TRACK E VERIFICATION — MECHANISM DISCRIMINATOR (E1)

Verifier scope: code + recorded result artifacts + TRAIN-partition reads (declared inline). No holdout row was read, computed over, or aggregated by this verification. All holdout quantities cited below come from audits/phase-b/b3-model-competition-result.json and experiments/results/science-gates-report.json — already-spent published numbers.

Headline: the design contains three genuinely valuable findings (E0.1, E0.3, E0.4), and one loud CONTRADICTION on the D5 firewall. It is not decisive, its verdict-bearing decision rule is unreachable at 19 motors by its own arithmetic, and its negative-control battery has zero power against the most likely implementation bug. Recommendation stated in §9.

---

1. CHECK 1 — DOES IT RELY ON THE BURNED MARK CHANNEL? YES. CONTRADICTED.

1.1 The entire discriminator is built on the burned channel

CHECKED_AGAINST_CODE · HOLDOUT_MARK_CHANNEL_BURNED_RETROSPECTIVE_ONLY

Every primary statistic in §1.1 is a function of direction / jump / nextStateN:

statistic	mark dependence
Δ_exit	d_i = direction of event i
Δ_mem	a_i = direction of event i−1
Π_mem	direction × direction joint
E1-C2	direction (global swap)
E1-C5(ii)	jump (restrict to |jump|=1)
E1-S D_a…D_d	every rung carries p(N' | N) = nextStateN
ICC	the only mark-free statistic in the protocol

So 3 of 4 primaries, 2 of 8 controls, and 4 of 4 ladder rungs consume the channel that D5 permanently burned.

1.2 The document contradicts itself on prospectivity

CONTRADICTED · CHECKED_AGAINST_CODE

• §0 E0.5 asserts: "The arrival-direction observable is genuinely unspent."
• §6.4 asserts: "It is quasi-prospective, not prospective."

Both cannot stand. The binding fact is §6.4, and even that is too generous. The correct label under the D5 contract is HOLDOUT_MARK_CHANNEL_BURNED_RETROSPECTIVE_ONLY — not "quasi-prospective". The prior UltraCode track's read of "empirical marginals of direction/jump per state" is precisely the marginal that Δ_exit and Π_mem are built from. Δ_exit and Π_mem are not fresh statistics at all; they are a re-slicing of an already-read marginal.

Δ_mem (the lag-1 joint) is the one functional the repo has genuinely never computed — I confirm the code side of E0.5:

• grep -c "direction\|jump\|nextStateN" audits/phase-b/b3-model-competition-runner.py → 0. CHECKED_AGAINST_CODE. B3 never touched the mark channel. Credit to the track: this part of E0.5 is correct and non-obvious.
• But scripts/run-science-gates.py:312-314 scores holdout events through event_scores(holdout, mechanism) → sourceLogLikelihood (lib/source-first-passage.js, branches on event.direction) and memoryless_scores(holdout, memoryless) (branches on event["direction"]). CHECKED_AGAINST_CODE. The holdout exit-direction channel is spent inside the released science gates, independently of the UltraCode track.

Net: the channel is burned twice over; only the lag-1 bivariate functional is unread. "Genuinely unspent" is wrong.

1.3 The track spent MORE of the burned channel while designing the protocol — and mislabelled it

CONTRADICTED · HOLDOUT_MARK_CHANNEL_BURNED_RETROSPECTIVE_ONLY

§1.4 states: "The holdout was touched for counts only — no effect estimate on any holdout motor was computed."

That defence is not available, because §4 reports these holdout mark-channel statistics:

§4 quantity	what it actually is
holdout exit counts: 150 on / 83 off	a holdout direction marginal — the exact object that burned the channel
holdout arrival-labelled rows: 215	a holdout lag-1 direction derived count
motors with both arrival arms: 16 / 19	a holdout arrival-direction per-motor contingency
motors with both exit directions: 16 / 19	a holdout exit-direction per-motor contingency
per-motor row counts: 49, 43, 30, 11, …, 2, 0	per-motor arrival-labelled counts

I can corroborate the last row's provenance without touching holdout data: b3-model-competition-result.json → cohorts.derived_eligible_1_to_8.holdoutMotorEventCounts gives {50, 44, 31, 12, 10, 10, 10, 8, 8, 7, 7, 6, 5, 5, 5, 4, 4, 4, 3} (sum 233). The §4 vector is exactly each of these minus one, plus a 20th motor at 0 — i.e. it is the per-motor count of events possessing a lag-1 arrival label. CHECKED_AGAINST_RESULTS. That is a mark-channel derivative, not a "count".

A marginal is not an effect estimate — the track is right about that narrow point. But "counts only" is the wrong characterisation, and it matters: the arm sizes were known before §2.4 registered one-sided hypotheses and before §4 declared the effective N. A pre-registration written with holdout arm sizes in hand is weaker than one written blind, even if arm sizes are ancillary under the conditional permutation test of §2.5.

1.4 E0.1 burned holdout reads to verify something provable from two lines of source

CONTRADICTED (method) · CHECKED_AGAINST_CODE

E0.1 reports: "Verified over all 1349 events: direction != sign(jump) in 0 cases." That computation ranged over holdout direction/jump/nextStateN.

