TRACK F VERIFICATION — Parity-Evidence Map vs. Recorded Gate Verdicts

Verifier scope: read-only. No file created or edited. No holdout-partition statistic computed. One declared TRAIN-partition count executed (§4.1), disclosed inline. Target: .../scratchpad/tracks/TRACK_F.md (Track F — Biological Parity Evidence Requirements) Repository: [redacted: operator-path]/Documents/UNI-Flagellum/UNI-FLAGELLUM, branch hierarchical-aif/motor-stack, HEAD 17a2f0e — CHECKED_AGAINST_CODE (git log -1 → 17a2f0e audit(phase-b): B4 identifiability & robustness result + PROSPECTIVE flip), NO_DATA_ACCESS_NEEDED.

Headline: Track F's parity-evidence map is substantially accurate — every gate id, status, and headline numeric I checked reproduces from the recorded artifacts, to the digit. It is not guilty of the failure modes the contract targets: it does not redefine the ladder, does not claim a design document raises a level, and does not launder M2_LOGNORMAL. Four defects survive verification, one of which is a genuine numeric CONTRADICTION traceable to a repository documentation error, and one of which is a scientifically material omission: Track F does not know about D5 or D6, and its single proposed runnable route to P3 is dead because of D5.

1. Ground-truth confirmation (as ordered by the task)

1.1 Cross-study parity — CONFIRMED

CHECKED_AGAINST_RESULTS | NO_DATA_ACCESS_NEEDED Source: [redacted: operator-path]/Documents/UNI-Flagellum/UNI-FLAGELLUM/experiments/results/cross-study-parity-report.json

summary.overall = "PARTIAL_PARITY_ONLY", summary.fullBiologicalParityAchieved = false, statusCounts = {PASS: 8, FAIL: 3, NOT_ESTABLISHED: 2, BLOCKED_EXTERNAL: 3}. Full enumeration as recorded:

X16 evidence verbatim: allPriorRequiredGatesPass: false, nonPassGateIds = ["X06_FINITE_LATTICE_COOPERATIVITY","X10_CROSS_STUDY_PARAMETER_TRANSFER","X11_STRUCTURAL_CONSISTENCY","X12_ACTIVE_INFERENCE_CAUSAL_IDENTITY","X13_LIVE_SIGNAL_CHAIN","X14_INDEPENDENT_WET_LAB_REPLICATION","X15_PRINTED_MODEL_VALIDATION"].

One precision note on the task's own ground-truth statement. The task framed cross-study as "X06, X11 FAIL". The recorded statusCounts.FAIL is 3, not 2 — X16_FULL_BIOLOGICAL_PARITY is itself carried as a FAIL cell, not merely as a derived verdict. Track F states this correctly at F.9 (X16_FULL_BIOLOGICAL_PARITY = FAIL). No disagreement with Track F; noted so the count is not later misquoted as "two failures".

1.2 Science gates — CONFIRMED

CHECKED_AGAINST_RESULTS | NO_DATA_ACCESS_NEEDED Source: .../experiments/results/science-gates-report.json

summary.overall = "PARTIAL_PARITY_ONLY", fullBiologicalParityAchieved = false, proofClaim = "No universal, causal, or biological Active-Inference identity was proved.", gateCount = 14, statusCounts = {PASS: 4, FAIL: 3, SOURCE_ONLY: 1, BLOCKED_EXTERNAL: 5, NOT_ESTABLISHED: 1}.

The three FAILs are exactly G03_PUBLIC_ARTIFACT_PARITY, G05_SYNTHETIC_RECOVERY, G06_HELDOUT_MECHANISTIC_PREDICTION — as the task stated. Full enumeration: G00 PASS, G01_OBSERVATION_BOUNDARY PASS, G02 PASS, G03 FAIL, G04 PASS, G05 FAIL, G06 FAIL, G07_H_STATE_REEXTRACTION SOURCE_ONLY, G08 BLOCKED_EXTERNAL, G09_SWITCH_COOPERATIVITY BLOCKED_EXTERNAL, G10 NOT_ESTABLISHED, G11 BLOCKED_EXTERNAL, G12 BLOCKED_EXTERNAL, G13 BLOCKED_EXTERNAL.

