B4C10 — Corrected Full-N Result Report

Gate: H-AIF-G5 · Cell: B4C10_M4_STRUCTURAL_IDENTIFIABILITY · Date: 2026-07-21 Branch: hierarchical-aif/motor-stack Maps to existing ladder: P1 equation/implementation, P0 provenance; a scoped identifiability statement that touches the P6 discussion without raising any P-level Prediction record (WRITTEN before execution; NOT committed before execution — it entered the repo in the SAME commit as the result, so prospectivity is NOT_SATISFIED and structurally unattainable, see §6): hierarchical-aif/protocols/B4C10-CORRECTED-FULL-PREDICTION.md Lane: LANE B (corrected B4 robustness cells)

No frozen artifact was modified. The runner is the committed, byte-identical audits/phase-b/b4-identifiability-robustness-runner.py. The only difference from the recorded run is the replicate count, moved from 100 to the frozen 2000.

1. Executive result

B4C10 ran to completion at the full frozen N_boot = 2000: 1994 replicates completed, 6 failed, resourceBoundPartial: false. All three frozen identifiability checks returned OK — U2_bootstrapCollapseFrac = 0.0050150451354062184 (fires at ≥ 0.25), U3 log10(lambda_3) 95% interval span = 0.43327087478677417 decades (fires at ≥ 2.0), and U4 omega_3 95% percentile interval = [0.030746372918754247, 0.28838105457744573] (fires only if it straddles, i.e. lo < 5/793 and hi > 0.25; the lower bound does not, so U4 does not fire). The runner therefore recorded M4_status = "IDENTIFIED_ON_THIS_COHORT (U1 DISTINCT from B3, U2/U3/U4 OK)", and the reduced-budget calibration shortfall was 8.7993612396530807e-11 nats against a 0.05-nat threshold (WITHIN_0.05_NATS). By the mapping committed in §5 of the prediction record, the frozen prediction UNIDENTIFIED_OR_WEAK is REFUTED AT FULL N, within B4C10 scope only: B4C10 at full frozen N=2000 supports M4 identifiability on the derived_eligible_1_to_8 cohort under the frozen U2/U3/U4 criteria. This is an identifiability statement about one model on one cohort. It is not evidence for any mechanism, it does not make M4 correct, and it may not be transferred to M7, to C11, or to any parity claim.

2. Provenance

Three independent hash witnesses agree (stdout at write time, the .sha256 sidecar, and a fresh sha256sum at report time). The result file has not drifted since it was written.

3. Planned vs actual N, and the 6 failures

The run is a full-N run: every one of the 2000 frozen replicates was attempted. Six did not yield a usable fit.

Does the JSON record a reason for the 6 failures? NO — NOT_RECORDED. The result object carries only the integer failed: 6. Reading the frozen runner (cell_C10, lines 1060–1092) shows exactly two code paths that increment it, and the runner distinguishes neither in the output:

1. b3.Cohort(f"C10_b{b}", …) raises while rebuilding the resampled cohort (line 1076–1079); or
2. _fit_m4_reduced returns None (lines 1081–1084), which happens when the differential-evolution objective is non-finite or >= 1e11 (line 1015), or when the canonicalised mixture violates the mean-one constraint by more than 1e-8 (line 1018).

What would be needed to attribute them. The frozen runner may not be edited, so attribution requires an out-of-tree replay under hierarchical-aif/: re-run the deterministic per-replicate seed sequence 20260717 + b for b = 0..1999, reconstruct each bootstrap cohort by the same recipe, and record which of the three exit conditions fired for each failing b. The seeds are arithmetic and the pipeline is deterministic, so this replay is exactly reproducible; it costs another full-N pass (≈2.3 h). It has not been done. Until it is, the failure cause is NOT_ESTABLISHED.

Why this is not waved away. Failure path 2 is not neutral: a non-finite objective or a mean-one violation is the kind of numerical outcome that a near-degenerate mixture can produce. Dropping such replicates is potentially informative (non-random) missingness, and it biases the U2 collapse fraction downward. That is a real, named risk, not a rounding detail.

Bounded sensitivity, computed rather than asserted. The recorded collapse fraction implies a collapsed count of 0.0050150451354062184 × 1994 = 10.0 exactly, i.e. 10 collapsed of 1994 completed. Under the adversarial worst case that all 6 failures were collapses, the fraction over all 2000 attempted replicates would be (10 + 6) / 2000 = 0.008, versus the U2 firing threshold of 0.25 — a factor of 31.25 below it. Under the benign case it is 10 / 2000 = 0.005. The U2 verdict is therefore insensitive to the attribution of the 6 failures. No comparable bound is available for U3 and U4, because the failed replicates contribute no log10(lambda_3) or omega_3 value at all and their quantile influence cannot be bounded without knowing what those values would have been — that remains NOT_COMPUTED.

