Universal Natural Intelligence

This is the document. Rendered from the repository at the commit above, with nothing rewritten for the web. A gate re-renders it on every deploy and fails the build if a single byte differs.

Verifier scope. Read-only. Nothing created or edited. audits/phase-c/** and audits/phase-d/** never opened. No statistic, table, marginal, fit, or plot computed over any row of experiments/data/wadhwa-2022-events.json with partition == "holdout". Every train-partition computation below was performed after a hard sha256(motorId) % 5 != 0 filter applied before any other operation, and is labelled TRAIN_ONLY.

Headline. Track D's central verdict — the full motor-stack AIF model is not identifiable on current Wadhwa data — survives independent check and is if anything understated. But six of its supporting receipts are wrong, including two that its own key argument is built on. The document also contains one internal self-contradiction about which stages retain prospectivity, and its proposed module list contains three concrete contract hazards, one of which would silently move every NLPD by up to 0.45 nats while appearing to be a model improvement.

0. Receipt audit — what checked out, what did not

0.1 CONFIRMED

Track claim	My verification	Label	Split boundary
F1 split `sha256(motorId)%5==0 → holdout`; `Cohort.__init__` HALTs on mismatch	`b3-model-competition-runner.py:89` (`sha256_mod5`), `:105-108` (halt `FAILED-SPLIT-MISMATCH`)	CHECKED_AGAINST_CODE	NO_DATA_ACCESS_NEEDED
F2 eligibility excludes every right-censored event	`:109-110` — `elig = [e for e in events if (not e["rightCensored"]) and e["stateN"] in self.states]`	CHECKED_AGAINST_CODE	NO_DATA_ACCESS_NEEDED
F3 B3 never reads `nextStateN`/`direction`/`jump`	Exhaustive field extraction over the runner returns exactly: `motorId`(11), `stateN`(14), `durationS`(7), `partition`(1), `rightCensored`(1) + 3 telemetry keys. Zero occurrences of `nextStateN`, `direction`, `jump` in all 1645 lines	CHECKED_AGAINST_CODE	NO_DATA_ACCESS_NEEDED
F4 M6 is 8 independent per-state Weibull shapes, no kernel	`fit_m6` at `:445-465` loops `for s in cohort.states`, fits a 1-D box `(0.05,5.0)` per state, `m6_logpdf_perstate = m1_logpdf`. No transition term exists	CHECKED_AGAINST_CODE	NO_DATA_ACCESS_NEEDED
F6 109 motors with events → 109 censored events, all terminal	`ingestion.exclusions` = `{leftTruncated:129, rightCensored:109, outOfRange:3, zeroOrNeg:0}`; `motorCount:129`, `eventCount:1349`, `uncensoredEventCount:1240` ⇒ 109 censored events, one per each of the 109 motors with ≥2 runs. Ingest appends the terminal censored dwell unconditionally (`ingest-wadhwa-data.py:64-76`) and does not `continue` past it	CHECKED_AGAINST_CODE	NO_DATA_ACCESS_NEEDED
F8 cohort counts 793/80 train, 233/19 holdout; 818/80 and 251/19	`b3-model-competition-result.json → cohorts.*.summary` reproduces all eight integers exactly	CHECKED_AGAINST_RESULTS	HOLDOUT_ALREADY_SPENT_DURATION_ONLY (published)