Gap in Track F (not a contradiction, an omission): G07_H_STATE_REEXTRACTION = SOURCE_ONLY and G09_SWITCH_COOPERATIVITY = BLOCKED_EXTERNAL appear nowhere in Track F. SOURCE_ONLY is a non-PASS status; under the frozen verdict algorithm it counts against the conjunction exactly as FAIL does. Track F's P2 section claims to enumerate observational evidence and does not mention that the transient H-state reproduction is source-only. CHECKED_AGAINST_RESULTS | NO_DATA_ACCESS_NEEDED.

1.3 Observed experiment H1/H2/H4 — CONFIRMED, with a status-string correction

CHECKED_AGAINST_RESULTS | HOLDOUT_ALREADY_SPENT_DURATION_ONLY (H1/H2) / HOLDOUT_ALREADY_SPENT_DIRECTION (H4) Source: .../experiments/results/observed-experiment-report.json → claims

• H1_OVERDISPERSION = SUPPORTED_WITHIN_PROTOCOL, observed 3.149531339591441, interval95 [1.5141240937, 3.5675042611]. ✓
• H2_HELDOUT_LOG_SCORE = SUPPORTED_WITHIN_PROTOCOL, observed 0.210181454187411 nats/event, interval95 [0.0691846224, 0.3247183938]. ✓
• H4_DIRECTION = INCONCLUSIVE_POINT_ESTIMATE_ONLY (not the bare string INCONCLUSIVE), observed 0.008775508769140683, interval95 [-0.0283851190, 0.0484046190]. ✓ in substance; the exact recorded token is longer than both the task's phrasing and Track F's.
• Not mentioned by Track F at all: H3_SURVIVAL_POSTERIOR = MODEL_CONSEQUENCE_CONFIRMED, with the fence "This is exact inference in the declared model, illustrated with observed durations; it is not evidence that a bacterium represents this posterior." This is the single most abuse-prone claim in the repository (it is the one that superficially reads as "the motor infers"), and Track F's P5 section — the section whose whole job is to prevent that misreading — never cites it. CHECKED_AGAINST_RESULTS.

1.4 Ladder authority — CONFIRMED, contract satisfied

CHECKED_AGAINST_CODE | NO_DATA_ACCESS_NEEDED

audits/coordinator/claude-three-phase-parity-completion-plan.v1.md: ## 3. Operational parity ladder at line 61; the P0 table row at line 68; P8 row at line 76; `P8` is conjunctive at line 78. Verdict algorithm Pn = PASS only if P0..P(n-1) PASS and all required Pn cells PASS at line 545 (Track F cites 544-546 — correct as a range). Level names match Track F's usage verbatim; Track F does not redefine the ladder. Contract clause satisfied.

Repo CLAUDE.md:163-182 restates the ladder in prose (P0..P8, conjunctive, "the report names the first unsatisfied level"). Track F's citation is accurate.

2. Per-level map verification (F.0 table and F.1–F.9)

Every numeric below was re-read from the artifact. ✓ = reproduces to the digit.

Citation-hygiene note (low severity, CHECKED_AGAINST_CODE). Track F's line anchors into scripts/run-science-gates.py are off by one for G02 (id at 415, cited 414), G03 (423 / 422), G04 (431 / 430), G05 (446 / 445), G06 (454 / 453), G08 (477 / 476), G10 (506 / 505), G12 (525 / 524) — consistent with citing the gates.append(gate( opening line — but G00 is cited at its id line (389). Same class of drift into audits/phase-b/b3-model-competition-runner.py: the censoring filter if (not e["rightCensored"]) and e["stateN"] in self.states is at line 109, cited as :110. Docs anchors drift ±2 lines. All cross-study anchors (366, 392, 395, 403, 408, 412, 418-422, 425) are exact. lib/source-first-passage.js:61 is exact — if (event.rightCensored) return Math.log(Math.max(EPS, densities.survival));. Not material to any claim; flagged so a reader chasing an anchor does not conclude the artifact moved.

3. The five questions

Q1 — Which level is the FIRST unsatisfied one, and therefore the binding constraint on P8?

