B4C11 — Corrected Full-N Prediction Record

PROSPECTIVITY: NOT_SATISFIED TODAY, BUT STILL FIXABLE — AND THE WINDOW IS OPEN (D9). Written 2026-07-21 prior to the corrected B4C11 full-N run, and its exact bytes are pinned at launch time by B4C11_CORRECTED_FULL_ENV.txt (predictionRecord=... sha256 5d0a1170b78a860ca971cb9227ab86d90d60665a3345d28dc1a56ce437526ea1, started_utc=2026-07-22T01:40:03Z). But this file is UNTRACKED, and an uncommitted prediction cannot satisfy the commit-graph test.

The B4C11 result does not exist yet (run projected ≈14 h). Committing this file before that result lands would make B4C11 a genuinely prospective cell. That requires the human principal, since nothing in this programme is committed without authorization. It is the single highest-value, time-limited action available in this batch. Gate: H-AIF-G5 · Cell: B4C11_M7_STRUCTURAL_IDENTIFIABILITY · Lane: B (corrected B4 robustness) Defect routed: D1_C11_CLUSTER_COLLAPSE (OPEN → this run is its repair attempt)

1. Purpose

Re-execute the C11 U4 motor-cluster bootstrap at the frozen N_boot = 2000 with the D1 cluster-collapse defect corrected, so that the withdrawn U4_OK is either re-established on a valid bootstrap or replaced by an honest UNSTABLE_DISPERSION_U4_FIRES.

The submitted C11 U4_OK is withdrawn and stays withdrawn regardless of what this run returns. It came from 30 of 2000 replicates (resourceBoundPartial_U4: true) built with a bootstrap that collapsed 80 motor draws into ~46 groups. Neither problem can be repaired retroactively.

Frozen recorded artifact under repair: audits/phase-b/b4-identifiability-robustness-result.v1.json sha256 f361e4dcf5fb8e1bf1724844b0581ebd1392027bfb694220944f3d048f1807f3, cell B4C11, U4_bootstrap = {completed: 30, failed: 0, collapseFraction_tau_lt_1e_3: 0, tauHatSummary: {median: 0.2196403657219113, p025: 0.17658259553140288, p975: 0.27020046374673834}, verdict: "U4_OK"}. That artifact is not modified by this run and is not overwritten.

2. Planned run

Exact planned command:

bash hierarchical-aif/scripts/launch_B4C11_corrected_full.sh 2000
# equivalently:
python hierarchical-aif/scripts/run_c11_corrected_full.py 2000 \
  hierarchical-aif/results/motor_stack_aif/B4C11_CORRECTED_FULL_RESULT.json

D3 does not apply to C11. The frozen C11 seeding is pure arithmetic (seed_base + b); there is no hash() anywhere in the U4 path. The corrected harness reproduces that arithmetic seeding byte-for-byte and does not substitute seeding.stable_seed.

3. The one and only deviation: corrected cluster-bootstrap semantics

The committed cell_C11 assembles each replicate's events and rebuilds b3.Cohort(f"C11_b{b}", …). Cohort groups train_by_motor by motorId, so a motor drawn K times collapses into one group holding K copies rather than K exchangeable groups. m7_train_nll iterates train_by_motor, so a K-fold group contributes L_m^K and over-sharpens that motor's latent.

Measured (not asserted), seed_base = 20260717:

Source: hierarchical-aif/results/motor_stack_aif/C11-PAIRED-DIAGNOSTIC.json sha256 fad54c32c1e0e6e7d75290f70926743b3b776e7fd088864f4285962ea362f630. At b = 0 the largest legacy group inflates 70 → 153 events (42.5% cluster loss).

The correction is motor_stack_aif.bootstrap.build_bootstrap_cohort: it builds the cohort identically to the legacy path (same split derivation, same bootstrap-resampled per-state scales, same normalised _y, bit-identical flat train_y) and then rebuilds only train_by_motor as one group per draw, with bootstrap_group_origin carrying the source motorId as metadata only.

