UNI Universal Natural Intelligence

Wiki · Hierarchical Active Inference

B4C10 — Corrected Full-N Prediction Record

Hierarchical Active Inference · hierarchical-aif/protocols/B4C10-CORRECTED-FULL-PREDICTION.md @ b909801f3db4 (hierarchical-aif/motor-stack) — opens the published snapshot 8b4b5935bcba

How to read this page

A Plain and a Clear version of this page have not been written yet. What follows is the document itself.

Precise — the source document

This is the document. Rendered from the repository at the commit above, with nothing rewritten for the web. A gate re-renders it on every deploy and fails the build if a single byte differs.

PROSPECTIVITY: NOT_SATISFIED — and not retroactively repairable (D9). This file was written 2026-07-21 prior to the full-N B4C10 run, but it was committed 2026-07-21T22:44:31Z in b9b5670, 2 h 33 min AFTER the run finished (~20:11:51Z) and after the result already existed on disk. Per CLAUDE.md, a prediction is prospective only if it was committed before its observation. This cell's result may never be labelled PROSPECTIVE. It is a retrospectively-graded refutation against a pre-written prediction, which is strictly weaker. Gate: H-AIF-G5 · Cell: B4C10_M4_STRUCTURAL_IDENTIFIABILITY

1. Purpose

Complete the M4 three-component-mixture identifiability bootstrap (U2/U3/U4) at the frozen N_boot = 2000. It was previously executed at 100 of 2000 replicates (5%) and labelled resourceBoundPartial, on a resource justification since measured at ≈2.1 h (D2). At that cost it should never have been partial.

2. Runs on the FROZEN runner, unmodified

Unlike B4C02, this cell requires no code change:

  • D1 does not apply. _fit_m4_reduced fits the flat pooled coh.train_y; duplicate motor draws enter correctly, so the bootstrap is a valid motor-cluster bootstrap for a pooled i.i.d. likelihood. Verified by test_m4_pooled_path_is_unchanged_by_the_fix.
  • D3 does not apply. C10 seeds with np.random.default_rng(seed_base) and seed_b = seed_base + b — arithmetic, no hash().

Therefore this run invokes the committed, byte-identical runner audits/phase-b/b4-identifiability-robustness-runner.py (sha256 3e21edac97a2b68f…) with --cells C10 --c10-boot 2000. Strongest possible provenance: the only change from the recorded run is the replicate count, which is the frozen value.

3. Frozen criteria (unchanged)

check criterion
U2 bootstrap collapse fraction of degenerate M4 fits
U3 width of the log10(lambda_3) 95% interval, in decades
U4 omega_3 95% interval

Recorded at 100/2000: U2_bootstrapCollapseFrac = 0; U3 span = 0.40821 decades; U4 omega_3 95% CI = [0.03813, 0.31659]; all three OK; M4_status = IDENTIFIED_ON_THIS_COHORT. The frozen prediction was UNIDENTIFIED_OR_WEAK, recorded as REFUTED_PARTIAL.

4. Hypotheses held alive

id claim
H_M4_IDENTIFIED M4 is identified on this cohort; the 100-replicate reading survives at full N.
H_M4_WEAK At full N the collapse fraction rises and/or the intervals widen, so U2/U3/U4 fires and the frozen UNIDENTIFIED_OR_WEAK prediction is upheld.
H_NOT_ESTABLISHED Intervals do not close against the frozen criteria.

5. Pre-committed outcome mapping

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.
any of U2/U3/U4 fires frozen prediction CONFIRMED at full N The 100-replicate reading was an artifact of partial sampling. The prior REFUTED_PARTIAL is withdrawn. This would be a second instance of a partial run misleading a verdict — materially reinforcing D2's lesson.
interval fails to close / run incomplete PARTIAL_NOT_ESTABLISHED No verdict.

Falsifier of the 100-replicate reading: any of U2/U3/U4 firing at full N.

Prediction I am willing to be wrong about: I expect U2/U3/U4 to remain OK at full N, because 0/100 collapses is reasonably strong evidence against a high collapse rate. But 100 replicates cannot bound a rare-event fraction tightly, so a nonzero collapse fraction at 2000 is entirely possible and would be the more interesting outcome.

6. Explicit boundary

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.

7. Expected runtime

≈2.1 h (2000 × 3.8 s measured). Frozen N; no reduced protocol required.

8. Wording

Allowed: "B4C10 at full frozen N=2000 supports / does not support M4 identifiability on the derived_eligible_1_to_8 cohort under the frozen U2/U3/U4 criteria."

Forbidden: "M4 is the correct model" · "mixture mechanism confirmed" · any transfer of this result to M7, to C11, or to mechanism/parity claims.

sha256 2c76973527ce1749 — of the original file, so what was ingested stays checkable.