Track F's answer (P0) is CORRECT. CHECKED_AGAINST_RESULTS + CHECKED_AGAINST_CODE | NO_DATA_ACCESS_NEEDED.

Under the frozen algorithm at audits/coordinator/claude-three-phase-parity-completion-plan.v1.md:545, Pn = PASS only if P0..P(n-1) PASS and all required Pn cells PASS. P0's required cells are not all PASS: AC4 = FAIL verbatim in audits/phase-d/d1-correction-package.v1.json, with creditedDetections: 20 of 23 cells and D1A10_LATTICE_NEXT_NEAREST_NEIGHBOUR_RING recorded SURVIVED against target scripts/run-cross-study-parity.py. Therefore P0 is the first unsatisfied level and is today's binding constraint on P8. Track F's two qualifications — that P0 is the cheapest constraint (closable without a laboratory) and that closing it does not bring P1–P7 near — are both sound.

One correction to Track F's own labelling. Track F's F.0 table assigns P0 the status NOT_ESTABLISHED. The governing cell is recorded as FAIL ("AC4": "FAIL"). Under the plan's status vocabulary, FAIL is the stronger, adverse status; NOT_ESTABLISHED means "the evidence does not decide". P0 is not undecided — an acceptance criterion was evaluated and failed. The correct P0 ladder label is FAIL, and Track F softens it. This is the one place in the document where the error runs in the self-favouring direction. CHECKED_AGAINST_RESULTS | NO_DATA_ACCESS_NEEDED.

Secondary methodological note: the plan's status vocabulary (PASS / FAIL / CONTRADICTED / NOT_ESTABLISHED / BLOCKED_EXTERNAL / NOT_RUN) governs cells and parity dimensions; the ladder itself is binary-with-ordering (Pn = PASS or not, plus "name the first unsatisfied level"). Track F's per-level statuses (P1=FAIL, P3=NOT_ESTABLISHED, P6=FAIL…) are legitimate cell-aggregation summaries but are not ladder verdicts, and a reader could mistake "P3 = NOT_ESTABLISHED" for a weaker statement than "P3 is not PASS and is unreachable while P0 is open". Track F does state the ordering explicitly, so this is presentation, not redefinition — the contract clause is not breached.

Q2 — Which levels can NEVER be closed by software alone?

Track F answers this correctly and plainly at F.10 §2. Confirmed and restated here without hedge.

• P5 (interventional) — cannot be closed by software. Ever. CHECKED_AGAINST_RESULTS | PROSPECTIVE_NEW_DATA_ONLY. X13_LIVE_SIGNAL_CHAIN.evidence.softwareCannotSubstitute = true; X12.evidence.discriminatingInterventions = 0; G10.limitation verbatim: "Exact Bayesian inference inside a declared software model is not evidence that a bacterium implements that representation." No model, no model class, no hierarchical AIF stack, no additional bootstrap replicates, no longer compute run, and no additional agent moves P5. It requires physical manipulation of living motors with the prediction committed before the outcome. Track F's compounding argument — that Lmotor-0 = {dt, N', direction, jump, censor} contains no motor-selected action, so q(π) = softmax(ln E(π) − γ G(π)) has no separable observable consequence and any G_motor term merely reparameterizes the same marked point-process density — is a correct in-principle identifiability argument and I endorse it. It means P5 is blocked twice: no wet lab, and no action variable.
• P6 (structural/mechanistic) — cannot be closed by software. Ever. CHECKED_AGAINST_RESULTS | INDEPENDENT_TRANSFER_REQUIRED / PROSPECTIVE_NEW_DATA_ONLY. Requires an independent molecular measurement (simultaneous reporter + rotation/occupancy trace). CLAUDE.md:215-218 is binding: predictive superiority is never promoted to mechanism. A better fit, of any margin, on any dataset, cannot move P6.
• P7 (independent replication) — cannot be closed by software. Ever. CHECKED_AGAINST_RESULTS | INDEPENDENT_TRANSFER_REQUIRED. G12.evidence.independentLaboratories = 0; G12.limitation verbatim: "Repository replay and an independent numerical implementation are not independent biological replication." Track F's explicit non-substitute list (independent JS/Python oracles, Codex verdicts, a second AI agent, a second runtime, a clean clone, a second bootstrap seed, another machine) is correct and should be carried verbatim into any deliverable.
• P4 (transfer) — not closable by modelling in this repository, but blocked by acquisition rather than by physics. X10.evidence.commensurateRawTransferTests = 0. Track F's distinction (acquisition-blocked, not permanently impossible) is fair.
• G13/X15 (printed mechanism) additionally require fabrication, dimensional metrology, sensor calibration, and safety review.