It was unnecessary. scripts/ingest-wadhwa-data.py:159-160:

"direction": None if next_state is None else ("on" if next_state > dwell["state"] else "off"),
"jump":      None if next_state is None else next_state - dwell["state"],

The redundancy is true by construction — both fields are pure functions of (next_state, state). No data read is required to establish it. E0.1's substantive conclusion is CORRECT and is the single most valuable finding in the document (see §5.1 below); its verification method spent holdout mark reads it did not need.

One edge case the track's blanket statement misses: jump == 0 maps to direction = "off" while sign(0) = 0. TRAIN_ONLY, I confirm 0 events with jump == 0 among 1048 train events and 0 direction/jump inconsistencies. Holdout: NOT_CHECKED — would require holdout mark access; requires prospective record first.

---

2. CHECK 2 — ARE THE ADVERSARIAL BASELINES AND THE CURRENT-DESIGN CONTROL NAMED?

Named: YES. Entered as competitors: NO. CHECKED_AGAINST_CODE + CHECKED_AGAINST_RESULTS · NO_DATA_ACCESS_NEEDED

The nine B3 models are confirmed present in the result artifact: M0_EXPONENTIAL, M1_WEIBULL, M2_LOGNORMAL, M3_TWO_TIMESCALE, M4_MIXTURE_K3, M5_GAMMA, M6_SEMI_MARKOV_STATE_DEPENDENT, M7_HIERARCHICAL_MOTOR, M8_EMPIRICAL_KDE.

§1.2 names all five required adversarial baselines (M0, M1, M2, M5, M8) under class (a), plus M4 and M3; assigns M6 to (b) and M7 to (c). The current-design control is named twice: the D-L-T model (lib/source-first-passage.js) and the memoryless competing-risk baseline (scripts/run-science-gates.py:176-196) both appear under class (b). Requirement met on naming.

But the naming is rhetorical, not operational. No E1 cell puts an adversarial baseline into a contest:

• E1 primary (Δ_exit / Δ_mem / Π_mem / ICC) is model-free. The baselines appear only as predicted zeros in a decision table. Since all seven class-(a) models predict Δ_exit = Δ_mem = Π_mem = 0 identically, the primary cannot rank them — it can only reject the class. The track says this (§3, "(a) is falsified as a complete account") and is correct to.
• E1-S ladder: D_a is "lognormal p(dt|N) × multinomial p(N'|N)" — per-state, in seconds-space. That is not M2. M2 is a single mean-one shape fitted in normalised y = dt/scale_N space and shared across all eight states (b3-model-competition-runner.py:196+; m0_logpdf(y) and siblings take no state argument — CHECKED_AGAINST_CODE). D_a is a new, untested, more flexible model. If D_b beats D_a, that is uninterpretable as evidence for state-dependent hazard unless D_a is at least as strong as the standing adversarial winner. Defect: M2_LOGNORMAL must be a floor rung of the ladder, or every ladder contrast is confounded with D_a's unknown quality.
• E1-X is the only cell that touches real baselines, and it uses exactly two (M2, M3).

---

3. CHECK 3 — CI-BOUND VERDICT WITH MOTOR AS RESAMPLING UNIT?

Resampling unit: YES. Estimand: NO — the point estimate is event-weighted. CONTRADICTED

Correct: §2.2 declares experimental unit = MOTOR; §2.5 resamples the 19 holdout motorIds with replacement carrying whole chains; §2.7 is CI-bound with "point estimates are never verdicts"; §5 E1-C3 explicitly runs the event-level bootstrap as a labelled pseudoreplication demonstration. That is correct discipline.

The contradiction: §2.2's registered primary estimator is Cochran–Mantel–Haenszel weighted,

Δ̂ = Σ_s w_s (ȳ_s,on − ȳ_s,off) / Σ_s w_s, w_s = n_on·n_off/(n_on+n_off)

The weights are event counts. B3's declared primary aggregation is "primary": "MOTOR_EQUAL (v3 R3)" (CHECKED_AGAINST_RESULTS, b3-model-competition-result.json → aggregation). E1's primary estimator therefore abandons B3's frozen motor-equal aggregation in favour of an event-weighted one. Three motors carry 125/233 = 54% of holdout events (CHECKED_AGAINST_RESULTS), so under CMH weights those three dominate the point estimate. A motor-cluster bootstrap corrects the variance; it does not correct the estimand.

This is not a stylistic quibble — it breaks §4:

§4 computes MDE from the per-motor unweighted between-motor sd (1.429, 1.443), which is the sampling sd of a motor-equal estimator. It does not describe the CMH estimator registered in §2.2. The power table and the registered primary are two different tests. One of them must change.

3.1 The corrected-bootstrap requirement is impossible as written

CONTRADICTED · CHECKED_AGAINST_CODE

§2.5: "E1 must import that corrected bootstrap or reproduce its test."

hierarchical-aif/src/motor_stack_aif/bootstrap.py cannot be imported for E1:

1. draw_motors(train_motors, rng) and _assemble(coh, sampled, …) resample coh.train and pass holdout_events = [dict(e) for e in coh.holdout] through unchanged. It resamples the wrong partition — E1 needs holdout motors resampled and train fixed.

2. _assemble rebuilds a full b3.Cohort, which recomputes scale_N from the resampled train (b3-model-competition-runner.py:115-120). E1 explicitly requires scale_N frozen. Importing it would unfreeze the normalisation per replicate.