4. Criteria and results

Criteria and thresholds are read from the frozen runner (cell_C10, lines 1094–1114); they were not chosen or restated for this report.

U4 read precisely. 5/793 = 0.006305170239596469 (five events' worth of weight out of the 793 training events). The observed lower bound 0.030746… is 4.88× above that point, so lo < 5/793 is False. The observed upper bound 0.288381… is above 0.25, so hi > 0.25 is True. Because the criterion is a conjunction, U4_OK here is carried entirely by the lower bound; the upper half of the straddle condition is satisfied. Anyone reading U4_OK as "the interval is comfortably away from both degeneracy edges" would be reading it wrong.

Reduced-budget calibration (the check that the reduced DE budget — maxiter=400, popsize=30 — does not under-fit relative to B3's full M4 fit):

These calibration numbers are bit-for-bit identical to the recorded 100-replicate run (see §5), which is expected: the calibration fit uses seed_base on the unresampled cohort and does not depend on n_boot. It is a reproducibility witness, not new independent evidence.

Units note: the calibration NLLs are training negative log-likelihoods produced by the B3/B4 pipeline on its own scale, and the shortfall is a difference of two such quantities, so it is in nats regardless. The collapse fraction is dimensionless; the U3 span is in decades of log10 rate; omega_3 is a mixture weight (dimensionless, in [0,1]). None of these is an NLPD, so the seconds-vs-normalised-y Jacobian question that governs B3 NLPD comparisons does not arise here.

5. Old (100/2000 partial) vs new (2000/2000 full)

Old column from audits/phase-b/b4-identifiability-robustness-result.v1.json, cell B4C10 (frozen, read-only). New column from B4C10_CORRECTED_FULL_RESULT.json.

The collapse fraction is the instructive one, and it is the D2 lesson in miniature. The partial run observed 0 collapses in 100 replicates and recorded U2_bootstrapCollapseFrac = 0. A count of zero cannot bound a rare-event fraction tightly: the exact one-sided 97.5% Clopper–Pearson upper limit for 0/100 is 0.0362, and the familiar rule-of-three bound is ≈0.03. The full-N value of 0.005015 sits comfortably inside that band. Indeed, if the true replicate-collapse rate is 0.005015, the probability of observing zero collapses in 100 draws is (1 − 0.005015)^100 = 0.605 — the partial reading was the more likely observation, not a fluke. So the two readings are entirely consistent; the partial one simply carried no information about where in [0, 0.036] the fraction lay, and reporting it as the bare point value 0 overstated what 100 replicates could say.

This is exactly the D2 failure mode — a partial run presented with the precision of a full one — and this time it happens to land in the benign direction: the true fraction is small, the verdict does not flip, and the earlier reading survives. That is luck about the direction, not vindication of the practice. The identical procedure applied to a cell where the rare event is common would have produced a wrong verdict, and the prediction record §5 explicitly reserved that outcome as the more interesting one.

The U3 span widened (+6.1%) and the U4 interval narrowed (−0.0208) on going from 100 to 2000 replicates. Neither movement is large relative to its threshold, and neither is treated here as a finding: with 100 replicates the 2.5%/97.5% empirical quantiles are each estimated from a handful of order statistics, so movement of this size is expected Monte-Carlo behaviour of the quantile estimator.

Statistical scope of every number in this section. These are bootstrap-replicate fractions and bootstrap intervals. A bootstrap replicate is not a biological replicate. The binomial arithmetic above is a statement about Monte-Carlo sampling noise in the resampler, not a biological inference. The experimental unit remains the motor, and the resampling unit here is the training motor (80 of them, drawn with replacement, n_tm = 80 per replicate). Nothing in this cell increases the biological sample size.

6. Prediction outcome

The frozen prediction for this cell was UNIDENTIFIED_OR_WEAK. Its falsifier, stated in the prediction record §5, was "any of U2/U3/U4 firing at full N" — no such firing occurred; the committed mapping's first row is the one that applies.

observed at N=2000	result	claim impact
U2/U3/U4 all OK	frozen prediction UNIDENTIFIED_OR_WEAK REFUTED at full N	The earlier REFUTED_PARTIAL is upgraded to a full-N refutation within C10 scope only. M4 identifiability on this cohort is supported. This is a P1/P6-scoped statement about one model's identifiability — not evidence for any mechanism.

