Resource-Bound Reclassification (D2)

Gate: H-AIF-G4 · Measured: 2026-07-21 · Maps to existing ladder: P0/P1 provenance Machine: Windows 11 Pro 10.0.26200, Python 3.12.10, numpy 2.3.5, scipy 1.16.3, CPU-only Cohort: derived_eligible_1_to_8 (80 training motors / 793 training events / 19 holdout motors)

The recorded resource estimates in the committed B4 result are withdrawn. A RESOURCE_BOUND status is only honest if the resource claim is true.

1. Measured per-unit runtimes

Each figure is a single timed call on the frozen cohort. Reproduce with python hierarchical-aif/scripts/../../<scratch>/verify_timing.py.

C01 and C02 do not run a 9-model competition. The runner comment reads "skip M4/M7/M8 as they are the slow ones", and the cells record skippedModels = [M4_MIXTURE_K3, M7_HIERARCHICAL_MOTOR, M8_EMPIRICAL_KDE]. The cost driver is the simple model set plus M6.

2. Reclassification

Total for the four schedulable cells: ≈45.4 h of CPU-only compute, versus the ≈433–650 h implied by the recorded reasons.

3. Consequences

• B4C10 should never have been partial. At 2.1 h it was always within reach. Its bootstrap method is sound (M4 fits the flat pooled train_y, so duplicates enter correctly), so only the replicate count was ever at issue.
• B4C02 is the priority. It is the HIGH-risk misspecified-world discriminator — the cell designed to test whether the adverse lognormal result reflects heavy-tailed dwell shape rather than one assumed mechanism. It was presented as unreachable at 150–250 h. It is ≈8.7 h. This is the most decisive missing evidence in the entire submission.
• B4C11 remains blocked on D1, not on cost. Running it at full N on the defective bootstrap would produce 2000 replicates of an invalid statistic.
• B4C01/B4C02 remain blocked on D3 (non-deterministic hash() seeding) regardless of cost.

4. Ordering rationale

1. B4C10   2.1 h   unblocked now; D1 does not affect it; method already sound
2. B4C11  20.1 h   requires the D1 bootstrap fix first
3. B4C02   8.7 h   requires the D3 seeding fix first; highest scientific value
4. B4C01  14.5 h   requires the D3 seeding fix first

Cheapest-and-unblocked first, then highest-value-once-unblocked. Every run at full frozen N; no reduced N is used, so no prospective reduced protocol is required. If any run must later be reduced, a prospective reduced protocol will be written before it starts.

5. Honest limits of this reclassification

• Timings are single measurements on one machine, not distributions. They are point estimates and could vary by a factor of ~2 on different hardware; they do not vary by the 17–29× needed to rescue the recorded figures.
• _fit_m4_reduced and _fit_m7_reduced were timed on the real cohort. Bootstrap replicates resample it, so per-replicate cost may differ modestly with the resampled event count.
• C01/C02 timings assume the synthetic-data generation cost is small relative to the fit cost. This was not separately measured.
• B4C09 was not re-timed. Its NOT_RUN status stands; only its stated reason is in question, and in the opposite direction (possibly understated). This is recorded as R7 in the defect ledger as REPORTED_UNVERIFIED.

None of these caveats changes the conclusion: the recorded justifications are wrong by more than an order of magnitude, and the cells are schedulable.