3. The D1 defect does not even bite E1's primary. The module docstring states the defect scope precisely: train_by_motor grouping for M7's grouped likelihood is AFFECTED; M4's pooled flat likelihood is UNAFFECTED. For a CMH-weighted stratum contrast, collapsing a motor drawn K times is algebraically identical to keeping K distinct clusters: doubling n_on and n_off doubles w = n_on n_off/(n_on+n_off) exactly and leaves the stratum mean difference unchanged. Δ_exit / Δ_mem / Π_mem are invariant to D1 by construction.

Where D1 does bite E1 is ICC (a variance component computed per motor group: collapsing duplicates inflates group size and shifts var_between/var_within) and E1-S rung D_c (motor-level frailty). §2.5's blanket requirement points the reader at the wrong target and leaves the two genuinely at-risk statistics unguarded. Defect.

---

4. CHECK 4 — IS THE FROZEN SPLIT PRESERVED?

Substantially YES, with one specification inconsistency.

CHECKED_AGAINST_CODE. Line references verified exactly:

track cites	actual	status
sha256_mod5 at :88-89	def sha256_mod5 lines 88-89	exact
FAILED-SPLIT-MISMATCH halt at :104-108	halt at lines 104-108	exact
censoring exclusion at :109-110	elig = [… not e["rightCensored"] …] lines 109-110	exact
per-state training scale at :117-121	by_state loop 116-118, scale_N 119-120, _y at 130	±1, substantively correct
ingest direction/jump at :159-160	exact	exact
run-science-gates.py:176-196 memoryless	fit_memoryless 176-185, memoryless_scores 187-196	exact
run-science-gates.py:454-467 G06	gate("G06_…") block 453-467	±1

Frozen counts CHECKED_AGAINST_RESULTS: summary = {train: 793, trainMotors: 80, holdout: 233, holdoutMotors: 19} for derived_eligible_1_to_8. Matches §2.1 exactly. E1-C7 (split mutation must HALT) and the reassertion requirement are correct.

Inconsistency: §2.1 freezes the cohort at 233/19, but §2.3 repairs B3's right-censoring exclusion and makes a stratified log-rank on censored-inclusive data the primary estimator. The primary therefore runs on up to ~252 events while the runner asserts 233. The cohort E1 actually analyses is not the frozen B3 cohort. This is defensible science — inheriting an undeclared exclusion is worse — but the artifact must declare two cohorts with two counts, and §2.1's table as written would make the runner assert a number it does not use.

4.1 Unflagged: D6 corrupts the mark channel on holdout, and E1 consumes it

CHECKED_AGAINST_CODE · HOLDOUT_MARK_CHANNEL_BURNED_RETROSPECTIVE_ONLY

scripts/ingest-wadhwa-data.py:141-143 range-checks dwell["state"] and continues on failure. next_state = dwell["nextState"] at :147 is never range-checked — confirmed, the only occurrences of next_state/nextState in the file are lines 59, 73, 147, 158, 159, 160, none validating. Per established fact D6, this writes nextStateN = -1 for 2 holdout events.

Those two rows carry direction = "off" and jump = -1 − stateN (i.e. |jump| ≥ 1, plausibly ≥ 2) — a fabricated "off" exit and a fabricated "off" arrival for the successor event.

TRACK E NEVER MENTIONS D6. These corrupted rows land directly in:

• Δ_exit (spurious off arm member),
• Δ_mem (spurious off arrival for the following event),
• Π_mem (spurious cell),
• E1-C5(ii), the veto control, which keys on |jump| = 1 — the corrupted rows are exactly the large-|jump| rows the artifact hypothesis (e) predicts.

With 16 effective motors and a median holdout motor carrying 5–8 events, two corrupted rows landing in a small motor can move that motor's per-motor Δ materially. The protocol contains no exclusion rule, no sensitivity analysis, and no mention. This must be a pre-registered exclusion with the count declared, not discovered at analysis time.

TRAIN_ONLY control read: 0 train events have out-of-range nextStateN, so the defect is holdout-side only in this cohort — which is the worst place for it.

---

5. THE THREE FINDINGS THAT ARE GENUINELY GOOD — AND ONE THAT IS BROKEN

5.1 E0.1 (mark triple-counting): CONFIRMED, and it is the most valuable line in the document

CHECKED_AGAINST_CODE · NO_DATA_ACCESS_NEEDED. Writing p(N'_i, direction_i, jump_i | ·) would count one degree of freedom three times. Given N_i, the pair (direction_i, jump_i) is a deterministic function of N'_i (ingest:159-160). Any F_motor built on the four-coordinate blanket inflates the mark contribution ~3× relative to the dwell contribution. This is a real defect found before implementation and it should be propagated to whichever track owns the target-stack specification.

5.2 E0.3 (undeclared 0.300 s left truncation): CONFIRMED and under-sold

CHECKED_AGAINST_CODE + TRAIN_ONLY.

• grep -rn "truncat" across audits/phase-b, scripts, lib, experiments, hierarchical-aif (.py/.js/.mjs/.ts) returns zero hits. No competitor carries a truncation term.
• TRAIN_ONLY (declared: 793 train, states 1–8, uncensored): min = 0.300 s, 22 events at exactly 0.300 s, 98 events below 0.5 s, and the support is quantised on a 0.02 s grid (0.30, 0.32, 0.34, …). The floor is 15 samples and it is present in the train partition alone — so t_min is freezable from train and E1-X is executable without any new holdout read.
• hierarchical-aif/protocols/B4C02-CORRECTED-FULL-PREDICTION.md names exactly the three generators the track claims (weibull_gamma_blend, three_timescale_heavy_tail, per_motor_heterogeneous_weibull) and contains no mention of truncation, floor, or t_min. CHECKED_AGAINST_CODE. The track is right that B4C02 structurally cannot detect this.