Q3 — Does the track anywhere imply a P-level moves UP as a result of design work?

NO. CHECKED_AGAINST_CODE (full read of TRACK_F.md) | NO_DATA_ACCESS_NEEDED. No CONTRADICTED finding on this axis.

The document self-declares at line 3: "DESIGN / EVIDENCE-MAP ONLY. It raises no P-level. It creates no receipt. Every status below is copied from an existing frozen artifact, not asserted." F.10 §2 closes with the required corollary: "building the hierarchical motor-stack AIF model raises no P-level by itself." Every Next experiment block is framed as an obligation or an acquisition task, never as an accomplished raise. F.6 goes further and states the P5 blocker is "a limitation in principle, not a compute limitation" — the opposite of design-as-evidence.

The closest thing to an exposure is F.10 §2's clause (b), "contribute to P2 by scoring the currently unscored mark channel". That is a claim about executing a scoring run, not about design — so it does not breach this clause. It does, however, breach the D5 firewall, which is §3 Q5 and §4.2 below.

Q4 — The strongest claim the current evidence actually licenses

Track F's F.10 §3 sentence is close to correct and materially better than anything else in the repository, but it has two defects: it asserts the mark channel is unscored (false at repository scope — §4.3), and it omits the P0 caveat, so it presents an evidence claim resting on artifacts whose raw→events ingestion has never been re-executed in this tree. My corrected sentence:

On the frozen sha256_mod5(motorId) motor split of the single-study, single-laboratory Wadhwa et al. 2022 E. coli post-electrorotation stator-remodeling event artifact (SHA-256 32ec7ebf45bf7776ebef239351d2304cc3cead6d1be7749a2478e44a323f25db, whose raw→events derivation is asserted by recorded hash and has never been re-executed in this tree because data/remodeling_data.mat is absent, and whose semantic-mutation coverage is credited at 20 of 23 cells with AC4 = FAIL and one surviving undetected lattice-adjacency corruption), restricted to 233 uncensored held-out dwell durations in states N=1..8 from 19 held-out motors fitted on 793 training events from 80 disjoint training motors, a training-only two-timescale mixture assigns higher held-out log predictive density than a homogeneous memoryless model by 0.210 nats/event with a motor-clustered 95% interval of [0.069, 0.325], and the mean held-out CV² is 3.150 with motor-clustered 95% interval [1.514, 3.568] lying entirely above the memoryless value of one — while a one-parameter lognormal baseline out-predicts that same mixture point-wise on the same holdout (recorded motor-clustered M3 − M2 interval [-0.0680, 0.0148], point -0.0369 nats/event) — so the evidence licenses only the statement that E. coli post-electrorotation stator-remodeling dwell timing is inconsistent with a homogeneous memoryless process at this observation resolution, and licenses nothing whatever about mechanism, latent-state identity, cross-laboratory transfer, intervention response, active inference, or biological parity.

CHECKED_AGAINST_RESULTS | HOLDOUT_ALREADY_SPENT_DURATION_ONLY.

Supporting receipt Track F missed, which strengthens its own adverse position: experiments/results/observed-experiment-report.json → pairedMixtureAdvantageInterval95.mixtureVsLognormal = {lower: -0.0680322535, upper: 0.0148179542} with pairedMixtureAdvantageNatsPerEvent.mixtureVsLognormal = -0.0368663986. Track F's P3 "Exact missing receipt" asks for "a motor-clustered 95% interval for [candidate model − M2_LOGNORMAL]" as if none existed. One already exists for M3, it is recorded, it contains zero, and its point estimate is negative. A second recorded adverse detail Track F omits: in the same report's calibration block the lognormal has the best KS statistic (0.0622) of all four models, better than the mixture (0.0775). The lognormal is not merely tying on log score — it is better calibrated. Both facts belong in the retained adverse record.

