diff --git a/specs/t27f/reel_gf16_nmse.t27 b/specs/t27f/reel_gf16_nmse.t27
index 31bb7a7f75..b00d95892a 100644
--- a/specs/t27f/reel_gf16_nmse.t27
+++ b/specs/t27f/reel_gf16_nmse.t27
@@ -10,8 +10,9 @@
 ; SOURCE: gHashTag/t27 at SOURCE_COMMIT, measured on MEASURED_ON. The protocol spec
 ; specs/benchmarks/gf16_bfloat16_nmse.t27 and its mirror docs/GF16_BFLOAT16_NMSE_PROTOCOL.md, the
 ; manifest schema schemas/nmse-protocol-v1.json, the roadmap docs/SCIENTIFIC_IMPROVEMENT_PLAN.md
-; (PUB-01, PUB-02), and what the repo already holds about GF16. Every cell names its source in
-; SOURCES; a cell with none is a gap and says in MISSING what would ground it.
+; (PUB-01, PUB-02), the PUB-02 run harness specs/benchmarks/gf16_bfloat16_nmse_run.t27 and the
+; manifest it wrote, RUN_MANIFEST (#10021), and what the repo already holds about GF16. Every cell
+; names its source in SOURCES; a cell with none is a gap and says in MISSING what would ground it.
 
 ; BRIEF is the first three sentences of the protocol spec's header, verbatim. route_brief
 ; counts gf16 twice and benchmark once, no other list is hit, so the domain is DOMAIN_RESEARCH.
@@ -19,28 +20,36 @@
 ; LENSES: one row per lens. Steps 1-9 are the formats and why they compare at all; 10-18 the protocol
 ; and the one run it asks for; 19-27 what may be said and what the repo has said so far.
 
-; MEASURED on 2026-10-17:
+; MEASURED on 2026-10-18:
 ;   - grounded: the source is in the repo at SOURCE_COMMIT (path and section in SOURCES).
 ;   - test: t27c test-report and t27b test --check on the Railway lab passed all 7 tests of the
 ;     protocol spec; a planted GF16_MANT_BITS = 8 turned 2 of them red under both, so they are real.
-;   - review: the PRs that wrote the protocol (#1399) and the spec rewrite (#5947) have 0 reviews.
+;     The harness's own tests check its procedure, not the committed manifest, so no step of the
+;     run (16, 17, 25) is marked as tested; on a DELIVER step a test would not count anyway.
+;   - review: the PRs that wrote the protocol (#1399), the spec rewrite (#5947) and the run (#10021)
+;     have 0 reviews.
 ;   - live: GET https://t27.ai/t27/files/<path> returned 200 for the protocol spec and the schema,
-;     and 404 for the doc, FORMAT-SPEC-001.json, gf16_bench_results.json, the crosswalk and the paper
-;     draft; a made-up path returned 404, so 200 is not a catch-all.
+;     and 404 for the doc, FORMAT-SPEC-001.json, gf16_bench_results.json, the crosswalk, the paper
+;     draft, the harness and the manifest; a made-up path returned 404, so 200 is not a catch-all.
 ; The engine asks a research LEARN step for a source link, a PRACTICE step for a review or a passing
 ; test, and a DELIVER step for a review. So: 15 published (sourced, earned, and live), 1 verified
-; (reporting rules: sourced, not live), 6 drafts (practice or deliver steps with neither a test nor a
-; review) and 5 gaps.
+; (reporting rules: sourced, not live), 9 drafts (practice or deliver steps with neither a test nor a
+; review, three of them the run's own cells) and 2 gaps.
 
-; THE GAPS are the point of this reel. PUB-02 asks for one sealed bench/results/nmse-*.json; the repo
-; holds none, so the run, its schema check, the ratio per distribution and the paper's Delta_dB (which
-; PUB-02 forbids before that manifest) are gaps, and so is the chip repo's manifest:
-; gHashTag/tt-trinity-gamma tests/vectors/nmse_gf16_bf16.golden.json at GAMMA_COMMIT pins the seeds
-; but says measured_silicon false and harness_committed false. The only GF16/BF16 numbers in the repo,
-; conformance/gf16_bench_results.json (2026-04-05, no seed, no seal), are drawn as a draft that the
-; protocol itself says does not count.
+; THE RUN. PUB-02 asks for one sealed bench/results/nmse-*.json. Until #10021 the repo held none, and
+; steps 16, 17 and 25 (the run, its schema check, the ratio per distribution) were gaps. They now cite
+; RUN_HARNESS and RUN_MANIFEST, and they are DRAFTS, not verified: each is a DELIVER step, a research
+; DELIVER step is verified by a review, and the run has none. The manifest is a software run on the
+; Railway lab (runner.host_arch x86_64), not a silicon measurement.
+; THE GAPS that remain: the chip repo's manifest, gHashTag/tt-trinity-gamma
+; tests/vectors/nmse_gf16_bf16.golden.json at GAMMA_COMMIT, pins the seeds but says measured_silicon
+; false and harness_committed false; and the paper draft has no Delta_dB yet (PUB-02 allowed one only
+; after the manifest, and attaching the manifest to the draft is still open). The older GF16/BF16
+; numbers, conformance/gf16_bench_results.json (2026-04-05, no seed, no seal), are drawn as a draft
+; that the protocol itself says does not count.
 
