ORCHESTRATE — Active-Flow Status

Snapshot: 2026-07-22T02:10Z · UPDATED: B4C02 landed at 02:01:15Z · HEAD: b9b5670602a2afde158b53ea5e8135180f8c02f5 Branch: hierarchical-aif/motor-stack · Two authorized commits this session — 897c8ab (B4C11 prediction record) and 28ce738 (B4C01 prediction record), one file each. Nothing pushed.

Alignment was the gate, not the destination. This records what was executed after it.

1. Executive verdict

The flow ran without hitting a STOP condition. B4C10 landed at full frozen N and was reported; B4C11 was prepared, smoke-tested, defect-fixed, and LAUNCHED; the F-side motor stack was built out, scored, and returned a CI-bound NOT_ESTABLISHED; and three new defects (D8, D9, plus a latent partial-run defect in the C11 harness) were found and routed. The test suite went from 71 passed / 1 xfailed to 438 passed / 1 skipped / 1 xfailed (three consecutive clean runs). Frozen evidence is byte-identical throughout. The claim guard reports 0 violations.

The two most important scientific statements produced, both adverse:

1. The F-side motor stack reproduces the frozen M7_HIERARCHICAL_MOTOR to 2.5e-7 nats. At the resolution this study can achieve it is not a new model but a re-derivation of the incumbent, and it buys nothing measurable over its own tau → 0 limit M1_WEIBULL (contrast +0.000615, half-width 0.0111). M2_LOGNORMAL and M8_EMPIRICAL_KDE still out-predict it on point estimate.
2. Prospectivity was being over-claimed (D9). Reports asserted predictions were "committed before execution." The commit graph falsifies that for B4C10 and for the F-side scoring. No measurement is affected; the epistemic grade was. This is now mechanically guarded. My own first D9 write-up was itself wrong and was corrected: I recorded the B4C10 result as uncommitted, when git log --diff-filter=A shows prediction and result entered in one commit (b9b5670). The conclusion is unchanged and stronger — same-commit introduction makes strict ancestry structurally unattainable, so there is no mechanical test left to launder. Found by the adversarial verification lane, and recorded rather than silently patched.

B4C11 was rescued from D9. On authorization, commit 897c8ab pre-registered its prediction record while the run was at 210/2000 replicates with no result on disk — so B4C11 will be a genuinely prospective cell, unlike B4C10 (prediction and result in one commit, structurally unrepairable) and the F-side scoring (result already existed). Two of the four cells in this batch now carry real prospective standing: B4C02 (satisfied, with a mid-run-commit caveat) and B4C11 (pending, ordering secured).

No P-level was raised by any of this work. P8 remains FULL_PARITY = false; the first unsatisfied level is still P4 transfer.

2. What ACTs were performed after alignment

3. Contract recheck

No STOP condition was encountered.

4. C02 status — COMPLETE (landed 2026-07-22T02:01:15Z)

Full frozen N = 200 per generator, 600 sims, ZERO failures, resourceBoundPartial: false, runStatus = ELIGIBLE_FOR_FROZEN_VERDICT. Runtime 29 409.1 s = 8.17 h against a recorded RESOURCE_BOUND claim of 150–250 h. Result sha256 0633988dbfd690c0c0d12075dba4e0d8c25ddd178125064bd01fbdaf4629e398.

gensWithM2overM3 = 1 of 3 → GENERATOR-SPECIFIC → the frozen prediction GENERATOR-ROBUST_ADVERSE is REFUTED at full N. H_SHAPE_ARTIFACT is weakened: the adverse M2-over-M3 result is not generic to heavy-tailed dwell shape. Per the committed mapping this strengthens the case that the M2-vs-M3 contrast carries mechanism-relevant information — and establishes no mechanism. The separation is extreme (0.0050 / 0.9400 / 0.0050), not marginal.

I predicted GENERATOR-ROBUST_ADVERSE and was wrong. That is the recorded outcome.

Prospectivity: SATISFIED — the only cell in this batch. Prediction committed 2026-07-21T22:44:31Z, observation 2026-07-22T02:01:15Z: committed 3 h 17 min before the observation. Caveat stated: the commit was mid-run, with generator 1 already complete and failing the threshold — so the prediction was committed against the partial evidence, not tuned to it. Graded PROSPECTIVE_WITH_MID_RUN_COMMIT_CAVEAT.

Report: reports/B4C02-CORRECTED-FULL-REPORT.md. The real-data adverse M2-over-M3 finding is unchanged.

5. C10 status — COMPLETE

Full frozen N_boot = 2000 on the byte-identical frozen runner (3e21edac…), --cells C10 --c10-boot 2000. Result sha256 959a00e974641eca1c0d6f3c2f7322b8c6c8411f68ca953c7e169492c4a53dde, runtime 8372.7 s, completed 1994 / failed 6, resourceBoundPartial: false.