Applying it:

• The frozen prediction UNIDENTIFIED_OR_WEAK is REFUTED AT FULL N, within B4C10 scope only.
• The earlier REFUTED_PARTIAL is upgraded to a full-N refutation, again within B4C10 scope only. The upgrade is licensed because the run reached the frozen N with resourceBoundPartial: false, not because the numbers looked favourable.
• Of the three hypotheses held alive in prediction record §4, H_M4_IDENTIFIED is supported on this cohort under these criteria; H_M4_WEAK is not supported at full N; and H_NOT_ESTABLISHED does not apply, because all three intervals closed against the frozen criteria.

Being wrong in the pre-committed direction is the recorded outcome, not a defect. The prediction was written before the run, committed, and then contradicted by the evidence. That is a prediction record working correctly. The prediction record itself said as much: "Prediction I am willing to be wrong about." No criterion, threshold, seed, or N was changed after seeing a number, and the mapping applied here is the one that was committed, not one selected afterwards.

Prospectivity: NOT_SATISFIED. This is a closed negative, not a pending check.

An earlier draft of this report graded the ancestry check NOT_VERIFIED, which wrongly implied a check that might still pass. It cannot pass. The commit graph was inspected:

The prediction and the result entered the repository in ONE commit, 2 h 33 min after the run had already finished. Per CLAUDE.md — "a prediction is prospective only if it was committed before its observation" and "prospectivity is decided by the commit graph, not by prose" — the required strict-ancestor relation is structurally unattainable for this pair: a commit cannot be its own strict ancestor. This result may not be labelled PROSPECTIVE, and no future commit can repair it.

Correction, recorded rather than silently fixed. An earlier draft of this section asserted that the result was "NOT COMMITTED — still an untracked working-tree file", and reasoned that a future commit would satisfy the mechanical ancestry test while failing the substantive one. That premise was false: git log --diff-filter=A shows the result was already committed, in b9b5670, alongside the prediction. The conclusion is unchanged and in fact stronger — there is no mechanical test left to launder, because same-commit introduction forecloses ancestry outright. Found by adversarial verification of this report.

What is verified, and all this report claims: the prediction file records UNIDENTIFIED_OR_WEAK with its outcome mapping fixed in advance of this report; the outcome contradicts it; and the mapping applied is the recorded one, not one selected afterwards. That makes this a retrospectively-graded refutation on a pre-written prediction, which is weaker than a prospective refutation and is labelled as such throughout.

Tracked as D9 in ledgers/HIERARCHICAL-AIF-DEFECT-CLOSURE-LEDGER.md, which also records the only receipt that fixes this going forward: commit the prediction record before launching the run it predicts.

7. Scope boundary — hard

Reproducing prediction record §6 verbatim:

C10 may not be used to repair C11. They concern different models (M4 vs M7) and different likelihood structures (pooled i.i.d. vs motor-grouped). A favourable C10 says nothing about M7 identifiability, and nothing about the withdrawn C11 U4_OK.

Spelled out:

• Different models. C10 tests M4_MIXTURE_K3, a three-component mixture. C11 tests M7_HIERARCHICAL_MOTOR, a motor-grouped hierarchy. They do not share parameters, a parameter count, or an identifiability geometry.
• Different likelihood structures. _fit_m4_reduced fits the flat pooled coh.train_y. M7's training NLL iterates coh.train_by_motor — the grouped structure that D1 corrupts. This structural difference is the reason C10 is valid on the frozen runner while C11 is not; it is the same fact, and it cuts in both directions.
• Therefore: this result does NOT restore, repair, support, or partially support the withdrawn C11 U4_OK. P6 for C11 U4 remains withdrawn pending the corrected full-N C11 run (defect D1, OPEN_UNTIL_CORRECTED_C11_FULL_RUN). Nothing in this report may be cited toward it.

Identifiability is not mechanism, and identifiability is not correctness. U2/U3/U4 ask only whether the fitted M4 parameters are recoverable and distinguishable on this cohort under bootstrap resampling. A model can be perfectly identifiable and completely wrong about the biology; an identifiable mixture is still a curve-fitting object until an independent mechanistic prediction survives. Specifically:

• M4 being identified on this cohort does not make M4 correct.
• M4 being identified does not make M4 the UNI model. It is a competitor in the B3 leaderboard.
• No mixture mechanism is confirmed, implied, or supported by this result.
• The finding is bounded to one cohort (derived_eligible_1_to_8), one model, one criterion set, and one resampling scheme.