-; Not claimed: no NMSE value, no ratio, no winner. No GIF or MP4; stills only, for the chat.
+; Not claimed: this reel quotes no NMSE value, no ratio and no winner; the manifest holds the numbers
+; and is the only place to read them. Nothing here is silicon. No GIF or MP4; stills only, for the chat.
 ; phi^2 + 1/phi^2 = 3 | TRINITY
 
 module reel_gf16_nmse;
@@ -63,6 +72,7 @@ use t27f::topic_engine::SALES_WORDS;
 use t27f::topic_engine::keyword_hits;
 use t27f::topic_engine::route_brief;
 use t27f::topic_engine::step_stage;
+use t27f::topic_engine::earned_status;
 use t27f::topic_reel::count_status;
 use t27f::topic_reel::count_true;
 use t27f::topic_reel::statuses_earned;
@@ -73,11 +83,14 @@ use t27f::topic_reel::stage_slides_bare;
 pub const TOPIC_ID : str = "gf16-nmse";
 pub const SOURCE_REPO : str = "gHashTag/t27";
 pub const SOURCE_SPEC : str = "specs/benchmarks/gf16_bfloat16_nmse.t27";
-pub const SOURCE_COMMIT : str = "de2cff88243c179e5bbf21f8cb06e5da1ab833e8";
+pub const SOURCE_COMMIT : str = "cc80a305dd5fef1cd21f930e1394ae47cbdacbfc";
 pub const GAMMA_REPO : str = "gHashTag/tt-trinity-gamma";
 pub const GAMMA_COMMIT : str = "de484c4ff249356624ec508127fcbab880ee25fc";
 pub const LIVE_BASE : str = "https://t27.ai/t27/files/";
-pub const MEASURED_ON : str = "2026-10-17";
+pub const MEASURED_ON : str = "2026-10-18";
+/// The PUB-02 run (#10021): the harness that wrote the manifest, and the manifest.
+pub const RUN_HARNESS : str = "specs/benchmarks/gf16_bfloat16_nmse_run.t27";
+pub const RUN_MANIFEST : str = "bench/results/nmse-railway-x86_64-0x2545F4914F6CDD1D-2026-10-17.json";
 
 pub const BRIEF : str = "GF16 against bfloat16 NMSE protocol (TRI-NET benchmark 0). The standard comparison of GoldenFloat GF16 and bfloat16 numeric fidelity, as normalised mean squared error over named reference distributions. Protocol only: no silicon number is asserted.";
 /// The domain route_brief gives that brief (see the_domain_is_routed_not_chosen).
@@ -132,8 +145,8 @@ pub const SOURCES : [27]str = [
     "specs/benchmarks/gf16_bfloat16_nmse.t27 test identity_witnesses_hold_in_f64",
     "specs/benchmarks/gf16_bfloat16_nmse.t27 test d_deep_mixture_weights_sum_to_one",
     "specs/benchmarks/gf16_bfloat16_nmse.t27 test nmse_non_negative_and_zero_baseline_undefined",
-    "",
-    "",
+    "specs/benchmarks/gf16_bfloat16_nmse_run.t27 nmse_manifest; bench/results/nmse-railway-x86_64-0x2545F4914F6CDD1D-2026-10-17.json samples_per_distribution, seal_hash",
+    "bench/results/nmse-railway-x86_64-0x2545F4914F6CDD1D-2026-10-17.json schema_version, under schemas/nmse-protocol-v1.json",
     "",
     "specs/benchmarks/gf16_bfloat16_nmse.t27 header, the ratio",
     "docs/GF16_BFLOAT16_NMSE_PROTOCOL.md section 8",
@@ -141,31 +154,28 @@ pub const SOURCES : [27]str = [
     "specs/benchmarks/gf16_bfloat16_nmse.t27 test ratio_direction",
     "conformance/gf16_bench_results.json",
     "docs/GF16_BFLOAT16_NMSE_PROTOCOL.md section 6",
-    "",
+    "bench/results/nmse-railway-x86_64-0x2545F4914F6CDD1D-2026-10-17.json results.*.ratio, from specs/benchmarks/gf16_bfloat16_nmse_run.t27",
     "",
     "docs/WHITEPAPER/gf_paper_v3_imrad_draft.md (PUB-01)"
 ];
 