I can strengthen the track's case quantitatively. Using scale_N from the published artifact (CHECKED_AGAINST_RESULTS): {1: 4.749, 2: 3.483, 3: 6.351, 4: 5.063, 5: 7.948, 6: 14.390, 7: 18.245, 8: 24.521}. The truncation point in the models' normalised y-space is y_min(N) = 0.300 / scale_N(N), ranging 0.0122 (N=8) to 0.0861 (N=2). Under a mean-one exponential the omitted renormalisation is −log S(y_min) ≈ y_min, i.e. 0.012–0.086 nats per event.

The retained adverse M2−M3 gap is 0.0369 nats/event (CHECKED_AGAINST_RESULTS, adverseLognormalRetention.derived_eligible_1_to_8). The omitted term is of the same order as, and at N=2 is 2.3× larger than, the effect it is invoked to explain. E1-X's #1 ranking in §7 is correct and I endorse it independently.

But E1-X has two defects the track did not catch:

(a) Leakage in t_min. §2.9 sets t_min = 0.300 s at "the observed detection floor" — observed over all 1240 uncensored dwells, i.e. train and holdout. §5 E1-C5 is scrupulous about this ("threshold frozen from the training partition"); §2.9 is not. A truncation parameter estimated with holdout durations in scope, used in a refit that is then scored on holdout, is a leakage path. TRAIN_ONLY verification above shows the floor is 0.300 s in train alone, so the fix is free: declare t_min train-frozen. As written, E1-X is HOLDOUT_ALREADY_SPENT_DURATION_ONLY-contaminated.

(b) The correction is written in the wrong space and breaks the model family. §2.9 writes log p(dt_i) → log p(dt_i) − log S(t_min | N_i). But every B3 competitor is a mean-one density in y, with a single shape shared across all eight states (m0_logpdf(y) etc. take no state argument — CHECKED_AGAINST_CODE). So (i) the correction must be − log S_y(0.300/scale_N(N_i)), which is state-dependent and therefore breaks the cross-state pooling that defines the fit; and (ii) a left-truncated mean-one density no longer has mean one, so the truncated refit is not a member of the frozen family. §2.9 must specify whether the mean-one constraint is imposed pre- or post-truncation. These are different models and they will give different answers.

5.3 E0.4 (per-state normalisation removed state-dependent means): CONFIRMED

CHECKED_AGAINST_CODE. b3-model-competition-runner.py:116-120, 130 divides each dwell by the training mean for its state before scoring. B3 therefore tested state-dependence of shape only, never of mean. Correct and material.

Consequence the track half-notices: because scale_N depends only on state and strata are (motor, state), the normalisation cancels exactly in every within-stratum difference. Δ_exit and Δ_mem are numerically identical whether computed on log y or log dt. Therefore E1-C8 ("compute scale_N from holdout → must HALT or flag LEAKAGE") is vacuous for all three primary contrasts — no leakage is possible through a constant that cancels. It is a real control only for E1-S. Registering an invariant-by-construction check as a mandatory leakage guard whose failure invalidates the primary is exactly the vacuous-success pattern the operating contract warns against. Demote E1-C8 to E1-S scope, or state explicitly that it is a code-path smoke test with no inferential content.

5.4 E0.2 (G06 already ran the naive test): numbers CONFIRMED, interpretation OVERSTATED

CHECKED_AGAINST_RESULTS. science-gates-report.json G06 reproduces every quoted figure to the last digit: status: FAIL, mechanistic −3.838083424564402, memoryless −3.8963417959275564, advantage +0.05825837136315377, interval [−0.015801079856724822, +0.13264344169039152], 2000 replicates, seed 20260717. Half-width = 0.0742 ✓. G05 status: FAIL, passedReplicates: 2 / 3 ✓.

CHECKED_AGAINST_CODE. The coupling argument is sound in structure: lib/source-first-passage.js sourceDensities gives plus = kPlus·survival and minus = Σ w_j r_j e^{−Λ_j t}, so direction and dwell are dependent under D-L-T; fit_memoryless gives constant cause-specific hazards, so they are independent under the baseline.

Overstatement: the track concludes "G06 is therefore already a dwell×direction-coupling test." It is not a clean one. D-L-T differs from the memoryless baseline in two ways simultaneously — the coupling and the dwell shape (multi-exponential mixture vs single exponential). The +0.058 nats is a sum of both contributions with no decomposition. G06 bounds the total predictive gain of the whole mechanistic model; it does not isolate coupling. The resolution figure (±0.074 nats/interval) is valid and useful; the causal reading is not. REPORTED_BY_TRACK, partially CONTRADICTED.

---

6. CHECK 5 — POWER. THE HONEST ANSWER IS: ALMOST NOTHING IS DETECTABLE, AND THE PROTOCOL'S OWN DECISION RULE IS UNREACHABLE.

NO_DATA_ACCESS_NEEDED — all of this is standard sample-size arithmetic over already-published interval widths.