8. Defect impact

D2 — resource-bound overestimate: second, independent confirmation. RESOURCE-BOUND-RECLASSIFICATION.md §2 projected B4C10 at 2.1 h from a measured per-unit _fit_m4_reduced cost of 3.8 s (2000 × 3.8 s = 7600 s = 2.111 h). The completed run measured 8372.7 s = 2.326 h — a ratio of 1.102, i.e. the projection was accurate to within 10.2%, with the excess attributable to CPU contention (the B4C02 full-N run was executing concurrently on the same machine). Effective per-replicate cost under contention was 4.186 s.

The first confirmation was the per-unit timing measurement itself (an extrapolation); this is the second and stronger one, because it is an end-to-end wall-clock measurement of the whole cell rather than a projection. Set against the runner's own stated cost model for this cell — the NOT_RUN branch text at lines 1033–1036, "≈ 33-100 h wall-clock; not feasible in this dispatch's compute budget" — the measured 2.326 h is 14.2× to 43.0× cheaper. Precision note: that 33–100 h text sits in the n_boot <= 0 branch, which this cell never took; the recorded partial artifact for B4C10 carries no hour figure of its own, only resourceBoundPartial: true and the verdictScope string. So the comparison is against the runner's stated cost model for the frozen N, not against a number in the C10 result object. Either way, B4C10 should never have been partial, and the record now shows that on measurement.

D1 — cluster-bootstrap collapse: does NOT apply to C10. _fit_m4_reduced fits the flat pooled coh.train_y (runner line 1008), so duplicate motor draws enter the likelihood correctly and the motor-cluster bootstrap is valid as written for a pooled i.i.d. likelihood. This is asserted by test_m4_pooled_path_is_unchanged_by_the_fix, which compares sorted train_y between the legacy and corrected arms, and it is recorded in the closure ledger's D1 unaffected_lanes field. No correction was applied to this run (correctionApplied = NONE), and none was needed.

D3 — hash-seed non-determinism: does NOT apply to C10. C10 seeds with np.random.default_rng(seed_base) and seed_b = seed_base + b (runner lines 1056, 1063) — arithmetic only, no hash(). Confirmed by reading the frozen runner. The D3 closure ledger entry lists C10 under unaffected_lanes for the same reason.

D7 — width field provenance. Not applicable in the direction that matters, but stated for discipline: the U3 span and the U4 interval reported here are percentile bootstrap quantities (np.quantile at 0.025/0.975, runner lines 1097–1098), computed directly by the C10 cell. They are not BCa intervals and are not the B3 width field that D7 concerns. Any width quoted from this report must be labelled percentile.

D5/D6 — held-out data firewall. This cell reads the frozen derived_eligible_1_to_8 cohort's duration channel and carries the original holdout events through cohort reconstruction unchanged (runner lines 1071–1074) purely so that b3.Cohort can re-derive the frozen split. No held-out mark field (nextStateN, direction, jump) is read, printed, or reasoned about anywhere in this cell or in this report. No new data channel was spent.

9. Lane impact

10. P-ladder mapping

Per hierarchical-aif/ledgers/HIERARCHICAL-AIF-GATE-TO-EXISTING-P-LADDER-MAP.md, this is a H-AIF-G5 receipt (corrected full B4 reruns), which that map routes to P3 held-out predictive and P6 structural/mechanistic per cell. For this cell the honest routing is narrower than the gate's general mapping:

Explicitly: NO P-LEVEL IS RAISED BY THIS RESULT. It is a P1/P6-scoped identifiability statement about one model on one cohort, exactly as the prediction record's committed mapping said it would be. P6 for C11 U4 remains withdrawn. P6 for duration-only B3/B4 remains unchanged. P8 remains FULL_PARITY = false with P4 transfer as the first unsatisfied level. The map's ledger rule §4 is satisfied here only in the sense that it licenses no update: the level definitions are named, the artifact is named, the scope is named, the falsifier is carried, and the partial/negative states are preserved — and the conclusion of applying it is that nothing moves.