 /// What would ground each gap; "" for a cell that has its source.
 pub const MISSING : [27]str = [
-    "", "", "", "", "", "", "", "", "", "", "", "", "", "", "",
-    "bench/results/nmse-*.json, the sealed run PUB-02 asks for: the repo holds 0 such files",
-    "a manifest to check against schemas/nmse-protocol-v1.json: none exists",
+    "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "",
     "tt-trinity-gamma nmse_gf16_bf16.golden.json pins seeds only: measured_silicon false, harness_committed false",
-    "", "", "", "", "", "",
-    "an NMSE ratio per distribution: no run has produced one",
-    "a Delta_dB figure: PUB-02 forbids one before the manifest lands",
+    "", "", "", "", "", "", "",
+    "a Delta_dB figure in the paper draft: the manifest is not attached to it yet",
     ""
 ];
 
 /// The measurements per cell (see the header for how each was taken).
-pub const GROUNDED : [27]bool = [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, true, true, true, true, true, true, false, false, true];
+pub const GROUNDED : [27]bool = [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, true, true, true, true, true, true, true, false, true];
 pub const TEST_PASS : [27]bool = [false, false, false, true, true, true, false, false, false, false, false, false, true, true, true, false, false, false, false, false, false, true, false, false, false, false, false];
 pub const REVIEWED : [27]bool = [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false];
 pub const LIVE : [27]bool = [true, true, true, true, true, true, false, false, false, true, true, true, true, true, true, false, false, false, true, false, true, true, false, false, false, false, false];
 
 /// What a host passes the viewer, per cell: the status code and the step number as a label.
-pub const STATUS : [27]u8 = [5, 5, 5, 5, 5, 5, 3, 3, 3, 5, 5, 5, 5, 5, 5, 0, 0, 0, 5, 4, 5, 5, 3, 3, 0, 0, 3];
+pub const STATUS : [27]u8 = [5, 5, 5, 5, 5, 5, 3, 3, 3, 5, 5, 5, 5, 5, 5, 3, 3, 0, 5, 4, 5, 5, 3, 3, 3, 0, 3];
 pub const LABELS : [27]str = ["1", "2", "3", "4", "5", "6", "7", "8", "9", "10", "11", "12", "13", "14", "15", "16", "17", "18", "19", "20", "21", "22", "23", "24", "25", "26", "27"];
 
 /// Slide text per demo scene, checked by topic_reel.t27 slides_fit (index scene*4+row).
@@ -174,7 +184,7 @@ pub const HEAD : [5]str = [
     "Steps 1-9: formats",
     "Witness first",
     "Evidence",
-    "The missing run"
+    "The sealed run"
 ];
 pub const ROWS : [20]str = [
     "", "", "", "",
@@ -184,13 +194,13 @@ pub const ROWS : [20]str = [
     "A failed witness aborts the run",
     "No manifest is written after a failure",
     "7 of 7 protocol tests pass on the lab",
-    "15 published, 1 verified, 6 drafts",
-    "5 gaps: no measured run exists yet",
-    "Measured 2026-10-17",
+    "15 published, 1 verified, 9 drafts",
+    "2 gaps: chip run, paper Delta_dB",
+    "Measured 2026-10-18",
     "PUB-02: one sealed NMSE manifest",
     "5 distributions, 10M samples each",
-    "Gaps stay gaps until a run lands",
-    ""
+    "A software run, not silicon",
+    "Draft until someone reviews it"
 ];
 
 /// The protocol spec's own test count, which the evidence slide quotes.
@@ -200,6 +210,29 @@ pub fn step_number(cell: i32) -> i32 {
     return cell + 1;
 }
 
+/// Whether a SOURCES entry names a path: `path` occurs in `src` byte for byte.
+pub fn cites(src: []const u8, n: usize, path: []const u8, m: usize) -> bool {
+    if (m == 0 || m > n) {
+        return false;
+    }
+    var i : usize = 0;
+    while (i + m <= n) {
+        var j : usize = 0;
+        var same : bool = true;
+        while (j < m && same) {
+            if (src[i + j] != path[j]) {
+                same = false;
+            }
+            j += 1;
+        }
+        if (same) {
+            return true;
+        }
+        i += 1;
+    }
+    return false;
+}
+
 test the_domain_is_routed_not_chosen {
     assert DOMAIN == DOMAIN_RESEARCH;
     // gf16, gf16, benchmark; "bfloat16" is not "gf16", and no other list is hit
@@ -229,7 +262,7 @@ test every_step_is_a_cell {
 