6.1 The empirical resolution floor at 19 motors, from published intervals

CHECKED_AGAINST_RESULTS. b3-model-competition-result.json → cohorts.derived_eligible_1_to_8.contrasts.NLPD_motor_equal, every model vs M3, motor-cluster BCa:

model	point	BCa	width	verdict
M0_EXPONENTIAL	−0.11365	[−0.29763, +0.03305]	0.3289	INCONCLUSIVE
M1_WEIBULL	+0.00103	[−0.06815, +0.07204]	0.1399	INCONCLUSIVE
M2_LOGNORMAL	+0.02502	[−0.04370, +0.08676]	0.1285	INCONCLUSIVE
M4_MIXTURE_K3	+0.01022	[−0.02446, +0.05968]	0.0835	INCONCLUSIVE
M5_GAMMA	−0.02940	[−0.12351, +0.05984]	0.1801	INCONCLUSIVE
M6_SEMI_MARKOV	+0.00444	[−0.06918, +0.08741]	0.1552	INCONCLUSIVE
M7_HIERARCHICAL	+0.00164	[−0.07740, +0.08035]	0.1580	INCONCLUSIVE
M8_EMPIRICAL_KDE	+0.01181	[−0.02814, +0.05767]	0.0864	INCONCLUSIVE

Eight of eight INCONCLUSIVE. Not one CI excludes zero. Best achievable half-width on a per-event log-score contrast at 19 motors is ±0.042 nats; typical is ±0.064–0.090.

Add G06's ±0.074 nats/interval on the joint dwell×direction score, also FAIL.

Two independent predictive-contest routes, 9 density models plus a mechanistic model plus a memoryless baseline, and zero CI-excluding-zero outcomes. That is the calibration any new discriminator must beat.

6.2 This breaks E1-X's registered decision criterion

CONTRADICTED. §2.9 registers: "if the M2-over-M3 advantage narrows by more than half, or reverses, the adverse result is an artifact."

The M2-over-M3 advantage on B3's motor-equal primary is +0.02502 with BCa [−0.0437, +0.0868]. A "50% narrowing" of a quantity whose 95% interval spans from −0.044 to +0.087 is not a measurable event. Monte Carlo noise and refit variation alone will move a point estimate of 0.025 by more than half. The criterion is a point-estimate rule on a statistically null quantity.

Compounding it: the retained adverse finding is the event-pooled number, 0.0369 nats (adverseLognormalRetention), which carries no interval at all in the artifact. §2.9 does not say which of the two it targets. Estimand ambiguous, criterion untestable as written.

Repairable, and worth repairing: score the paired difference (M2−M3)_truncated − (M2−M3)_untruncated per motor, under the same motor-cluster bootstrap with common random numbers. Paired contrasts on the same holdout motors are dramatically tighter than either marginal, and — given §5.2's finding that the omitted term is 0.012–0.086 nats against a 0.037-nat gap — a paired design plausibly does have resolution here. This is the one E1 cell I judge to be worth running, and only after this repair plus the train-frozen t_min fix.

6.3 E1 primary: MDE arithmetic verified, then pushed through the actual decision rule

Track's numbers reproduce. Their MDE uses two-sided z = 1.96 + 0.842 = 2.802: 2.802 × 1.429 / √19 = 0.9186 ✓ (0.918); 2.802 × 1.443 / √19 = 0.9276 ✓ (0.927); @16 → 1.0009 ✓ (1.000). Required-N figures also reproduce: 0.44 → 84, 0.50 → 65, 1.32 → 9, 1.00 → 16, 0.92 → 19. G06 route (0.074/0.058)² × 19 = 30.9 ✓ → ~31 holdout → 155 total at the 1-in-5 split.

Two errors:

1. Sidedness mismatch. §2.4 registers one-sided hypotheses and §2.7 uses a one-sided lower bound, but the MDE table is computed two-sided. The one-sided MDE is 2.487 × 1.429/√19 = 0.815. The power table does not describe the registered test.
2. Design-effect slip. §4 writes 1 + (m̄−1)·ICC then computes 1 + 11.3(0.26) ≈ 3.9 — using m̄ where the formula says m̄−1. With m̄ = 215/19 = 11.32, the correct DE is 1 + 10.32(0.26) = 3.68, effective n ≈ 58, not ≈ 55. §5 E1-C3 applies the same formula correctly (1 + (12.3−1)(0.26) = 3.93). Internally inconsistent.

The error that matters — §4 never propagates power through §2.7. §2.6 explicitly designates the Bonferroni-0.9875 interval, not the nominal one, as verdict-bearing. One-sided α = 0.0125 → z = 2.2414:

Verdict-bearing MDE = (2.2414 + 0.8416) × 1.429 / √19 = 1.011 log units.

The training effect is +0.923. The MDE of the interval that actually decides the verdict is 1.011. The protocol is underpowered against its own registered effect size, before any control is applied.

Carrying it through all four §2.7 conditions, at the optimistic assumption that the holdout effect equals training exactly:

condition	power
(1) Bonferroni one-sided lower bound > 0, n = 19	z = 0.923·√19/1.429 − 2.2414 = 0.574 → 72%
(1) at effective n = 16 (16/19 motors have both arms)	z = 0.343 → 63%
(3) E1-C5 restricted, nominal, tolerating 50% attenuation	z = 0.4615·4/1.429 − 1.645 = −0.353 → 36%
(2), (4)	~1 (permutation p and null control track condition 1)

Joint probability of emitting MEMORY_SUPPORTED ≈ 0.63 × 0.36 ≈ 23%, at the optimistic effect size. If the holdout effect is half of training — entirely ordinary, and expected here because the direction was selected from training by inspecting 8/8 sign consistency, i.e. winner's curse — joint power falls below 5%.