Q5 — Does the track incorporate D5 (mark→P4/P7) and D6 (limits P2 for mark fields)?

NO to both. This is the most consequential finding in this verification. CHECKED_AGAINST_CODE (full text read of TRACK_F.md) | NO_DATA_ACCESS_NEEDED.

The strings D5 and D6 appear nowhere in Track F. Neither does any equivalent concept: no mention that the held-out mark channel was read by a prior track, no mention that mark-process claims on Wadhwa-2022 can no longer be prospective, and no mention that ingest writes an out-of-range nextStateN for holdout events. Consequences in §4.2 and §4.4.

4. Findings, ranked by severity

4.1 CONTRADICTED — Track F's G04 censoring counts are wrong, and so is the repository document it copied

Track F F.3 states: "G04_CENSORED_JOINT_LIKELIHOOD = PASS … 76 training + 16 holdout censored intervals contribute log S(t)."

What is actually true. CHECKED_AGAINST_RESULTS — experiments/results/science-gates-report.json → G04_CENSORED_JOINT_LIKELIHOOD.evidence:

trainIntervals: 836   trainRightCensored: 43
holdoutIntervals: 244 holdoutRightCensored: 11

Root cause — not Track F's invention. CHECKED_AGAINST_CODE: docs/SCIENCE-GATES.md:58 reads "| G04 censored joint likelihood | PASS | 76 training and 16 held-out censored intervals in eligible states contribute survival likelihoods. |". Track F faithfully copied a repository document that disagrees with its own machine-readable artifact. This is a live resonance defect under the operating contract clause 7 (report vs artifact must agree).

Independent adjudication.

• Train leg — CHECKED_AGAINST_CODE, split boundary TRAIN_ONLY, computation declared: I counted rows of experiments/data/wadhwa-2022-events.json with partition == "train" only. Result: 1048 train rows; rightCensored true across all states = 89; rightCensored true and 1 ≤ stateN ≤ 8 = 43; uncensored and 1 ≤ stateN ≤ 8 = 793. The artifact's trainRightCensored: 43 is correct and reconciles exactly with B3's train: 793 (43 + 793 = 836 = trainIntervals). The doc's "76" matches nothing — not the eligible-state count (43), not the all-state count (89).
• Holdout leg — CHECKED_AGAINST_RESULTS by arithmetic on already-published numbers only, no holdout row read: holdoutIntervals 244 − B3 holdout: 233 = 11, matching holdoutRightCensored: 11. The doc's "16" matches nothing. I did not count holdout rows; that would have been a new statistic over the holdout partition. Split boundary: HOLDOUT_ALREADY_SPENT_DURATION_ONLY (both operands are published figures).

Impact. Low on any scientific conclusion — G04 PASSes either way and no downstream number depends on the count. High on evidence hygiene: it is a documentation-vs-artifact contradiction sitting in the summary table of the primary gate document, and Track F propagated it without checking. Any deliverable quoting "76/16" is quoting a defect.

Minimal correction (for the owner, not applied by me — read-only run): amend docs/SCIENCE-GATES.md:58 to 43 training and 11 held-out, and add a failing-first test asserting the doc table row equals science-gates-report.json → G04.evidence.{trainRightCensored, holdoutRightCensored}. That class of test does not currently exist, which is why the drift survived.

4.2 CONTRADICTED (in consequence) — Track F's only runnable route to P3 is dead under D5

Track F F.4 states: "Two honest routes only: 1. More experimental units — external. 2. More information per unit — score the currently unscored mark channel and the currently excluded censored events. This adds likelihood terms without adding motors and is the only route runnable in this repository. It must be pre-registered…" — echoed at F.3's missing-receipt (b), F.10 §2 clause (b), and F.11 item 5 ("the single largest piece of runnable unexploited information in the repository — and the only route to P3 that does not require new motors").

Why this is contradicted. P3 requires held-out predictive parity — a frozen model scored on untouched motors, with the prediction committed before the scoring run. Per the established facts governing this verification, a prior UltraCode track read the held-out mark channel while answering a request for empirical direction/jump marginals per state. That read is irreversible. No mark-process claim on Wadhwa-2022 can ever again be labelled PROSPECTIVE. A mark-channel likelihood scored on this holdout is therefore HOLDOUT_MARK_CHANNEL_BURNED_RETROSPECTIVE_ONLY and cannot carry P3. Track F instructs the reader to pre-register a contrast that can no longer be pre-registered.