 test a_gap_is_named_not_filled {
     assert gaps_are_named(SOURCES, MISSING, GROUNDED);
-    assert count_true(GROUNDED) == 22;
+    assert count_true(GROUNDED) == 25;
     // a gap is drawn as a gap, and only a gap is
     var c : i32 = 0;
     while (c < CELL_COUNT) {
@@ -237,21 +270,32 @@ test a_gap_is_named_not_filled {
         assert (STATUS[i] == STATUS_GAP) == (GROUNDED[i] == false);
         c = c + 1;
     }
-    // every gap is a DELIVER step: what is missing is the run and what follows it, not the protocol
-    assert step_stage(15) == STAGE_DELIVER;
-    assert step_stage(16) == STAGE_DELIVER;
+    // the two gaps left are DELIVER steps: the chip repo's run and the paper's Delta_dB
+    assert STATUS[17] == STATUS_GAP;
+    assert STATUS[25] == STATUS_GAP;
     assert step_stage(17) == STAGE_DELIVER;
-    assert step_stage(24) == STAGE_DELIVER;
     assert step_stage(25) == STAGE_DELIVER;
+    // the run's own cells cite the run, by both paths
+    assert MISSING[15].len == 0;
+    assert MISSING[16].len == 0;
+    assert MISSING[24].len == 0;
+    assert cites(SOURCES[15], SOURCES[15].len, RUN_HARNESS, RUN_HARNESS.len);
+    assert cites(SOURCES[15], SOURCES[15].len, RUN_MANIFEST, RUN_MANIFEST.len);
+    assert cites(SOURCES[16], SOURCES[16].len, RUN_MANIFEST, RUN_MANIFEST.len);
+    assert cites(SOURCES[24], SOURCES[24].len, RUN_MANIFEST, RUN_MANIFEST.len);
+    assert cites(SOURCES[24], SOURCES[24].len, RUN_HARNESS, RUN_HARNESS.len);
+    // and cites is not a yes-machine: the protocol's own cells do not name the run
+    assert cites(SOURCES[0], SOURCES[0].len, RUN_MANIFEST, RUN_MANIFEST.len) == false;
+    assert cites(SOURCES[16], SOURCES[16].len, RUN_HARNESS, RUN_HARNESS.len) == false;
 }
 
 test a_status_is_earned_not_chosen {
     assert statuses_earned(STATUS, DOMAIN, GROUNDED, TEST_PASS, REVIEWED, LIVE);
-    // the evidence slide's "15 published, 1 verified, 6 drafts" and "5 gaps"
+    // the evidence slide's "15 published, 1 verified, 9 drafts" and "2 gaps"
     assert count_status(STATUS, STATUS_PUBLISHED) == 15;
     assert count_status(STATUS, STATUS_VERIFIED) == 1;
-    assert count_status(STATUS, STATUS_DRAFT) == 6;
-    assert count_status(STATUS, STATUS_GAP) == 5;
+    assert count_status(STATUS, STATUS_DRAFT) == 9;
+    assert count_status(STATUS, STATUS_GAP) == 2;
     assert count_true(REVIEWED) == 0;
     // every passing test is one of the protocol's own tests, on a PRACTICE step
     assert count_true(TEST_PASS) == PROTOCOL_TESTS;
@@ -267,6 +311,23 @@ test a_status_is_earned_not_chosen {
     assert STATUS[22] == STATUS_DRAFT;
 }
 
+test the_run_is_a_draft_until_reviewed {
+    // steps 16, 17 and 25 now have the run as their source, and are DELIVER steps
+    assert STATUS[15] == STATUS_DRAFT;
+    assert STATUS[16] == STATUS_DRAFT;
+    assert STATUS[24] == STATUS_DRAFT;
+    assert step_stage(15) == STAGE_DELIVER;
+    assert step_stage(16) == STAGE_DELIVER;
+    assert step_stage(24) == STAGE_DELIVER;
+    assert REVIEWED[15] == false;
+    assert REVIEWED[16] == false;
+    assert REVIEWED[24] == false;
+    // a passing test and a live page do not lift a research DELIVER step: only a review does
+    assert earned_status(15, DOMAIN, true, true, false, true) == STATUS_DRAFT;
+    assert earned_status(15, DOMAIN, true, false, true, false) == STATUS_VERIFIED;
+    assert earned_status(15, DOMAIN, true, false, true, true) == STATUS_PUBLISHED;
+}
+
 test the_camera_is_the_demo_camera {
     assert DEMO_COUNT == 5;
     // scene 1 shows lens 0, the row HEAD[1] calls "Steps 1-9"