oliviermattelaer · Oct 27, 2026 · Oct 27, 2026
diff --git a/.gitattributes b/.gitattributes
@@ -6,24 +6,13 @@
 .bzrignore export-ignore
 
 # pyamplicol tests and development tooling stay in the repository for its own
-# PyPI releases. The MadGraph7 installer builds the wheel without them, except
-# for the legacy-oracle modules that the wheel bundles.
+# PyPI releases. The MadGraph7 installer builds the wheel without them.
 pyamplicol/tests/ export-ignore
-pyamplicol/tools/developer/** export-ignore
-pyamplicol/tools/developer/legacy_amplicol.py -export-ignore
-pyamplicol/tools/developer/legacy_oracle -export-ignore
-pyamplicol/tools/developer/legacy_oracle/** -export-ignore
-pyamplicol/tools/fft_profiling/ export-ignore
-pyamplicol/umami_example/ export-ignore
-pyamplicol/flake.nix export-ignore
-pyamplicol/flake.lock export-ignore
-pyamplicol/run_fft_profiling_render.sh export-ignore
+pyamplicol/tools/developer/ export-ignore
 
 # Generated pyamplicol data: GitHub collapses these in PR diffs.
 pyamplicol/Cargo.lock linguist-generated
-pyamplicol/flake.lock linguist-generated
 pyamplicol/dependencies/python-runtime-lock.toml linguist-generated
 pyamplicol/tests/fixtures/** linguist-generated
 pyamplicol/src/pyamplicol/assets/selftest/** linguist-generated
 pyamplicol/src/pyamplicol/assets/models/json/** linguist-generated
-pyamplicol/src/pyamplicol/_profiling_campaign/results/*.json linguist-generated
diff --git a/docs/source/conf.py b/docs/source/conf.py
@@ -32,13 +32,7 @@ def _copy_pyamplicol_docs():
         "_config.yml", "_includes", "css", "results", "archive", "arxiv", ".*"
     )
 
-    def ignore(directory, names):
-        # Only the report index is tracked; campaigns are local output.
-        if os.path.basename(directory) == "performance_reports":
-            return {name for name in names if name != "README.md"}
-        return jekyll_only(directory, names)
-
-    shutil.copytree(source, target, ignore=ignore)
+    shutil.copytree(source, target, ignore=jekyll_only)
 
 
 _copy_pyamplicol_docs()

diff --git a/pyamplicol/README.md b/pyamplicol/README.md
@@ -228,152 +228,15 @@ The public C ABI is version 1. Every generated artifact can include standalone
 Python, C11, C++17, Fortran 2008, and dependency-free Rust 2021 drivers backed
 by the wheel-owned static Rusticol SDK.
 