11. Reproduction

cd [redacted: operator-path]/Documents/UNI-Flagellum/UNI-FLAGELLUM
python audits/phase-b/b4-identifiability-robustness-runner.py \
  --cells C10 --c10-boot 2000 \
  --out hierarchical-aif/results/motor_stack_aif/B4C10_CORRECTED_FULL_RESULT.json

Expected artifact sha256:

959a00e974641eca1c0d6f3c2f7322b8c6c8411f68ca953c7e169492c4a53dde

Verify with:

sha256sum hierarchical-aif/results/motor_stack_aif/B4C10_CORRECTED_FULL_RESULT.json
sha256sum audits/phase-b/b4-identifiability-robustness-runner.py
# runner must be 3e21edac97a2b68faec087e73de307439348e89e778b6e97d84809bbf1e135a7

Environment required for byte-identity: python 3.12.10, numpy 2.3.5, scipy 1.16.3, CPU-only, single worker (workers=1, updating="deferred" are fixed in _fit_m4_reduced). Expected wall-clock ≈2.1–2.4 h depending on machine contention. All seeds are arithmetic (20260717 + b), so the result should be reproducible bit-for-bit on the same scipy version; bit-identity across scipy versions is not claimed and has not been tested — differential_evolution's internals are version-sensitive.

12. Limitations and adverse notes

1. 1994 of 2000, not 2000 of 2000. Six replicates failed with no recorded reason (NOT_RECORDED). Attribution requires the out-of-tree replay described in §3 and has not been done. The U2 verdict is bounded against the worst case (0.008 vs a 0.25 threshold); the U3 and U4 quantiles are not so bounded — that sensitivity is NOT_COMPUTED.
2. One cohort. derived_eligible_1_to_8 only. 80 training motors, 793 training events, 19 holdout motors. Right-censored events are excluded from the frozen cohort entirely, so this result says nothing about censored dwells.
3. 19 holdout motors is a small number of experimental units, and the experimental unit is the motor. This cell does not use the holdout for scoring, but the surrounding evidence body is built on those 19 units, and no result here enlarges them. Underpowered is not equivalence.
4. Bootstrap replicates are not biological replicates. Every interval here is a Monte-Carlo interval over resampled training motors, not a confidence statement about a population of motors, cells, or cultures.
5. The adverse M2-over-M3 result stands, untouched and unremediated. M2_LOGNORMAL — an adversarial baseline, never the UNI model — out-predicts the reference M3_TWO_TIMESCALE by ≈0.0369 nats event-pooled on held-out data. Nothing in B4C10 addresses, weakens, or explains that finding. It is reported alongside this result, never instead of it.
6. Related resolution note (D7). The narrowest motor-equal B3 contrast is M4_MIXTURE_K3 at a percentile width of 0.083461 (BCa width 0.084141), and the corrected resolution floor is ≈0.042 nats (BCa half-width 0.042070; percentile half-width 0.041730) — not the 0.064 figure that came from the mislabelled width field. Every width here carries its interval label, because quoting 0.083461 bare would repeat the exact D7 defect this note exists to correct. That B4C10 concerns M4 and that M4 holds the narrowest B3 contrast are separate facts about the same model; neither supports the other.
7. Identifiability ≠ correctness ≠ mechanism. Restated because it is the single most likely misreading of this report. M4_status = IDENTIFIED_ON_THIS_COHORT is a statement about parameter recoverability under resampling, full stop.
8. The U1 component is inherited, not re-derived. U1_alreadyDone_in_B3: "DISTINCT (per B3 result)" is copied from the B3 artifact. The M4_status string asserts U1; this run did not re-test it.
9. Prospectivity is NOT_SATISFIED (§6) — a closed negative, not a pending check. The prediction record was committed in b9b5670 at 2026-07-21T22:44:31Z, 2 h 33 min after the run finished at ~20:11:51Z. This refutation is retrospectively graded against a pre-written prediction and may never be labelled PROSPECTIVE. Tracked as D9.
10. Concurrency caveat on the timing claim. The 8372.7 s figure was measured while B4C02 ran concurrently. It is therefore an upper bound on the uncontended cost of this cell, which strengthens rather than weakens the D2 conclusion, but it is a single measurement on one machine and not a distribution.
11. B4C02 is still running. Its partial log is not a result and is not referenced, quoted, or relied upon anywhere in this report.

NEXT_ACT = Update the H-AIF-G5 row and the B4C10 row of RESOURCE-BOUND-RECLASSIFICATION.md / HIERARCHICAL-AIF-NEGATIVES-AND-PARTIALS.md to record B4C10 as FULL_N_DELIVERED with this report and sha256 959a00e9… as its receipt; run claim_guard over this report; then, once B4C02 lands, queue B4C11 at the frozen N=2000 on the CORRECTED build_bootstrap_cohort — that run, and only that run, can address the withdrawn C11 U4 and defect D1. Do NOT let this favourable C10 shorten, reduce, or pre-judge it.