Correction Notice — UNI Flagellum B3/B4 Review Package

Issued: 2026-07-21T16:39:52Z Issued against: commit 17a2f0e18c09c762ab1cefe854c0d68698803eac (branch phase-2/b3-model-competition) Original package: uni-flagellum-b3-b4-evidence-17a2f0e.zip, sha256 ae42f203c2d156aca1d1345a27da3d522bccd6ae3bbb58cf780cef0c413087f1 Correction work branch: hierarchical-aif/motor-stack Gate: H-AIF-G2

Provenance note. This work was originally requested as Phase-C closure work, but was namespaced under hierarchical-aif/ to avoid collision with frozen audits/phase-c/ evidence. The frozen Phase-C audits remain read-only source evidence.

1. Status

The package previously submitted for review is corrected in place by this notice. It is not deleted and not silently replaced. It remains a historical artifact, but parts of its B4 interpretation are invalidated.

The previously submitted package is not withdrawn for all purposes, but its B4 robustness
interpretation is corrected.

B3 leaderboard: still standing unless later evidence changes it.
B4 C11 U4: invalidated by verified cluster-collapse bootstrap defect.
B4 resource-bound labels for C01/C02/C10/C11: invalidated / require correction.
B4 C01/C02 future runs: blocked until deterministic seeding is fixed.
Full parity / mechanism claims: not licensed.

These defects were found by an internal read-only audit of the repository against a stacked active-inference specification, and were then independently reproduced and verified before this notice was issued. They were not reported by the reviewer.

2. What still stands

• Package integrity receipts remain valid historical receipts. The 43-artifact manifest, the determinism re-run (f361e4dc…), the 43/43 independent-oracle agreement, the live-remote ancestry proofs, and the plan §7 validation baseline were all re-verified at H-AIF-G1 on 2026-07-21 and are unaffected by D1–D4.
• The B3 leaderboard and B3 held-out scoring are not invalidated by D1–D4 unless later evidence shows otherwise. D1 is a defect in a bootstrap resampling path used only by B4C11; B3's own M7 fit runs on the real 80-motor cohort with no resampling. The adverse headline — a simple lognormal (M2) out-predicting the two-timescale mixture (M3) on held-out data — is untouched and remains the retained headline.
• B4C11 U2 (the 61-point τ profile) is unaffected: it runs on the real cohort at the full frozen contract, not on bootstrap replicates.
• B4C11 U1 and U3 derive from B3 and are unaffected.
• B4C10's bootstrap method is sound. M4 is fitted on the flat pooled train_y, so duplicate motors enter correctly — it is a valid motor-cluster bootstrap. Only its replicate count and its resource justification are at issue (see D2).
• The current-design control remains available and runnable.

3. What does not stand

D1 — B4C11 U4 U4_OK does not stand

The C11 U4 bootstrap resamples motors with replacement, concatenates the sampled motors' events into a flat list, and rebuilds the cohort. Cohort groups training events into train_by_motor keyed by motorId, so a motor drawn K times collapses into one group holding K copies of its events instead of K separate exchangeable groups.

Reproduced at the declared seed_base = 20260717:

• Group count equals the number of distinct motors exactly → duplicates collapse.
• Cluster loss at b=0: 42.5% (80 intended → 46 actual exchangeable clusters).
• Largest group inflates 70 → 153 events (a 51-event motor drawn 3×).
• Theoretical E[distinct] = 80(1-(1-1/80)^80) = 50.8, consistent with the observed range.

This reaches M7's likelihood: m7_train_nll(k, tau, train_by_motor) iterates for ym in train_by_motor, and _fit_m7_reduced consumes coh.train_by_motor. A motor drawn K times therefore contributes L_m^K, over-sharpening that motor's latent while the number of exchangeable clusters silently drops by roughly 40%.

The direction of bias is an inference, not a measurement, but it plainly favours the conclusion that was reported ("τ is well determined").

Consequently:

• the reported U4_OK verdict does not stand;
• the reported τ bootstrap CI [0.17658259553140288, 0.27020046374673834] must not be used as evidence about between-motor dispersion;
• the collapseFraction_tau_lt_1e_3 = 0 result from the defective runner must not be cited;
• the B4 headline sentence "B4C11 … U4 at 30 of 2000 replicates 0/30 collapsed (U4_OK)" is withdrawn;
• the C11 contribution to M7_status = IDENTIFIED_ON_THIS_COHORT (U1/U2/U3/U4 all OK) is reduced to U1/U2/U3 only, and the aggregate status is NOT_ESTABLISHED pending a corrected full run.

