0])
+
+    def test_cli_beats_switch_and_card(self):
+        decision = select("change backend madtrex\n" + LAUNCH,
+                          requested=[('--reweight-backend', 'classic'),
+                                     ('reweight switch', 'madtrex')])
+        self.assertEqual((decision.backend, decision.source),
+                         ('classic', '--reweight-backend'))
+        self.assertEqual(len(decision.warnings), 2)
+
+    def test_auto_and_none_requests_are_not_explicit(self):
+        decision = select("change backend classic\n" + LAUNCH,
+                          requested=[('--reweight-backend', 'auto'),
+                                     ('reweight switch', None)])
+        self.assertEqual(decision.backend, 'classic')
+        self.assertEqual(decision.warnings, [])
+
+    def test_agreeing_requests_do_not_warn(self):
+        decision = select("change backend madtrex\n" + LAUNCH,
+                          requested=[('reweight switch', 'madtrex')])
+        self.assertEqual(decision.warnings, [])
+
+    def test_unknown_request(self):
+        with self.assertRaises(rb.ReweightBackendError):
+            select(LAUNCH, requested='fortran')
+
+    def test_conflicting_card_directives_under_auto(self):
+        with self.assertRaisesRegex(rb.ReweightBackendError, 'mixes'):
+            select("change particle_in_density_matrix 3\nset PDF_SET X\n" + LAUNCH)
+
+
+class ProbeEnvironmentTest(unittest.TestCase):
+
+    def setUp(self):
+        self.dir = tempfile.mkdtemp()
+        os.mkdir(os.path.join(self.dir, 'Cards'))
+
+    def tearDown(self):
+        shutil.rmtree(self.dir)
+
+    def test_mg7(self):
+        with open(os.path.join(self.dir, 'Cards', 'run_card.toml'), 'w') as f:
+            f.write('[run]\noutput_format = "lhe" # options\n')
+        madtrex = os.path.join(self.dir, 'mtx')
+        os.makedirs(os.path.join(madtrex, 'madtrex'))
+        found = rb.probe_environment(self.dir, {'mg5_path': '/mg5'},
+                                     madtrex_install_dir=madtrex, has_f2py=True)
+        self.assertEqual((found.output_type, found.output_format, found.nlo_events),
+                         ('mg7', 'lhe', False))
+        self.assertTrue(found.classic_available)
+        self.assertTrue(found.madtrex_installed)
+
+    def test_madevent_without_f2py(self):
+        os.mkdir(os.path.join(self.dir, 'bin'))
+        open(os.path.join(self.dir, 'bin', 'madevent'), 'w').close()
+        found = rb.probe_environment(self.dir, {}, madtrex_install_dir=self.dir,
+                                     has_f2py=False)
+        self.assertEqual(found.output_type, 'madevent')
+        self.assertFalse(found.classic_available)
+        self.assertIn('f2py', found.classic_reason)
+        self.assertFalse(found.madtrex_installed)
+
+    def test_amcatnlo(self):
+        open(os.path.join(self.dir, 'Cards', 'FKS_params.dat'), 'w').close()
+        found = rb.probe_environment(self.dir, {}, has_f2py=True)
+        self.assertEqual(found.output_type, 'aMC@NLO')
+        self.assertTrue(found.nlo_events)
+
+    def test_no_mg5_path(self):
+        found = rb.probe_environment(self.dir, {'mg5_path': None}, has_f2py=True)
+        self.assertFalse(found.classic_available)
+
+
+class ClassicOptionsTest(unittest.TestCase):
+    """The classic option names come from reweight_interface."""
+
+    def test_every_classic_option_is_classified(self):
+        from madgraph.interface.reweight_interface import (REWEIGHT_CHANGE_OPTIONS,
+                                                           DENSITY_CHANGE_OPTIONS)
+        for name in REWEIGHT_CHANGE_OPTIONS | DENSITY_CHANGE_OPTIONS:
+            value = 'auto' if name == 'backend' else 'x'
+            traits = rb.classify_card("change %s %s\n%s" % (name, value, LAUNCH))
+            if name in rb.SHARED_CHANGE_OPTIONS:
+                self.assertEqual(traits.classic_only, [], name)
+            else:
+                self.assertEqual([d.line for d in traits.classic_only], [1], name)
+
+
+class ClassicSideTest(unittest.TestCase):
+    """The classic module reads the shared vocabulary: it accepts 'change
+    backend classic|auto', and says clearly what only MadtRex does."""
+
+    def change(self, line):
+        from madgraph.interface import reweight_interface as ri
+        cmd = ri.ReweightInterface.__new__(ri.ReweightInterface)
+        cmd.has_standalone_dir = False
+        cmd.exitted = True  # nothing to clean up when it is collected
+        return cmd.do_change(line)
+
+    def test_backend_line(self):
+        from madgraph.interface.extended_cmd import Cmd
+        InvalidCmd = Cmd.InvalidCmd
+        self.change('backend classic')
+        self.change('backend AUTO')
+        with self.assertRaisesRegex(InvalidCmd, 'MG7 output'):
+            self.change('backend madtrex')
+        with self.assertRaises(InvalidCmd):
+            self.change('backend fortran')
+
+    def test_madtrex_only_options(self):
+        from madgraph.interface.extended_cmd import Cmd
+        InvalidCmd = Cmd.InvalidCmd
+        for line in ('mode pdf', 'fptype d', 'max_batch 10', 'alpha_s_power 2'):
+            with self.assertRaisesRegex(InvalidCmd, 'MadtRex only', msg=line):
+                self.change(line)
+
+    def test_madtrex_only_options_match_the_selector(self):
+        from madgraph.interface.reweight_interface import MADTREX_ONLY_CHANGE_OPTIONS
+        madtrex_change, _, _ = rb._madtrex_card_vocabulary()
+        self.assertEqual(MADTREX_ONLY_CHANGE_OPTIONS, madtrex_change)
+
+    def test_default_f2py_is_the_running_pythons(self):
+        """f2py3 on the PATH may build for another Python: the f2py installed
+        with this one (e.g. a virtual environment's) comes first."""