Nothing else changes: not the threshold, not the criterion, not N_boot, not seed_base, not the 26-start L-BFGS-B _fit_m7_reduced budget or tolerances, not the cohort, not the model.

4. Frozen criteria (unchanged — do not touch)

• collapseFraction_tau_lt_1e_3 = #{tau_hat < 1e-3} / completed
• collapseFraction >= 0.25 → UNSTABLE_DISPERSION_U4_FIRES; otherwise U4_OK.
• Per replicate: _fit_m7_reduced = L-BFGS-B only, exactly 26 starts (5×5 product of k = linspace(0.05, 5.0, 5) × tau = exp(linspace(log 1e-4, log 5.0, 5)), plus the full-data optimum as the 26th), bounds log k ∈ [log 0.05, log 5.0], log tau ∈ [log 1e-4, log 5.0], ftol 1e-12, gtol 1e-10, maxiter/maxfun 20000, maxls 100, finite_diff_rel_step 1e-6, minimising b3.m7_train_nll(k, tau, coh.train_by_motor).
• k_full / tau_full (the 26th start) come from frozen B3 fitted.M7_HIERARCHICAL_MOTOR.kTau in audits/phase-b/b3-model-competition-result.json (sha256 5d7a0589e94de6b10f425f2d483e1e2a8f899d336aa59c335990209795e6b2bd).
• NO FLOOR. A non-finite log density or a None fit increments failed. It is never replaced, clipped, or floored.

5. U1 / U2 / U3 — decision and written justification

Why U2 is carried forward, not rerun. U2 is a deterministic 61-point profile-likelihood scan (tau_grid = exp(linspace(log 1e-4, log 5.0, 61)), flat-set offset 1.9207 = 0.5·chi2_{1,0.95}, logspan_normalized >= 0.50 → UNIDENTIFIED_U2_FIRES) computed on the full, unresampled derived_eligible_1_to_8 cohort. It never calls the bootstrap, never constructs a resampled cohort, and never touches train_by_motor grouping other than the frozen 1:1 motor grouping. D1 is a bootstrap-resampling defect and therefore cannot reach U2. Recomputing it could only reproduce the same numbers at real cost, and a recompute-then-relabel would be an untraceable duplicate of frozen evidence.

The corrected harness therefore copies the U2_profile object byte-for-byte out of the frozen artifact at run time (via _bridge.b4_result()["cells"]["B4C11"]["U2_profile"]), records the source path and its sha256, and labels it "provenance": "CARRIED_FORWARD_FROM_FROZEN_ARTIFACT". It is not recomputed and not relabelled as new evidence. Recorded values being carried forward: verdict = "U2_OK", flatLogspan_normalized = 0.06666666666666665, flatSetTauRange = [0.11332624602040679, 0.2331283928719889], nllStar = 575.7067684471608.

Consequence, stated plainly: the corrected run's M7_status is a composite of one newly computed check (U4) and three carried-forward checks (U1/U2/U3). It is not a fresh full identifiability assessment and must never be reported as one.

6. Pre-committed outcome branches

Falsifier of branch (a): any corrected full-N run with collapseFraction >= 0.25. Falsifier of branch (b): any corrected full-N run with collapseFraction < 0.25. There is no outcome under which the withdrawn artifact is restored.

7. Directional prediction I am willing to be wrong about — with its mechanism

7.1 The mechanism argument (direction)

The defective builder concatenates a motor drawn K times into a single group whose likelihood is L_m^K. That over-sharpens each duplicated motor's latent eta_m, pushing the per-motor posterior mass toward its own mode and inflating the apparent between-motor spread that tau must explain. Removing the collapse should therefore reduce fitted tau, and lower tau is closer to the collapse boundary. Mechanistically, the correction should make tau-collapse MORE likely, not less.