F9 train events/motor: min 1, median 7, mean 9.9, max 70; 22 < 5; 55 < 10	Recomputed: min 1, median 7.0, mean 9.9125, max 70; 22 motors < 5; 55 < 10. Exact	CHECKED_AGAINST_CODE	TRAIN_ONLY
F10 leaderboard span 0.1387 nats (M0 3.5480 → M2 3.4093); M3 at 3.4343; widest contrast 0.3289; all 8 motor-equal contrasts INCONCLUSIVE	All exact. Span = 3.54798328 − 3.40931416 = 0.13866913. M0 width 0.32889593. 8/8 INCONCLUSIVE	CHECKED_AGAINST_RESULTS	HOLDOUT_ALREADY_SPENT_DURATION_ONLY
F11 `nominalElectrorotationSpeed` 7 levels, per-motor constant, no timing; B4C06 3400 = `BLOCKED_EXTERNAL`	Train motors carry exactly 7 levels `{50,100,150,200,250,272,300}`; field lives on `motors[]` not `events[]`, so it has no timing. `B4C06.analysisStartIndex.3400.status == "BLOCKED_EXTERNAL"`, raw `c14de12c…` absent	CHECKED_AGAINST_CODE + CHECKED_AGAINST_RESULTS	TRAIN_ONLY
§1.10 calibration target M1 = `3.4333068483897238`	Exact match in `leaderboards.NLPD_motor_equal.motorEqual`	CHECKED_AGAINST_RESULTS	HOLDOUT_ALREADY_SPENT_DURATION_ONLY
§1.11 target M3-vs-M2 BCa `[−0.04369843…, +0.08676352…]`, width `0.12847493…`, INCONCLUSIVE	Exact match in `contrasts.NLPD_motor_equal.M2_LOGNORMAL`	CHECKED_AGAINST_RESULTS	HOLDOUT_ALREADY_SPENT_DURATION_ONLY
§1.12 adverse M2−M3 gap `0.0369` event-pooled	`adverseLognormalRetention.derived_eligible_1_to_8.lognormalMinusMixtureLogDensity = 0.03686651825627818`	CHECKED_AGAINST_RESULTS	HOLDOUT_ALREADY_SPENT_DURATION_ONLY
§3.1 quote "does NOT establish … that the organism performs the inference"	Verbatim in `notEstablished[0]`	CHECKED_AGAINST_RESULTS	NO_DATA_ACCESS_NEEDED
Ladder map citations (P5 `NOT_ESTABLISHED` line 77, P6 weakened line 78, P8 `FULL_PARITY=false` first unsatisfied `P4` line 80, six-condition rule 59-66)	All four verbatim	CHECKED_AGAINST_CODE	NO_DATA_ACCESS_NEEDED
S0's censoring motivation: 26.1 % of normalized exposure; Weibull shape 0.6251 → 0.5852; exponential rate 1.000 → 0.739; uncensored median 2.32 s	Recomputed independently: censored share of normalized exposure = 0.2607; mean-one Weibull shape 0.624971 → 0.585122 (Δ = 0.0398, grid 2e-4); mean-one exponential rate 1.000000 → 0.739312; uncensored median 2.3200 s. All confirmed	CHECKED_AGAINST_CODE	TRAIN_ONLY
S1 support: train `on`: 544, `off`: 249	Exact	CHECKED_AGAINST_CODE	TRAIN_ONLY
S2: "training jump support stops at +4"	Train eligible jump support = `{−4:2, −3:6, −2:35, −1:206, +1:395, +2:128, +3:16, +4:5}`. Max = +4. Confirmed	CHECKED_AGAINST_CODE	TRAIN_ONLY
Established fact: F, G, policies, E, Π, up/down messages absent from science pipeline	Independent grep for `expected_free_energy`, `variational_free_energy`, `policy_posterior`, `free_energy` across `audits/phase-b/.py`, `scripts/.py`, `lib/.js`: zero hits*	CHECKED_AGAINST_CODE	NO_DATA_ACCESS_NEEDED
Cited module line numbers (`_bridge.py:16` `dont_write_bytecode`, `:57-65` `frozen_cohort`; `seeding.py:30-44` `stable_seed`; `status.py:49-59` `verdict_from_ci`; `bootstrap.py:73-96` `build_bootstrap_cohort`)	All correct within ±1 line	CHECKED_AGAINST_CODE	NO_DATA_ACCESS_NEEDED

