Phase Closure — Hierarchical-AIF Corrected B4 Programme

Date: 2026-07-22 · Gate: H-AIF-G5 complete, H-AIF-G9 recorded · Scope: the corrected B4 programme (C01/C02/C10/C11), the F-side motor-stack build and scoring, and the eleven defects routed along the way. This is a status closure, not a release. It introduces no new claim.

Full biological parity is not a current status. It is the target world defined by the receipts. Nothing in this phase is evidence of it.

1. What was corrected

• D1 — C11 cluster-collapse bootstrap. The committed runner regrouped resampled motors by motorId, collapsing 80 draws into 46–59 groups and over-sharpening each duplicated motor's latent. Repaired by bootstrap.build_bootstrap_cohort (one exchangeable group per draw), and the corrected full-N run landed.
• D3 — hash-seed nondeterminism. C01/C02 seeded via process-varying hash(str). Replaced by seeding.stable_seed (SHA-256). Both affected cells completed deterministically.
• D2 — resource-bound overestimate. Four cells were NOT_RUN/partial on cost claims of 150–400 h; all four ran in 2.3–21.1 h. Every RESOURCE_BOUND reason is discharged by measurement.
• D4 — C01 reason mismatch (superseded in new reports; frozen artifact not edited).
• D7 — width = percentile not BCa (forward guard + corrected 0.042-nat floor).
• D8 — ledger claimed undelivered tests (delivered + mutation-tested).
• D10 — CI rule has no minimum effect size (added interpretation layer; frozen verdicts unaltered).
• D11 — fabricated/transplanted numbers in a generated report (corrected + a mechanical numeric_provenance_guard).

2. What now stands

• All four corrected B4 cells landed at full frozen N, 0 failures each (C11 had 0/2000 failed; C10 had 6/2000). Each hashed, verified, and reported against its committed prediction.
• B3 stands unchanged. The corrected work touched none of the B3 leaderboard.
• The F-side motor-stack model is built, scored, and reproduces six frozen models at exact-zero oracle residual. Determinism proven byte-identical.
• Two mechanical guards now defend the namespace: claim_guard (wording) and numeric_provenance_guard (declared numbers).
• Test suite: 501 passed, 2 skipped, 1 xfailed. Frozen evidence: 250 files, no drift.

3. What was withdrawn

• The submitted C11 U4_OK (30 of 2000 replicates under the defective bootstrap). It is withdrawn and stays withdrawn — not restored by the corrected run. A defect that produced a right-looking answer is still a defect.
• The PROSPECTIVE label from B4C10 and the F-side scoring — both are NOT_SATISFIED by the commit graph (D9); their prediction and result entered the repo in one commit (B4C10) or the protocol was untracked while the result existed (F-side). The measurements stand; the label does not.
• The 250–400 h and 150–250 h RESOURCE_BOUND cost claims — superseded by measured runtimes.

4. What was replaced

• C11 U4 — the withdrawn 30-replicate reading is replaced by the corrected full-N U4_OK (collapse 0.0055, N=2000, groups 80/80), within C11 U4 scope only.
• The "0.064-nat resolution floor" — replaced by the corrected BCa half-width 0.042 nats (D7), which drives the materiality reading.
• The frozen C01/C02/C10/C11 NOT_RUN/partial statuses — replaced by full-N results, without editing the frozen artifacts.

5. What remains NOT_ESTABLISHED

• P4 transfer — the first unsatisfied level. No independent dataset.
• P5 intervention — the action set is empty; structural, not sample-size.
• P6 — carried per scope: C11 U4 re-established (corrected run only); Wadhwa mark-process retrospective-only (D5/D6); full motor-stack AIF pending a mechanism discriminator.
• P7 independent replication — external review in progress, not complete.
• The candidate's superiority over any serious adversary — NOT_ESTABLISHED throughout.

6. What C01 says about assay power

Under correct specification, the 19-motor motor-equal assay:

• resolves structurally distinct models (M3 self-wins 0.935, M2 0.885), but
• is nesting-blind — a true exponential wins its own competition only 0.290 of the time, barely above the ~0.25 that near-random selection among ~4 indistinguishable models gives.

Parameter recovery is intact (all four parameterised generators 1–2 orders inside tolerance), so this is not a fitter defect — it is a measured statement that this design cannot separate nested near-equivalent dwell families at this cohort size. "Underpowered is not equivalence" now has a number.

7. What C11 says about M7 hierarchy stability

On a valid bootstrap, M7's dispersion parameter tau is identified on this cohort (U4_OK, collapse 0.0055) — it does not silently degenerate to tau → 0 under resampling. The corrected distribution is shifted toward zero relative to the withdrawn artifact (p025 0.1766 → 0.0853), so the old number was mildly over-optimistic, but the effect is ~85× from the collapse boundary. Identification is not necessity and is not mechanism.

8. What F-side scoring says about the candidate model

The constrained F-side motor stack:

• scored NOT_ESTABLISHED against CONTROL_CURRENT (M3) and every serious adversary;
• reproduces frozen M7 to 2.5e-7 nats — at this resolution a re-derivation of the incumbent, not a new model;
• buys nothing measurable over its own tau → 0 limit M1_WEIBULL (contrast +0.000615, half-width 0.0111);
• ranks 5th of 10 on the combined leaderboard;
• exposed D10 via a 2.5e-7-nat "resolved" contrast against M7.

The hierarchy is architecture, not evidence, on the current data — a design signal.

9. Why M2 remains a serious adversary

M2_LOGNORMAL leads the held-out motor-equal leaderboard (3.4093), ahead of every mechanistic candidate including the F-side stack (3.4327). B4C02 showed this advantage is generator-specific (it appears under a three-timescale world, 0.94, and vanishes under the others, 0.005) — so "it is just heavy-tailed shape" is not an adequate account, which makes the M2-over-M3 result more interesting, not less. M2 is a one-parameter adversarial baseline, never the UNI model, and it is kept alive precisely because it is not yet beaten. Keeping it alive is strong inference.

10. Why full biological parity is not inferred

• Every B4C01/B4C02 dataset is SYNTHETIC; none may be labelled OBSERVED.
• B4C10/B4C11 are identifiability checks on frozen fits, not observations, not mechanism.
• The F-side model beats no serious adversary and is a re-derivation of M7.
• P4, P5, P7 are irreducibly external and unsatisfied.
• P8 is conjunctive and FULL_PARITY = false; the first unsatisfied level is P4.

No result in this phase licenses a claim of biological parity, active inference demonstrated, or G-side biological policy selection. The dataset is passive; the action set is empty.

11. Existing P0–P8 mapping

No P-level was raised by this phase. One P6 scope (C11 U4) was lowered on defect and re-established on the corrected run — the gate working end to end.

12. Next required receipts for P4 / P5 / P6 / P7 / P8

P4, P5, P7 cannot be closed by any modelling in this repository. The next-gate plan (protocols/NEXT-GATE-TRANSFER-AND-INTERVENTION-PLAN.md) specifies the external receipts.

NEXT_ACT = write hierarchical-aif/protocols/NEXT-GATE-TRANSFER-AND-INTERVENTION-PLAN.md