7.2 The measured evidence (magnitude)

Paired diagnostic, identical draws in both arms, N = 5 (C11-PAIRED-DIAGNOSTIC.json, sha256 fad54c32…):

Corrected tau is lower in 5 of 5 paired replicates (medians 0.19906 → 0.15627). The mechanism's direction is confirmed by measurement. But the magnitude is ~0.03–0.08 in tau, while the collapse threshold is 1e-3. The smallest corrected tau observed is 0.1186 — roughly 119× the collapse threshold. A shift of that size does not come close to producing a collapse.

7.3 What I commit to

1. Direction (committed, high confidence): the corrected tauHatSummary.median at N = 2000 will be strictly less than the frozen recorded 0.2196403657219113. I expect it in [0.12, 0.20].
2. Collapse fraction (committed, this is the risky one): collapseFraction_tau_lt_1e_3 will be < 0.05, and most likely exactly 0.000. I commit to a pre-registered interval of [0.000, 0.02]. → I am predicting branch (a).
3. Corrected p025 of tau will remain > 1e-2, i.e. two orders of magnitude above the collapse boundary.
4. Paired legacy-vs-corrected block: corrected tau will be lower than legacy tau in > 80% of the paired subset, and corrected group count will be exactly 80 in 100%.

7.4 How I could be wrong, and what that would mean

Five replicates cannot bound a rare-event fraction. A collapse fraction of 0.25 requires a heavy left tail of the corrected tau distribution that N = 5 has no power to detect; 0/5 above a 119× margin is suggestive, not conclusive. If prediction 2 fails and branch (b) fires, the honest reading is that the corrected bootstrap exposes a bimodal tau landscape — a subpopulation of resamples in which the 26-start optimiser finds the degenerate tau → 0 corner — that the diagnostic's five draws never sampled. That outcome is recorded as an outcome, not as a failure, and it is the more informative branch per §6.

I am predicting against the mechanism's own direction of concern. I am doing so on measured magnitude, and I am naming that tension rather than hiding it.

8. Old-vs-new comparison plan — the D1 effect size must be MEASURED, not asserted

The corrected run additionally executes a paired subset (default 25 replicates, seeds b = 0 … 24 — the same seeds, the same draws) through bootstrap.build_bootstrap_cohort_LEGACY_DEFECTIVE, and records both arms side by side under

"legacyVsCorrectedPaired": {
  "label": "LEGACY_DEFECTIVE_FOR_COMPARISON_ONLY",
  "disclaimer": "The legacy arm reproduces defect D1. It is a defect-magnitude measurement. It is NOT evidence, NOT a result, and licenses NO verdict."
}

Recorded per paired replicate: legacyGroups, correctedGroups, nDistinctDrawn, legacyTau, correctedTau, legacyK, correctedK, deltaTau = correctedTau - legacyTau; plus medians, the sign count, and the fraction of paired replicates in which corrected < legacy.

No verdict may ever be drawn from the legacy arm. It exists only so the D1 effect size is a measured number in the record.

9. Lane scoping — explicit

Affected: LANE B, C11 U4 only.

• The withdrawn U4_OK for B4C11_M7_STRUCTURAL_IDENTIFIABILITY.
• M7_status for the derived_eligible_1_to_8 cohort, insofar as it depends on U4.
• Defect D1_C11_CLUSTER_COLLAPSE.

Unaffected, and stated as unaffected:

• LANE A — duration-only B3/B4 held-out scoring. Unchanged. The M2-over-M3 adverse result (~0.0369 nats event-pooled) is unchanged and is reported alongside whatever this run returns.
• B4C10 (B4C10_CORRECTED_FULL_RESULT.json, sha256 959a00e974641eca1c0d6f3c2f7322b8c6c8411f68ca953c7e169492c4a53dde). M4 fits the flat pooled coh.train_y, which the correction leaves bit-identical (test_m4_pooled_path_is_unchanged_by_the_fix). Unaffected unless a C11 result actively contradicts it, which it structurally cannot.
• C11 U1, U2, U3 — settled in B3 without any bootstrap. Unaffected by D1 and unaffected by this run (see §5).
• LANE C (mark process) — untouched. This run is duration-only and never reads a mark channel.
• LANE D (motor-stack AIF) — untouched. M7 is a frozen B3 competitor, not the F-side model.
• LANE E (parity ladder) — no P-level moves on this run alone.