Every observed value sits far from its threshold — this is a well-separated result, not a marginal call. The frozen prediction UNIDENTIFIED_OR_WEAK is REFUTED at full N, within B4C10 scope only.

Boundary, stated hard: C10 may NOT be used to repair C11. Different model (M4 pooled-i.i.d. mixture vs M7 motor-grouped hierarchy), different likelihood structure. Identifiability is not correctness and is not mechanism.

Prospectivity: NOT_SATISFIED — prediction committed 2 h 33 min after the run finished (D9). Recorded as a retrospectively-graded refutation, never as PROSPECTIVE.

6. C11 — prepared, corrected, and LAUNCHED

RUNNING. PID 26756, started_utc=2026-07-22T01:40:03Z, frozen N_boot = 2000 + 25 paired legacy-vs-corrected diagnostic replicates.

• Seed equivalence PROVEN and enforced: the harness reproduces the frozen inline draw sequence (seed_b = 20260717 + b) and aborts if it does not.
• Frozen spec reproduced exactly: threshold collapseFraction ≥ 0.25, 26-start L-BFGS-B via b4._fit_m7_reduced (not reimplemented), no floor, N = 2000.
• U1/U2/U3 carried forward verbatim from the frozen artifact with source path + sha256, labelled CARRIED_FORWARD_FROM_FROZEN_ARTIFACT. U2 is a deterministic profile scan that never touches the bootstrap, so D1 cannot reach it.
• The submitted U4_OK — 30 of 2000 replicates under the defective bootstrap — remains withdrawn regardless of what this run returns, and the harness records that in its own output.
• Measured ≈24.7 s per M7 refit → ≈14 h projected, against a recorded 20.1 h estimate.

Its progress log is not a result. No C11 number is quoted or interpreted in this report.

Prospectivity: NOT_SATISFIED today — but FIXABLE, and the window is open. The prediction record is untracked; the C11 result does not exist yet. Committing protocols/B4C11-CORRECTED-FULL-PREDICTION.md before that result lands would make B4C11 a genuinely prospective cell. That requires the human principal. See §17.

7. F-side motor-stack build status

BUILT. 2 free parameters (mu, tau); per-motor latents integrated by 33-node Gauss-Hermite, never estimated, so parameter count is independent of motor count. Censoring-correct hazard/survival, NO_FLOOR, motor-equal scoring, motor-resampling bootstrap. G-side remains fenced: a test asserts expected_free_energy does not exist, and the reason recorded is structural (empty action set in a passive dataset), not sample-size-limited.

Verified in the main loop, not merely asserted:

• motor_log_marginal agrees with an independent adaptive-quadrature oracle to 4.7e-14 in log space, and Gauss-Hermite is converged (n=17 off by 1.5e-8; n=33…257 stable to 4.6e-14).
• A hierarchy.py domain guard changed during this session was audited: k ≥ 1/170 is derived from the IEEE overflow bound of the mean-one Weibull scale (exp(lgamma(172)) overflows, exp(lgamma(171)) = 7.26e306 is finite), not tuned. The previous 1e-6 guard could only ever crash, never return a value, so no previously-working result can change.

8. F-side scoring status — EXECUTED

Result b3b12720f32c0aee3bfa456f52ae0901976e59e3b43c0f2690fa7a17386ab297.

• Determinism PROVEN: full pipeline executed twice, canonical bytes compared, BYTE_IDENTICAL.
• Independent oracle PASS with residual exactly 0.0 on three frozen models (M3, M0, M8), tolerance 1e-12. B3 stores only aggregates, so every per-motor array was recomputed — which is why this check is mandatory. The harness halts rather than warning if it fails.

Pre-committed branch (B). Not (A) — it does not beat the control or the serious adversaries. Not (C) — no adversary resolved below 0, so the hypothesis is not falsified either. NOT_ESTABLISHED is not equivalence; with 19 holdout motors most contrasts were expected to be inconclusive, and were.

Prediction scorecard: 7 of 8 hits, 1 recorded miss. The miss (item 6) is instructive and is reported, not buried: the M5_GAMMA contrast resolved at a half-width of 0.0250, below the ~0.042 nat "floor" — because that floor is a heuristic from the narrowest frozen B3 contrast, not a bound, and a paired motor-cluster bootstrap resolves on consistency of sign, not magnitude alone. A calibration caveat is recorded: the competitors' published holdout scores were already legitimately in hand, so most contrasts followed by arithmetic once the point location was predicted. The scorecard measures the reasoning, not new evidence.

9. Defect closure ledger status

test_defect_closure_ledger.py now mechanically enforces that every defect names an affected_lane and unaffected_lanes, carries a routed closure_status, and claims only test files that exist — the guard that would have caught D8 on the day it was written.

D2 pattern, measured not asserted: every re-measured cost has come in below the figure that justified a partial run — C10 at 2.3 h, C11 projecting ≈14 h against 20.1 h.