§2.7 condition 3 is internally self-contradictory. It tolerates ≥50% attenuation while demanding the attenuated estimate's CI still exclude zero. Detecting 0.46 log units at the nominal one-sided bound needs n = (2.487×1.429/0.46)² = 60 motors; at the Bonferroni bound, n = (3.083×1.429/0.4615)² = 91. The ≥50% clause does no work: the only way condition 3 passes is if the effect does not attenuate at all. And the |jump| = 1 restriction of C5(ii) discards ~25–30% of rows and dissolves strata that lose an arm, shrinking effective N further — a cost §4 never books.

6.4 What N would actually be required

Per-motor sd from training: 1.429 (Δ_mem), 1.443 (Δ_exit). Frozen split is 1-in-5.

target	holdout motors	total motors	available (99–109)
Δ_mem @ 0.923, Bonferroni one-sided, 80%	23	115	✗ (19 / ~99–109)
Δ_mem @ 0.46 (the attenuation §2.7 tolerates), Bonferroni	91	456	✗
Δ_exit @ 0.44, two-sided nominal (track's own)	84	420	✗
Δ_exit @ 0.44, Bonferroni	102	510	✗
G06-sized predictive gain (+0.058 nats)	31	155	✗
MEMORY_ABSENT at ±0.25 margin	≳ 300	≳ 1500	✗

Not one target is reachable. The single closest — 23 holdout motors for the un-attenuated optimistic Δ_mem — still exceeds the 19 available, and it is exactly the case in which the winner's-curse correction bites hardest.

Plain answer to the question asked: at 19 holdout motors, almost nothing is detectable. The one statistic whose training effect even approaches the resolution floor (Δ_mem) is powered at roughly 63–72% on its first condition and ~23% jointly, under assumptions that are systematically optimistic. MEMORY_SUPPORTED is a pre-registered near-certain NOT_ESTABLISHED. No reanalysis of Wadhwa 2022 changes this. INDEPENDENT_TRANSFER_REQUIRED / PROSPECTIVE_NEW_DATA_ONLY.

Credit where due: §4 and §6.6 register MEMORY_ABSENT as unattainable in advance, so no null can be laundered into an absence claim. That is exactly right and is the strongest governance feature of the document. The failure is that §4 stops one step short — it never applies the same honesty to MEMORY_SUPPORTED.

---

7. CHECK 6 — NEGATIVE CONTROLS

All three required controls present. The battery has a structural blind spot.

required	present	assessment
label scrambling	E1-C1 within-stratum arrival-label permutation	present; see defect below
direction reversal	E1-C2 global on↔off swap	present but near-vacuous — a difference of means flips sign by algebra; this is an implementation smoke test, not a scientific control. Referenced file tests/semantic/orientation-direction-and-score-sign.test.mjs exists (CHECKED_AGAINST_CODE).
event-level bootstrap (pseudoreplication)	E1-C3	best control in the document. Correctly framed as a demonstration, never a result; predicts CI ≈ 2.0× too narrow (√3.93 = 1.98 ✓); and mandates disclosure if the event-level CI would flip the verdict. Exemplary.

Plus C4 (lag profile), C5 (artifact veto), C6 (motor-ID scramble), C7 (split mutation → HALT), C8 (scale mutation).

7.1 The blind spot: eight negative controls, zero positive controls

This is the most serious methodological gap after the D5 issue.

Consider the single most likely implementation bug in this design: the lag-1 arrival join is wrong — off-by-one in the within-motor ordering, a silent None from the contiguity filter, or a merge that drops every arrival label. Such an estimator returns Δ ≈ 0 always. Walk it through the battery:

• C1 (scramble → null centred on 0): passes — an always-zero estimator gives a null centred on 0.
• C2 (reversal → sign flips, magnitude preserved to 1e-12): passes — 0 = −0.
• C3 (event bootstrap narrower): passes — both widths shrink toward 0.
• C4 (lag profile): passes, and worse — a flat profile is explicitly interpreted in §5 as scientific evidence ("evidence against a first-order latent and for a motor-level or artifact explanation"). A bug is converted into a finding.
• C5, C6, C7, C8: unaffected or pass.

An always-zero estimator passes all eight controls and emits NOT_ESTABLISHED, which is indistinguishable from the expected outcome. The battery has zero power against the failure mode most likely to occur.

Required addition: a positive control — inject a synthetic Δ_mem of known magnitude (e.g. 0.92 log units) into holdout-shaped surrogate data with the real motor/state/arm structure, and require the estimator to recover it within tolerance at the claimed MDE. Without this, a null from E1 is uninterpretable.

7.2 Missing control: within-motor non-stationarity

There is no control for time-varying covariates within a motor's recording. Stratifying on (motor, state) removes the motor's mean level and the state composition; it does not remove within-motor drift. Wadhwa 2022 is a stator-remodelling-under-load-change experiment (nominalElectrorotationSpeed is recorded per motor at ingest:130, CHECKED_AGAINST_CODE). If load, PMF, or bead attachment drifts across a recording, then events with arrival = on (concentrated during the motor's climb in N) systematically differ from arrival = off events (concentrated during descent) — at the same N, in the same motor, with no cross-event latent whatsoever.