Independently corroborated without invoking D5. CHECKED_AGAINST_RESULTS | HOLDOUT_ALREADY_SPENT_DIRECTION: the direction mark was already spent on the holdout before any UltraCode track existed — observed-experiment-report.json → claims[H4_DIRECTION] records a held-out state-conditioned direction log-loss (0.6464397923 vs global 0.6552153011), Brier scores, and a motor-clustered interval. docs/OBSERVED-EXPERIMENT.md:139-140 narrates it. So even on a strict reading, the direction sub-channel of the mark process was consumed by H4 and is retrospective-only regardless of D5.

What survives. The censoring/log S(t) leg of Track F's proposal is not mark-process evidence and is not burned — right-censored held-out intervals are already scored by G04/G06 under the correct treatment at lib/source-first-passage.js:61, and restoring them inside the B3 competition is a legitimate, runnable P2/P3 contribution. Track F conflated two proposals with very different evidential standing into one bullet. The correct rewrite of F.4 route 2 is: censoring-inclusive rescore = runnable and prospective-eligible; mark-channel scoring = runnable but RETROSPECTIVE_ONLY on this dataset, hence INDEPENDENT_TRANSFER_REQUIRED / PROSPECTIVE_NEW_DATA_ONLY for any P3 or P4/P7 credit.

Corollary Track F does not state and must: the mark-process mechanism question has been pushed off this dataset entirely. It is now a P4 (independent transfer dataset) or P7 (independent laboratory) question. That relocation strengthens Track F's overall thesis — the last runnable-looking lever was smaller than advertised — but it must be stated, or a reader will spend effort on a route that cannot pay.

4.3 CONTRADICTED — "the mark process of the observation is entirely unscored" is false at repository scope

Track F F.3 states: *"the B3 competition reads none of the mark fields … The mark process of the observation is entirely unscored."* Repeated in F.3's Allowed wording and embedded in the F.10 §3 strongest-claim sentence as "the transition-mark channel unscored".

First clause: TRUE. CHECKED_AGAINST_CODE — audits/phase-b/b3-model-competition-runner.py:108-109:

elig = [e for e in events
        if (not e["rightCensored"]) and e["stateN"] in self.states]

B3 filters on rightCensored and stateN only; nextStateN, direction, jump are never read by the nine-model competition. ✓

Second clause: FALSE. The direction mark is scored, on the holdout, in two separate places:

1. CHECKED_AGAINST_CODE — lib/source-first-passage.js:59-65, sourceLogLikelihood: censored → log S(t); direction === "on" → log P₊; direction === "off" → log P₋; and it throws on an uncensored event with no direction. This is the likelihood used by G04 and G06. docs/SCIENCE-GATES.md:117-118 says so in prose: "Every eligible held-out interval, including its censoring indicator and observed exit direction, was then scored."
2. CHECKED_AGAINST_RESULTS — H4_DIRECTION scores state-conditioned direction on the holdout (§4.2).

What is genuinely unscored is narrower and must be stated that way: nextStateN and jump are unscored anywhere; direction is unscored within the B3 nine-model competition but is scored by G04, G06, and H4. As written, Track F's sentence would let a reader believe the direction channel is fresh. It is not — it is spent. Split boundary: HOLDOUT_ALREADY_SPENT_DIRECTION.

4.4 Omission with a code-verified mechanism — D6 limits P2/P3 for mark fields, and the contamination reaches G04/G06

Track F never mentions D6. Its P2 section lists nextStateN, direction, jump, rightCensored as clean available channels.

Mechanism confirmed in code. CHECKED_AGAINST_CODE | NO_DATA_ACCESS_NEEDED — scripts/ingest-wadhwa-data.py:141-160:

if dwell["state"] < 0 or dwell["state"] > 11:      # entered state IS range-checked
    exclusions["outOfRangeDwells"] += 1
    continue
...
next_state = dwell["nextState"]                    # NOT range-checked
...
"nextStateN": next_state,
"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 asymmetry the established fact describes is exactly present: state is range-checked, next_state is not, and the row is emitted regardless.