10. Biological parity receipt map status

Updated for LANE B (C10 landed) and LANE D (built and scored), each with current receipts, next receipt, falsifier, and status. The required framing is intact:

Full biological parity is not a current status. It is the target world defined by these receipts. — and every lane asks "what would make this lane true?", not "why this lane fails."

LANE D's next receipt is now concrete and cheap: per-motor arrays for M4, M6 and M7 so the three models that match or out-score the candidate can be contrasted under the same paired bootstrap. It needs no new data.

11. Claim guard result

CLAIM GUARD: 0 violation(s) across 6 path(s).

Report: reports/CLAIM-GUARD-ACTIVE-FLOW-REPORT.md. A working CLI was added — the documented python -m hierarchical-aif.src... path is not importable (hyphenated directory) and the module had no entry point. It exits non-zero on violation so it can be gated.

It caught 8 real violations in the draft C11 prediction record: a forbidden-wording catalogue that soft-wrapped, defeating the guard's own same-line use/mention rule. Fixed by referencing the canonical list rather than loosening the guard.

Passing this guard is necessary, not sufficient. It checks wording, never whether evidence supports a claim.

12. Existing P0–P8 mapping

The unscoped "P6 is weakened" formulation that still sat in the ladder map was replaced with the four scoped statements the contract requires.

13. What got repaired

• C11 harness reporting IDENTIFIED_ON_THIS_COHORT on a partial run (the D2 failure mode).
• An independent-oracle test whose own precondition was unsatisfiable — fixed by tightening the oracle (epsrel=1e-12), not by loosening the comparison, with the tolerance derived from measurement and a negative control added.
• Two ledger-claimed tests that did not exist (D8).
• D7's contradictory status, reconciled on a substantive change of state.
• The unscoped P6 statement in the ladder map.
• Prospectivity over-claims across four prediction records and two reports (D9).
• A missing claim-guard CLI.

14. What got scored

The F-side motor stack, against CONTROL_CURRENT (M3) and five adversaries, on the frozen duration-only holdout split, motor-equal, CI-bound. Verdict NOT_ESTABLISHED against the control and every serious adversary. The retained adverse lognormal finding is extended, not overturned.

15. What got launched

B4C11 corrected full run, N_boot = 2000, 2026-07-22T01:40:03Z, PID 26756.

16. What remains running

17. What is blocked, and the exact receipt needed

18. Exact next ACT

NEXT_ACT = [B4C02 DONE - reported, hashed, ledger updated 2026-07-22T02:01:15Z]
           SUPERSEDED BLOCK, retained for the record:
           monitor PID 29340 (B4C02) to completion; on completion:
             sha256sum results/motor_stack_aif/B4C02_CORRECTED_FULL_RESULT.json
               -> results/motor_stack_aif/B4C02_CORRECTED_FULL.sha256
             write reports/B4C02-CORRECTED-FULL-REPORT.md, applying the outcome mapping
               already committed in protocols/B4C02-CORRECTED-FULL-PREDICTION.md
               (frozen criterion: gensWithM2overM3 >= 2 of 3 -> GENERATOR-ROBUST_ADVERSE)
             record that B4C02 IS genuinely prospective - its prediction record was committed
               2026-07-21T22:44:31Z, before any C02 observation existed (the only such cell)
             update the D2 row in the closure ledger; append a FLOW card; rerun the claim guard
           THEN continue supervising PID 26756 (B4C11, ~14-16 h) and, while it runs, recompute
             per-motor NLPD arrays for M4_MIXTURE_K3, M6_SEMI_MARKOV_STATE_DEPENDENT and
             M7_HIERARCHICAL_MOTOR from their frozen fitted params so the three models that
             match or out-score the F-side candidate can be contrasted under the same paired
             motor-cluster bootstrap - the cheapest remaining receipt for LANE D, needing no
             new data and touching no held-out mark channel

Updated NEXT_ACT after B4C02 landed

NEXT_ACT = (1) TIME-LIMITED, REQUIRES THE HUMAN PRINCIPAL: commit
               protocols/B4C11-CORRECTED-FULL-PREDICTION.md BEFORE the running B4C11 result
               lands (~14 h window). It is the only way B4C11 becomes a prospective cell;
               after the result exists the opportunity is gone permanently (D9).
           (2) supervise PID 26756 (B4C11, 30/2000 at snapshot). On completion: sha256 the
               result, write reports/B4C11-CORRECTED-FULL-REPORT.md against the committed
               prediction record, and move D1 from OPEN to its routed closure.
           (3) while it runs: recompute per-motor NLPD arrays for M4_MIXTURE_K3,
               M6_SEMI_MARKOV_STATE_DEPENDENT and M7_HIERARCHICAL_MOTOR from their frozen
               fitted params, so the three models that match or out-score the F-side candidate
               can be contrasted under the same paired motor-cluster bootstrap. Cheapest
               remaining LANE D receipt; no new data; touches no held-out mark channel.
           (4) B4C01 queued (~14.5 h measured). Commit its prediction record BEFORE launching.

This report is not a stopping condition. No STOP condition is in force.