D2 — B4 RESOURCE_BOUND labels are not reliable

Measured on the frozen cohort on the builder's machine:

Projected against frozen N versus what was recorded:

A RESOURCE_BOUND status is only honest if the resource claim is true. Three of these are wrong by more than an order of magnitude, which converted recoverable gates into permanent-looking NOT_RUNs and understated how much of B4 was actually within reach. This is a truth-contract issue, not a scheduling one.

B4C02 is the most consequential instance: it is the HIGH-risk misspecified-world discriminator — the cell designed to test whether the adverse lognormal result is a feature of heavy-tailed dwell shape rather than of one assumed mechanism. It was reported unrunnable at 150–250 h. It is approximately 8.7 h. The single most decisive piece of missing evidence in the submission was reachable and was not attempted.

D3 — B4C01/B4C02 seeding is non-deterministic across processes

Both cells seed via np.random.default_rng(seed_base + sim + hash(gen) % 100000). CPython randomizes str hashing per process and PYTHONHASHSEED is unset. Three consecutive processes produced seeds 14565/95125, 59809/55025, 89866/26054 for the same two generator strings.

Any future C01/C02 result would be unreproducible byte-for-byte and could never satisfy the protocol's own determinism discipline. The defect is latent only because those cells were never executed. C01/C02 must not be run until this is fixed.

D4 — B4C01's recorded reason misdescribes what the cell computes

The recorded reason states "~15-25 min per M4/M7-inclusive competition", but the cell's own code explicitly skips those models ("skip M4/M7/M8 as they are the slow ones"), and the cell records "skippedModels": ["M4_MIXTURE_K3", "M7_HIERARCHICAL_MOTOR", "M8_EMPIRICAL_KDE"]. The stated justification does not describe the implemented computation. This is a provenance-integrity defect in the recorded reason text, independent of the arithmetic error in D2.

4. Claim impact

No full parity claim is licensed.
No biological equivalence claim is licensed.
No active-inference-demonstrated claim is licensed.
No mechanism claim is licensed from the invalidated cells.
B4 must be rerun where affected.

Mapping onto the existing, unchanged P0–P8 ladder in CLAUDE.md:

No P8 movement. No parity level rises as a result of this notice. The pre-existing repository verdicts are unchanged: cross-study X16_FULL_BIOLOGICAL_PARITY = FAIL; science-gates overall PARTIAL_PARITY_ONLY.

5. Corrective action

The repository will, in this order:

1. preserve all old artifacts unmodified as historical evidence (no deletion, no overwrite, no history rewrite);
2. add failing reproducer tests for D1 and D3 before any fix;
3. fix the C11 cluster-bootstrap semantics so that K draws of a motor yield K exchangeable groups — an implementation fix to match the frozen protocol's intended cluster-bootstrap semantics, not a change to the frozen scientific protocol;
4. replace hash() seeding with a stable SHA-256-derived seed;
5. reclassify the resource-bound cells from measured runtime;
6. re-run the feasible cells at full frozen N on the corrected runner, under new corrected artifact names that do not overwrite the originals;
7. issue a corrected handoff package with an old-vs-new status table.

Corrected results will be written to B4C10_CORRECTED_FULL_RESULT.json, B4C11_CORRECTED_FULL_RESULT.json, B4C02_CORRECTED_FULL_RESULT.json, and B4C01_CORRECTED_FULL_RESULT.json. No corrected result is claimed until its corrected run exists.

6. Instruction to the reviewer

Please treat the following sections of the submitted package as superseded by this notice:

• HANDOFF-B3-B4.md §5 bullet "B4C10 / B4C11 REFUTE weak-identification prediction on partial replicates" — the C11 half is withdrawn;
• HANDOFF-B3-B4.md §6 rows B4C10 and B4C11;
• HANDOFF-B3-B4.md §3 and §13 items 1–2 (the RESOURCE_BOUND and partial-replicate justifications);
• the corresponding outcome.headline, outcome.resourceBoundLimitations, and outcome.notEstablished text in experiments/predictions/b4-identifiability-robustness.prediction.json.

Everything else in the package — B3, the integrity receipts, C03/C04/C05/C06/C07/C08, and the declared limitations — stands as submitted pending your review.

The reviewer's own G1 objection (that the git proofs were a text dump rather than live-repository verification) has since been closed independently: git ls-remote against the origin returns 17a2f0e18c09c762ab1cefe854c0d68698803eac, matching local HEAD, with all five anchor commits confirmed as ancestors of the remote branch and a pristine working tree. See H-AIF-G1-REPO-AND-FROZEN-EVIDENCE-INTEGRITY.md.