C10's favourable result may NOT be transferred to C11. They are different models (M4 mixture vs M7 hierarchical) with different likelihood structures (pooled i.i.d. vs motor-grouped). A favourable C10 says nothing whatsoever about M7 identifiability or about the withdrawn C11 U4_OK.

10. Required pre-run checks — each with the test that proves it

Command that must be green before launch:

python -m pytest hierarchical-aif/tests/motor_stack_aif -q
# baseline at HEAD b9b5670: 71 passed, 1 xfailed

11. Expected runtime — an operational estimate, not a measurement

The recorded figure is ≈20.1 h. Its basis: hierarchical-aif/ledgers/HIERARCHICAL-AIF-DEFECT-LEDGER.md line 50, D2 evidence row — a measured 36.2 s for a single _fit_m7_reduced, projected as 2000 × 36.2 s = 72 400 s ≈ 20.1 h. That is a projection from a single-fit measurement, not a measurement of the run.

D2's lesson is precisely that unmeasured cost estimates were wrong by 17–29× (C01 recorded 250–400 h vs ≈14.5 h measured; C02 recorded 150–250 h vs ≈8.7 h measured), and those bad estimates caused real cells to be recorded NOT_RUN and resourceBoundPartial. So this record does not rely on the projection.

Independent measurement already in hand: the paired diagnostic ran 5 replicates × 2 arms = 10 _fit_m7_reduced fits in 307.5444803237915 s ⇒ 30.75 s per fit (C11-PAIRED-DIAGNOSTIC.json). Projection from that: 2000 × 30.75 s = 61 509 s ≈ 17.1 h.

Smoke-test measurement (recorded before the full run — see §11.1). A N_BOOT = 3 smoke run of the corrected harness to a throwaway path measured the real per-replicate cost of this harness, including cohort assembly and the paired legacy arm. The projection below is re-derived from that measurement and supersedes both figures above.

11.1 MEASURED — smoke test of run_c11_corrected_full.py

Two smoke runs of the corrected harness, N_BOOT = 3, to throwaway scratch paths (never to the canonical result path). Both completed 3/3 with failed = 0.

Per-arm cost, derived from A and B:

corrected arm : 94.99591636657715 / 3           = 31.6653 s per replicate
legacy arm    : (130.4045786857605 - 94.99592) / 2 = 17.7043 s per fit

The legacy arm is cheaper per fit, as expected: it iterates ~46–53 collapsed groups instead of 80, so m7_train_nll does less work. That asymmetry is itself a signature of D1.

Re-derived full-run projection (this is the operative estimate):

corrected arm : 2000 × 31.6653 s = 63 330.6 s = 17.59 h
paired legacy :   25 × 17.7043 s =    442.6 s =  0.12 h
projected total ≈ 63 773 s ≈ 17.7 h   (single process, 1 core, no parallelism)

The recorded 20.1 h is a 1.13× overestimate of this projection — the same direction as D2, but nowhere near D2's 17–29× magnitude. The 20.1 h figure was, for this cell, roughly right.

The projection assumes per-replicate cost is stationary across seeds. A resample whose 26-start optimiser wanders costs more, and the smoke measured only 3 seeds (b = 0,1,2), so treat 17.7 h as a central estimate with a plausible band of roughly 14–24 h, not a bound.