-## Profiling campaigns
+## Performance reports
 
-An installed wheel can populate a self-contained campaign workspace:
-
-```console
-pyamplicol profiling-campaign copy ./pyamplicol-profiling-campaign --force
-cd ./pyamplicol-profiling-campaign
-./steer_performance_campaign.py run \
-  --workers 1 --table matrix --process-id 1 --multiplicity 1 \
-  --color-approximation lc --generation-mode non-union-flow \
-  --generation-engine recurrence --model built_in
-```
-
-That deliberately small real campaign measures only the final-state-
-multiplicity-one `d d~ > Z` recurrence cell. Broader campaign selections are
-intended for dedicated profiling hosts. The
-[documentation](https://madgraph.org/pyamplicol/) covers selection,
-continuation, optional original-AmpliCol comparisons, artifact retention, and
-PDF generation.
-
-The source checkout also contains a thin orchestrator for the dedicated
-FullColor FFT comparison. It delegates generation and timing to the existing
-pyAmpliCol profiling commands and schedules independent measurement children:
-
-```console
-just dev-install --with-legacy-amplicol --with-reference-fft
-.venv/bin/python tools/fft_profiling/fft_profiling.py \
-  --multiplicities 2 3 4 5 \
-  --lines reference-fft amplicol pyamplicol-recurrence pyamplicol-otf madgraph \
-  --cores 8 --candidate-cores 1 \
-  --memory-limit-gib 30 --time-limit-seconds 3600 \
-  --amplicol-root /path/to/AmpliCol \
-  --reference-fft-root /path/to/AllGluonsMultipletFFT \
-  --madgraph-root /path/to/MG5_aMC \
-  --build-amplicol
-```
-
-`--multiplicities` adds the selected values to the persistent fill history and
-defaults to `2 ... 9`. `--lines` similarly adds any combination of
-`reference-fft`, `amplicol`, `pyamplicol-recurrence`, `pyamplicol-otf`, and
-`madgraph`; the two pyAmpliCol groups each schedule their direct and FFT
-companion curves together so they reuse the same generation lane. Dependencies
-are selected automatically.
-Repeating a command against the same output unions both selections, resumes
-unfinished cells, and skips completed cells;
-`--resume` is an explicit alias for that default. `--cores` is the total
-scheduler budget, while `--candidate-cores` is one candidate child's core
-claim and evaluator setting. The memory and time limits are strict per-child
-cutoffs. `--retry` reruns only failed/skipped cells in the active selection.
-`--overwrite` reruns every selected cell and replaces each old result only when
-that cell's worker is about to launch; queued or blocked cells retain their old
-results. Use `--output PATH` for an independent run directory. `--refresh`
-removes only that exact recognized output directory and restarts it, so a
-custom output also scopes the refresh; a path that does not exist simply starts
-cleanly. Without `--output`, fixed and summed
-workloads use separate `IMPLEMENTATION_DOCS/RESULTS/fft-profiling/runs/`
-directories named `cluster-fullcolor-n2-n9` and
-`cluster-fullcolor-helicity-sum-n2-n9`. Refresh also shares the persistent
-MadGraph cache-writer lock and refuses to delete a run while a standalone
-MadGraph profiler is using that cache.
-
-Add `--compare-helicity-sums` for the independent complete physical-helicity-
-sum workload. The fixed-helicity MadGraph lane selects the shared helicity
-through the generated `MATRIX(P,NHEL,IC)` entry point. The summed lane instead
-calls the generated `SMATRIX(P,ANS)` with `USERHEL=-1`; MadGraph applies its
-native IDEN normalization and may reuse its warmed `GOODHEL` pruning. Fixed-
-helicity and summed overlays carry distinct workload identities and cannot be
-mixed. `just dev-install` omits the developer-only AmpliCol and Reference FFT
-repositories unless they are requested with `--with-legacy-amplicol` and
-`--with-reference-fft`; either opt-in also installs the `fft-profiling` Python
-extra into `.venv`. Their profiler roots default to
-`dependencies/checkouts/legacy-amplicol` and
-`dependencies/checkouts/reference-fft`; `--build-amplicol` may build the
-AmpliCol probe once. Both paths can be overridden explicitly with
-`--amplicol-root` and `--reference-fft-root`. The MadGraph root defaults to
-`PYAMPLICOL_MADGRAPH_ROOT` or a recognized developer checkout and may be set
-explicitly with `--madgraph-root`.
-
-The published fixed-helicity MadGraph series currently has measured points
-through `n=5` for pure gluons and `n=6` for
-`d d~ > d d~ + gluons`. Pure-gluon `n=6` retains its measured resource cutoff;
-`n=7..9` are explicit protocol-scope not-applicable cells for both families.
-The independent helicity-sum MadGraph series has measured points for both
-families at `n=2..5`. Every admitted point passed the same-workload numerical
-gate before entering the PDFs.
-
-The rolling plot frontiers are per process and implementation, rather than a
-claim that every curve reaches the same `n`. Both fixed-helicity and helicity-
-sum OTF curves are requested only through final-state `n=6`; beyond that the
-publication protocol retains recurrence, AmpliCol, and Reference FFT where
-applicable. Within that frontier, cutoffs are annotated rather than hidden or
-interpolated. Pure-gluon OTF FFT reached the 3,600 s first-use runtime cap at
-`n=6` and retains its measured `n=5` point. The helicity-sum comparison extends
-the `d d~ > d d~ + gluons` curves through `n=6` using an authenticated isolated
-30 GiB extension. At that point every requested curve measured successfully.
-For OTF, direct and FFT setup took 1,725.8 s and 1,607.9 s, while warmed runtime
-was 5.321 and 3.759 ms/point respectively (a 1.416x FFT speedup). The measured
-peak RSS values were 1.17 and 1.22 GiB.
-
-The plotted *setup time* is deliberately method-specific. pyAmpliCol includes
-artifact generation, a fresh load, the first requested evaluation, and OTF
-family warm-up where applicable. Reference FFT includes its build,
-initialization, and first pass. AmpliCol includes process/color-object
-generation for the fixed workload, or process/raw-library generation and build
-plus the immutable snapshot for the summed workload. Warmed runtime is measured
-separately after those setup boundaries. Resource limits apply individually to
-each child. The retained isolated high-frontier extensions use a 30 GiB
-process-tree guard.
-
-After the first snapshot is published, rendering never waits for workers and
-uses the latest available data. Concurrent renders and refresh publication are
-serialized so an older render cannot replace a newer one:
-
-```console
-python tools/fft_profiling/fft_profiling.py --render
-python tools/fft_profiling/fft_profiling.py --render --compare-helicity-sums
-python tools/fft_profiling/fft_profiling.py --render --output /path/to/run
-```
-
-For a custom output, the saved manifest determines whether the workload is
-fixed-helicity or helicity-summed, so a render does not need to repeat
-`--compare-helicity-sums`. Default outputs still use that flag to select the
-helicity-sum workspace. Default outputs also refresh the corresponding
-canonical PDF; custom outputs keep their PDF inside the selected run directory.
-During a scan, the progress display reports the active cell, total live RSS,
-and occupied core slots.
-
-An older local MadGraph overlay that predates the node-fingerprint field may
-appear only in a nonterminal anytime render, only when its system, machine, and
-Python version match the current workstation. Such plots carry an explicit
-provenance note; the strict terminal publication merger still requires the
-complete host identity produced by a fresh profiler run.
-
-The public performance index retains four selected rendered snapshots. Raw
-JSON, generated tables, attempts, and campaign workspaces stay untracked:
+Selected manual measurement snapshots are published in the upstream repository:
 
 - [MacBook M3 report](https://github.com/MadGraphTeam/pyamplicol/blob/main/docs/performance_reports/macbook_M3_pyAmpliCol.pdf)
 - [AMD EPYC report](https://github.com/MadGraphTeam/pyamplicol/blob/main/docs/performance_reports/EPYC_pyAmpliCol.pdf)
 - [FullColor FFT selected-helicity snapshot](https://github.com/MadGraphTeam/pyamplicol/blob/main/docs/performance_reports/summary_plots_final.pdf)
 - [FullColor FFT helicity-sum snapshot](https://github.com/MadGraphTeam/pyamplicol/blob/main/docs/performance_reports/summary_plots_final_helicity_sum.pdf)
 
-These are manual measurement snapshots rather than release-CI results; raw
-campaign data remains local. The general host report format is reproducible
-from an installed package, while the FullColor FFT snapshots use the source-
-checkout orchestrator above.
-
 ## Documentation
 
 Read the complete [pyAmpliCol documentation](https://madgraph.org/pyamplicol/).