A consequence Track F could not have drawn but which follows from the code alone, and which I have not seen stated anywhere: because direction and jump are derived from next_state, a corrupt next_state does not merely corrupt nextStateN — it silently manufactures a plausible-looking direction. A next_state of -1 is less than every valid state, so the row is emitted with direction: "off" and a large negative jump, and nothing downstream can distinguish it from a genuine off-transition. sourceLogLikelihood will not throw (it throws only on direction === null); it will score the event as an off-exit via log P₋.

Therefore, given the established fact that 2 holdout events carry nextStateN = -1: those 2 events are inside G04's 233 uncensored held-out intervals and inside G06's held-out score, contributing a fabricated log P₋ term. G06 already FAILs, so this cannot flip a pass to a fail; but it means G06's recorded advantage +0.05826 and interval [-0.01580, 0.13264] are computed over a holdout set containing 2 rows with synthesized direction marks. That is a P2 provenance limitation on every direction-dependent held-out number in the repository, and it is invisible in the current ledger.

Quantifying the effect is NOT_CHECKED — it would require re-scoring holdout rows, i.e. holdout access. Requires a prospective record first. The correct disposition is: record the limitation, add a failing-first range check on next_state in scripts/ingest-wadhwa-data.py mirroring the state check, re-ingest when data/remodeling_data.mat is obtained, and mark the G04/G06 direction-dependent figures RETROSPECTIVE, MARK-CONTAMINATED (n=2) until then. Track F's P2 section must carry this; at present it does not.

4.5 Overstatement — "the lattice topology is not equation-locked by any control" is too strong

Track F F.1 item 2 / F.2 / F.11 item 1 present D1A10 as showing the lattice interaction topology "is detected by nothing in the suite" and "not equation-locked by any control".

What the artifact actually records. CHECKED_AGAINST_RESULTS — audits/phase-d/d1-semantic-remediation-result.v1.json → outcomes:

• D1A10_LATTICE_NEXT_NEAREST_NEIGHBOUR_RING: propertyId: D1P10_PERIODIC_LATTICE_TOPOLOGY, targetPath: scripts/run-cross-study-parity.py, expectedFailingTest: "periodic thirteen-site lattice retains its wraparound bond", predictedClassification: "SURVIVED" — it was predicted to survive and did; the gate run exited 0.
• D1A10B_LATTICE_SINGLE_BOND_DOUBLE_COUNTED: same property, same target, predictedClassification: "DETECTED_SEMANTIC", and its d1-semantic gate command exited 1 — i.e. the wraparound-bond test does fire on a different corruption of the same topology.

So the property is partially covered: one control exists and detects one topology corruption; a next-nearest-neighbour rewrite escapes it. Track F never mentions D1A10B. The honest statement is "the periodic-lattice topology property has one control, which detects a double-counted-bond corruption but not a next-nearest-neighbour rewrite; the adjacency set itself is unpinned." Track F's stronger phrasing overstates the hole.

Counterweight that partially rescues Track F. CHECKED_AGAINST_RESULTS — d1-correction-package.v1.json records of the addendum: "Its revised D1A10 prediction and its added D1A10B cell were authored with knowledge of the implemented gates AND entered history in the very commit used as the replay base. They carry materially lower evidential weight than the ten exact replays", with ruleForReaders: "D1A10B may not be presented as a pre-implementation frozen prediction, and may not be presented as having strict prospective ancestry." So the one control that does fire is itself post-hoc. Track F's conclusion (the topology is under-locked, and this weakens X06's verification standing) stands; its supporting sentence is too absolute and should be replaced with the two-part statement above.

4.6 Omissions of record (each CHECKED_AGAINST_RESULTS, NO_DATA_ACCESS_NEEDED)

5. NOT_CHECKED register — verifications deliberately not performed

Each of the following would have required computing a new statistic over experiments/data/wadhwa-2022-events.json rows with partition == "holdout". Per the firewall, not done, and this is the correct outcome, not a gap in effort.