Additional smoke observations (all measured, all pre-run):

• seedEquivalenceCheck = IDENTICAL_TO_FROZEN_INLINE_DRAW_SEQUENCE for b = 0,1,2.
• Corrected groupCountMin = groupCountMax = 80 — one group per draw, every replicate.
• nDistinctDrawn ranged 46–53, so duplicates were genuinely present and genuinely un-collapsed.
• Corrected tauHats = [0.1186009353493865, 0.20020183204985909, 0.15626676308339552] — bit-identical across runs A and B, and bit-identical to the pre-existing paired diagnostic. Determinism holds.
• Paired arm at b = 0,1: legacy groups 46, 53 vs corrected 80, 80; legacy tau 0.19906013275874299, 0.23470934095126772 vs corrected 0.1186009353493865, 0.20020183204985909; fracCorrectedBelowLegacy = 1.0.
• verdict = PARTIAL_NOT_ESTABLISHED, runStatus = PARTIAL_NOT_ESTABLISHED, predictionOutcome = NOT_ESTABLISHED at N = 3. The harness correctly refuses to emit a U4 verdict from a partial run. The smoke collapseFraction = 0.0 is a partial-run number and licenses nothing.

Checkpointing every 25 replicates to B4C11_CORRECTED_FULL_PROGRESS.json (verified written by the smoke run, self-labelled PARTIAL_PROGRESS_CHECKPOINT_NOT_A_RESULT) bounds worst-case crash loss to ≈13 minutes of compute.

Runtime is an operational number: filling this section in before launch changed no criterion, threshold, N, seed, model, or outcome branch. §4, §6, and §7 are untouched.

Checkpointing every 25 replicates to B4C11_CORRECTED_FULL_PROGRESS.json bounds worst-case loss to ~11 minutes of compute. Checkpoints are progress, never results — the file self-labels PARTIAL_PROGRESS_CHECKPOINT_NOT_A_RESULT.

12. What this cell cannot do, whatever it returns

• It cannot establish biological parity, mechanism, or active inference.
• It cannot promote M7 to "the correct model". U4 is an identifiability/stability check on one parameter of one frozen competitor on one cohort — not a predictive claim and not a mechanism.
• It cannot move P6 structural/mechanistic on its own.
• It cannot repair, restore, or re-credit the withdrawn 30-replicate artifact.
• It cannot transfer to C10, to M4, to the mark process, or to the F-side motor-stack AIF model.
• The adverse B3 result (M2_LOGNORMAL out-predicting M3 by ≈0.0369 nats event-pooled; narrowest motor-equal contrast M4_MIXTURE_K3 at 0.083461 against a resolution floor of ≈0.042 nats half-width) is retained and reported alongside this result, never instead of it. M2 is an adversarial baseline, never the UNI model.

13. Wording

Allowed: "The corrected full run of B4C11 at frozen N_boot = 2000 supports / does not support M7 tau stability (U4) on the derived_eligible_1_to_8 cohort under the frozen collapseFraction >= 0.25 criterion." · "candidate model" · "target hypothesis" · "CI-bound verdict" · "not established" · "retrospective-only" · "transfer required" · "intervention required" · "mechanism discriminator pending" · "corrected full run" · "U4 re-established on the corrected run only" · "the withdrawn U4_OK remains withdrawn".

Forbidden (mechanically checked by claim_guard.py). The canonical forbidden-wording list is maintained in claim_guard.FORBIDDEN and in the H-AIF operating contract; it is referenced here rather than re-transcribed. Re-transcribing it across wrapped lines defeats the guard's own use/mention rule — the guard only accepts a negation or catalogue cue on the same line as the phrase, so a catalogue that soft-wraps reads as a bare assertion on every line after the first. This section previously carried such a wrapped catalogue and the guard correctly flagged 8 violations in it; the fix is to reference the list, not to loosen the guard.

Forbidden in this document specifically: any claim that the corrected run restores the withdrawn artifact; any transfer of B4C10's favourable M4 result to M7; any statement that U4 is settled before the run completes; and any unscoped sentence of the form "U4 is fine" or "U4 is broken". Every U4 statement names the run it came from and the N it ran at.

14. NEXT_ACT

launch_B4C11_corrected_full.sh 2000 — launched by the orchestrator, not by the preparing agent. On completion: record sha256, update hierarchical-aif/ledgers/HIERARCHICAL-AIF-DEFECT-CLOSURE-LEDGER.md D1 routing, append a FLOW card, and run claim_guard.py over the report batch.