+        from madgraph.various import misc
+        tmp = tempfile.mkdtemp()
+        try:
+            venv_bin = os.path.join(tmp, 'venv', 'bin')
+            other_bin = os.path.join(tmp, 'usr', 'bin')
+            for directory, name in ((venv_bin, 'f2py'), (other_bin, 'f2py3')):
+                os.makedirs(directory)
+                path = os.path.join(directory, name)
+                open(path, 'w').close()
+                os.chmod(path, 0o755)
+            with unittest.mock.patch.object(sys, 'executable', os.path.join(venv_bin, 'python')), \
+                    unittest.mock.patch.dict(os.environ, {'PATH': other_bin}):
+                self.assertEqual(misc.default_f2py(), os.path.join(venv_bin, 'f2py'))
+                self.assertTrue(misc.has_f2py())
+                os.remove(os.path.join(venv_bin, 'f2py'))
+                self.assertEqual(misc.default_f2py(), 'f2py3')
+                os.remove(os.path.join(other_bin, 'f2py3'))
+                self.assertIsNone(misc.default_f2py())
+                self.assertFalse(misc.has_f2py())
+        finally:
+            shutil.rmtree(tmp)
+
+    def test_do_reweight_backend_option(self):
+        from madgraph import InvalidCmd
+        from madgraph.interface.common_run_interface import CommonRunCmd
+        cmd = unittest.mock.Mock(InvalidCmd=InvalidCmd, options={'mg5_path': '/mg5'},
+                                 switch={'reweight': 'ON'})
+        with self.assertRaisesRegex(InvalidCmd, 'MG7 output'):
+            CommonRunCmd.do_reweight(cmd, '--backend=madtrex -from_cards')
+        with self.assertRaises(InvalidCmd):
+            CommonRunCmd.do_reweight(cmd, '--backend=fortran -from_cards')
+        # classic: the option is dropped and the reweighting goes on (here it
+        # stops at the missing card)
+        cmd.me_dir = tempfile.mkdtemp()
+        try:
+            self.assertIsNone(CommonRunCmd.do_reweight(cmd, '--backend=classic -from_cards'))
+        finally:
+            shutil.rmtree(cmd.me_dir)
+
+
+class DefaultCardTest(unittest.TestCase):
+    """MG7 outputs get their own default reweight card; madevent keeps its."""
+
+    MG7_CARD = MG_ROOT / 'madgraph' / 'iolibs' / 'template_files' / 'mg7' / \
+        'reweight_card_default.dat'
+
+    def test_mg7_card_runs_with_either_backend(self):
+        traits = rb.classify_card(self.MG7_CARD.read_text())
+        self.assertEqual((traits.classic_only, traits.madtrex_only, traits.backend),
+                         ([], [], None))
+        self.assertEqual(traits.launches, 1)
+        from madgraph.iolibs.template_files.mg7 import param_reweight_card
+        param_reweight_card.parse_reweight_card(self.MG7_CARD.read_text())
+
+    def test_madmatrix_output_copies_the_mg7_card(self):
+        text = (MG_ROOT / 'madmatrix' / 'output.py').read_text()
+        self.assertIn("relative_path_list(mg7_templates, ['reweight_card_default.dat'])", text)
+        classic = (MG_ROOT / 'Template' / 'Common' / 'Cards' / 'reweight_card_default.dat')
+        self.assertIn('change mode NLO', classic.read_text())
+
+
+class ImportHygieneTest(unittest.TestCase):
+    """The selector never imports madtrex or madspace."""
+
+    def test_selector_imports_neither(self):
+        code = textwrap.dedent(f"""
+            import sys
+            sys.path.insert(0, {str(MG_ROOT)!r})
+            from madgraph.various import reweight_backend as rb
+            traits = rb.classify_card("change backend auto\\nlaunch\\nset MT 175\\n")
+            env = rb.probe_environment({str(MG_ROOT)!r}, has_f2py=True)
+            rb.select_backend(None, traits, rb.Env(output_type='madevent'))
+            bad = [m for m in sys.modules if m.split('.')[0] in ('madtrex', 'madspace')
+                   or m.endswith('.mg7.madtrex') or m.endswith('.mg7.madtrex_driver')]
+            assert not bad, bad
+        """)
+        with tempfile.TemporaryDirectory() as tmp:
+            result = subprocess.run([sys.executable, '-I', '-c', code], cwd=tmp,
+                                    capture_output=True, text=True)
+        self.assertEqual(result.returncode, 0, result.stderr)
+
+
+if __name__ == '__main__':
+    unittest.main()