Add: stratify on (motor, state, trajectory-half), or include the within-motor event index as a covariate, or permute arrival labels within time-blocks. Without it, explanation (f) non-stationarity is fully confounded with (d) memory and is not in the decision table at all.

7.3 E1-C1 is not computable as specified

E1-C1 demands "the observed-null 95% coverage within [0.93, 0.97]." Coverage is a property measured over repeated datasets, not over repeated permutations of one dataset. Measuring it requires a bootstrap nested inside each permutation — 20000 × 20000 resamples. §7 costs "E1 primary + controls C1–C8" at "minutes." Either the coverage criterion means something undefined, or the cost estimate is off by orders of magnitude. Given established defect D2 (resource estimates already found overstated 17–29× in this repo), unsupported cost estimates are a live governance issue here — this one runs the other way.

---

8. CHECK 7 — IS IT DECISIVE? NO. Brutally: it separates two of five explanations, and its veto control cannot do the job it exists to do.

8.1 What the design does separate

Genuinely correct and well-argued: within-(motor, state) stratification does structurally neutralise (c) motor heterogeneity (the motor effect is exchangeable within its own events, so it cancels in every within-stratum contrast) and (b) composition effects across states. Δ_mem ≠ 0 surviving stratification really would falsify a memoryless-in-N Markov account. That much is sound.

Also correct: §1.3 registers the adverse discovery that plain serial correlation of durations is dead (lag1 r = −0.0706, lag2 +0.0958, lag3 −0.1217, sign-alternating, all within noise — REPORTED_BY_TRACK, TRAIN_ONLY per the track's own declaration; I did not recompute) and correctly refuses to make it the primary. Retaining a dead statistic as C4 rather than deleting it is exactly the discipline the operating contract demands.

8.2 Defect 1 — the veto control cannot distinguish the two hypotheses it exists to distinguish

This is the central decisiveness failure.

E1-C5 has veto power: if Δ_mem loses >50% of its magnitude after excluding dt < 0.460 s, the signal is reassigned to (e) segmentation artifact and MEMORY_SUPPORTED is vetoed.

But (d) a genuine mechanistic memory would plausibly live in short dwells too. The repository's own gate G07_H_STATE_REEXTRACTION is about "the 43 short hidden-state wells and the reported k_h and k_-h" (CHECKED_AGAINST_CODE, run-science-gates.py:468+). A short-lived hidden intermediate state is a named mechanistic hypothesis in this very repository, and it predicts that the memory signal concentrates in exactly the short dwells C5 excludes.

So C5 is a magnitude filter that both (d) and (e) predict will behave the same way, and it is asymmetrically wired to veto. Under it:

• real short-lived latent → effect collapses → vetoed as artifact (false negative),
• segmentation flicker → effect collapses → vetoed (true negative).

The control has ~100% power against (e) and ~0% specificity against a short-timescale (d). A control that vetoes the hypothesis it is testing is not a discriminator.

C5(ii) — restrict to |jump| = 1 — is a genuine artifact discriminator (merged/split segments produce |jump| > 1; a stator-kinetic mechanism should not care). It should be promoted to the primary artifact control and C5(i) demoted to a reported sensitivity. And C5(ii) must first exclude the D6-corrupted rows (§4.1), which are precisely large-|jump| fabrications.

8.3 Defect 2 — the decision table omits two live explanations

explanation	in table?	predicts Δ_mem > 0?
(a) shape-only	yes	no
(b) state-dependent hazard	yes	no
(c) heterogeneity	yes	no
(d) policy/memory	yes	yes
(e) segmentation artifact	yes (added by track — good catch)	yes
(f) within-motor non-stationarity	NO	yes (§7.2)
(g) boundary/cohort-edge effects	NO	yes (below)

(g): the cohort is states 1–8, but ingest admits states 0–11. An event at N = 8 whose predecessor was N = 9 carries arrival = off derived from an event outside the cohort; an event at N = 1 whose predecessor was N = 0 carries arrival = on likewise. §1.1 says arrival is "defined only for events whose predecessor is contiguous and uncensored" — it never says whether the predecessor must be cohort-eligible. This is not cosmetic: it changes the estimand, and it likely explains the unreconciled 215-vs-214 count in §4 (233 − 19 = 214 arrival-labelled rows if predecessors must be in-cohort; §4 reports 215). The definition must be pinned before freezing, and the two variants will not give the same Δ_mem.

At the boundary states the arrival label is mechanically correlated with position in the state range, and the state range is a cohort-definition artifact, not a physical boundary. This produces Δ_mem ≠ 0 with no memory.

8.4 Defect 3 — the "every model predicts exactly 0" claim proves too much

§1.2 is the load-bearing argument: "Every model in the repository predicts Δ_mem = 0 exactly, for structural reasons."

That is correct — and it is the problem. It means Δ_mem is not a model discriminator; it is a goodness-of-fit residual test against the entire model class at once. A rejection tells you all nine B3 models plus D-L-T plus the memoryless baseline are incomplete. It does not tell you which replacement class is right, because (d), (e), (f), and (g) all predict the identical observable — a positive stratified Δ_mem — and the protocol's only tool for separating them is E1-C5, which §8.2 shows cannot do it, plus C4's lag profile, which is confounded by the same drift that generates (f).