1. Direct count of holdout right-censored intervals (the "16" in §4.1). NOT_CHECKED — would require holdout access; requires prospective record first. Adjudicated instead by arithmetic on two already-published figures (244 − 233 = 11). Split boundary: HOLDOUT_ALREADY_SPENT_DURATION_ONLY.
2. Count/identity of the D6-affected holdout rows (nextStateN = -1). NOT_CHECKED — would require holdout access; requires prospective record first. Mechanism confirmed from scripts/ingest-wadhwa-data.py alone (§4.4); the count is taken as given from the established-facts block, REPORTED_BY_TRACK.
3. Magnitude of the D6 contamination on G04/G06 held-out log scores. NOT_CHECKED — would require holdout re-scoring; requires prospective record first. Split boundary: HOLDOUT_ALREADY_SPENT_DURATION_ONLY for the existing figures; any new estimate is PROSPECTIVE_NEW_DATA_ONLY.
4. Any empirical marginal, table, or correlation over holdout direction / jump / nextStateN. NOT_CHECKED and will remain so. Split boundary: HOLDOUT_MARK_CHANNEL_BURNED_RETROSPECTIVE_ONLY.
5. Whether D1A10's escape actually changes the X06 J estimate. NOT_CHECKED — would require executing the mutated scripts/run-cross-study-parity.py, i.e. a write/run in a frozen tree. Split boundary: NO_DATA_ACCESS_NEEDED but out of scope for a read-only verification.
6. Outcomes of the in-progress corrected runs B4C02 (misspecified-world discriminator) and B4C10 (M4 bootstrap). NOT_CHECKED — not returned. No claim in this document assumes either outcome. Track F likewise makes no assumption about them, correctly.
7. X15 keys physicalPrintRuns / measuredBacklashRuns. NOT_CHECKED — not present as keys in cross-study-parity-report.json; substance (softwareCannotSubstitute: true, missing fabrication/inspection/calibration) confirmed instead.

6. Verdict on Track F

ACCEPT WITH REQUIRED CORRECTIONS. The parity-evidence map is faithful: every gate id, status, and headline numeric that I re-read reproduces from the frozen artifacts, generally to ten significant figures. The forbidden-language contract is honoured throughout — no "parity achieved", no "active inference demonstrated", no promotion of M2_LOGNORMAL to model status, no design-raises-a-level claim, and the F/G-absence statement is scoped precisely to the science pipeline where a looser claim would have been false. The P5 in-principle identifiability argument (no action channel ⇒ G_motor observationally equivalent to a reparameterized marked point process) is the document's strongest original contribution and survives scrutiny.

Required before Track F is used as a basis for any deliverable:

1. Fix the G04 counts to 43 training / 11 held-out right-censored, and raise the underlying docs/SCIENCE-GATES.md:58 doc-vs-artifact contradiction as a separate delta with a failing-first test. (§4.1)
2. Withdraw "score the mark channel" as a P3 route. Split F.4 route 2 into (i) censoring-inclusive rescore — runnable, prospective-eligible; (ii) mark-channel likelihood — HOLDOUT_MARK_CHANNEL_BURNED_RETROSPECTIVE_ONLY, credit relocated to P4/P7. (§4.2)
3. Replace "the mark process is entirely unscored" with the accurate scoping: nextStateN/jump unscored anywhere; direction unscored within B3 but already spent on the holdout by G04, G06, and H4. (§4.3)
4. Add D6 to P2 with its code-verified mechanism and the previously unstated consequence that a corrupt next_state manufactures a plausible direction that flows undetected into G04/G06 holdout scoring. (§4.4)
5. Soften the D1A10 claim to acknowledge D1A10B and the one (post-hoc, lower-weight) wraparound-bond control that does fire. (§4.5)
6. Relabel P0 FAIL, not NOT_ESTABLISHED — the governing cell records "AC4": "FAIL". (§3 Q1)
7. Add G07 (SOURCE_ONLY) and G09 (BLOCKED_EXTERNAL) to the map, and cite H3's fence in P5. (§4.6 a, b)

None of these corrections moves any P-level. P0 remains the first unsatisfied level and the binding constraint on P8; P5, P6, and P7 remain permanently closed to software, and no model, no additional compute, and no additional agent will ever change that.