So the answer to "would all four predict nearly the same observable?" is: yes for (d)/(e)/(f)/(g), and structurally-zero for (a)/(b)/(c). The design cleanly separates {a,b,c} vs {d,e,f,g} and then cannot separate within the second group — which is the only group where a mechanism claim could live. The track is honest that (e) is live and gives it veto power; it does not see (f) or (g) at all.

8.5 What the design would actually license, at best

§6.1 states it correctly and I endorse it verbatim: a confirmed Δ_mem ≠ 0 licenses exactly one statement — "a memoryless-in-N model is falsified for this dataset" — and licenses nothing about F, G, policy priors E, precision Π, message passing, or active inference. §6.8's note that a positive Δ_mem is fully consistent with a two-conformation chemical stator model with no inference-theoretic content is correct and important.

Given the burned channel, even that statement is retrospective-only and raises no P-level. It is a within-P3 residual diagnostic on a re-used cohort, not new predictive evidence. The document's §3 headline framing ("Kills (b): Markov-in-N is then falsified regardless of how well D_b scores") is stronger than §6 permits, and §6 should govern.

---

9. VERDICT

REJECT AS A MECHANISM DISCRIMINATOR. PARTIALLY SALVAGE.

Kill or hold:

• E1 primary (Δ_mem / Δ_exit / Π_mem / ICC) — DO NOT RUN as a verdict-bearing cell. It rests entirely on the burned mark channel (§1), its verdict-bearing Bonferroni MDE (1.011) exceeds its own registered effect (0.923) (§6.3), joint power is ~23% at the optimistic effect and <5% at half of it, and its veto control cannot separate (d) from (e) (§8.2) while (f) and (g) are unmodelled (§8.3). Running it consumes the last usable read of a burned channel to buy a near-certain NOT_ESTABLISHED.
• E1-S ladder — DO NOT RUN. The track itself declares it "expected to return NOT_ESTABLISHED for most or all rungs"; §6.1's published intervals (8/8 INCONCLUSIVE, half-widths 0.042–0.165 nats) confirm it; and D_a is not M2, so the contrasts would be uninterpretable even if powered (§2).

Salvage — one cell, after three repairs:

• E1-X (left-truncation repair) is the highest-value item in the document and I endorse its #1 ranking independently: the omitted renormalisation is 0.012–0.086 nats/event against an adverse gap of 0.037 nats (§5.2), no competitor carries a truncation term (grep "truncat" → 0 hits), and B4C02's three generators structurally cannot detect it. Required before running:
  1. freeze t_min = 0.300 s from the train partition (verified available: TRAIN_ONLY min = 0.300 s, 22 events at the floor, 0.02 s grid) — as written it is holdout-contaminated;
  2. specify the correction in normalised y-space with state-dependent y_min(N) = 0.300/scale_N(N), and state whether the mean-one constraint is imposed pre- or post-truncation — otherwise the refit is a different model family;
  3. replace the untestable "narrows by more than half" rule with a paired motor-cluster BCa contrast on (M2−M3)_truncated − (M2−M3)_untruncated under common random numbers, and name which estimand (motor-equal +0.025, or event-pooled +0.0369) is the target.
• Propagate E0.1 immediately to whichever track owns the target-stack blanket. The p(N', direction, jump | ·) triple-count is a real pre-implementation defect and is free to fix now (§5.1).

Mandatory before any E1 cell is unfrozen:

1. Relabel the entire protocol HOLDOUT_MARK_CHANNEL_BURNED_RETROSPECTIVE_ONLY; delete "genuinely unspent" from E0.5; record that §4's 150 on / 83 off and the per-motor arrival vector are themselves holdout mark reads made during design (§1.3).
2. Add a positive control with an injected known Δ_mem — the current 8-control battery is passed in full by an always-zero estimator (§7.1).
3. Add a within-motor non-stationarity control and add explanations (f) and (g) to the decision table (§7.2, §8.3).
4. Pre-register exclusion of the D6 nextStateN = -1 holdout rows with counts declared; they land directly in Δ_mem, Π_mem, and the |jump|=1 veto (§4.1).
5. Reconcile §2.2 (CMH event-weighted) with §4 (per-motor sd) and with B3's frozen MOTOR_EQUAL primary — pick one estimand (§3).
6. Drop the build_bootstrap_cohort import requirement (it resamples train, unfreezes scale_N, and is a no-op for the CMH contrasts); apply D1 discipline where it actually bites: ICC and D_c (§3.1).
7. Demote E1-C8 to E1-S scope — scale_N cancels exactly in within-(motor,state) contrasts, so it is invariant by construction (§5.3).
8. Fix: §4's design-effect slip (3.68, not 3.9); the two-sided MDE table under one-sided registered tests (0.815, not 0.918); §2.1's 233/19 assertion versus §2.3's censoring-inclusive primary; §7's "minutes" cost against C1's nested coverage requirement.

Standing constraints unchanged by anything in Track E: M2_LOGNORMAL remains an adversarial baseline (§6.7 correctly says E1-X does not promote it). One study, no transfer set, raw MAT archive absent — P4 and P7 unreachable. B4C02 and B4C10 have not returned; nothing here anticipates their outcomes. A design document is not evidence and raises no P-level.