0.2 CONTRADICTED — six findings

C1 — the narrowest B3 contrast is 0.0835 nats, not 0.1285. The resolution floor the whole document rests on is wrong by 54 %. CONTRADICTED / HOLDOUT_ALREADY_SPENT_DURATION_ONLY Track F10 and §2.2, §2.3-reason-2, §5.5 and compare.minimum_detectable_effect (§1.11 test 5) all assert the narrowest 95 % BCa motor-equal contrast width is 0.1285 nats (M2), and derive a resolution floor of ≈ 0.064 nats. Full sorted widths from b3-model-competition-result.json → cohorts.derived_eligible_1_to_8.contrasts.NLPD_motor_equal:

M4_MIXTURE_K3                    0.083461   <-- narrowest
M8_EMPIRICAL_KDE                 0.086377
M2_LOGNORMAL                     0.128475   <-- track called this the narrowest
M1_WEIBULL                       0.139926
M6_SEMI_MARKOV_STATE_DEPENDENT   0.155219
M7_HIERARCHICAL_MOTOR            0.157996
M5_GAMMA                         0.180140
M0_EXPONENTIAL                   0.328896

M4 and M8 are both narrower than M2. On the track's own half-width heuristic the floor is 0.0417 nats, not 0.064. Consequence: §2.3's "factor of 2.6 below the noise floor" becomes a factor of 1.67. The verdict direction survives (0.025 < 0.0417), but the margin is 36 % of what was claimed, and §1.11 test 5 would hard-code a wrong number into a proving test.

Separately, the half-width heuristic itself is not a minimum detectable effect. A 95 % CI half-width is the MDE at ~50 % power; the conventional 80 %-power MDE is 2.80·SE ≈ 1.43 × half-width. minimum_detectable_effect(per_motor_ref, per_motor_chal, n_rep) as specified takes no power argument, so it cannot compute an MDE at all — it can only return a half-width under a different name.

C2 — §2.1's hidden-state kernel row contradicts its own formula; all three totals are understated. CONTRADICTED / NO_DATA_ACCESS_NEEDED The table gives the kernel q(z′|z,N) the formula S·K·(K−1) and the value 8 at S=8, K=2. 8 · 2 · 1 = 16. A K-state kernel conditioned on (z, N) has K−1 free entries per (z, N) pair, i.e. 16. The three totals (72 / 104 / 184) are arithmetically consistent with the wrong 8 and are each low by 8.

C3 — §2.1 omits the 8 per-state normalization scales, which are estimated from the training data. (this is the largest single count error) CONTRADICTED / TRAIN_ONLY b3-model-competition-runner.py:121-130 estimates scale_N[s] = mean(durationS) over train uncensored events per state, then normalizes e["_y"] = durationS / scale_N[stateN]. That is 8 free parameters fitted from the training split, and they carry the entire location structure — the proof is B4C05 → treatments.a_frozen_exclusion.fitted.M0_EXPONENTIAL.trainNLL == 793 exactly, i.e. the mean-one exponential's training NLL equals the event count identically, because scale_N has already absorbed the per-state mean. M0 is reported with params: [] but is really an 8-parameter model.

Consequence: §1.3's proving test "degenerate spec n_hidden=1, mark="none", motor_effect_dim=0 reproduces M1's single free parameter, proving the counter is calibrated against a known model" calibrates the counter against a wrong reference. M1 has 1 + 8 = 9 fitted parameters. A counter that passes that test is miscalibrated by construction, and every downstream count inherits the error.

Corrected totals (C2 + C3 together), same instantiation S=8, K=2, |Π|=4, d=1:

	track	corrected
direction-mark stack	72	88
jump-mark (geometric)	104	120
jump-mark (saturated)	184	200

C4 — S4's data-support figures are computed on a different cohort than S4 is defined on, overstating support by ~33 %. CONTRADICTED / TRAIN_ONLY §4-S4 claims "61 distinct (state, next) transitions observed in training, 35 with ≥ 5 observations, 28 with ≥ 10." On the frozen derived_eligible_1_to_8 cohort (train, uncensored, states 1–8) the counts are 46 distinct, 27 with ≥ 5, 23 with ≥ 10. The track's 61/35/28 reproduce exactly when the state restriction is dropped (train, all states 0–11). S4 is explicitly scoped to the 1..8 cohort, so the figures motivating it belong to a cohort S4 does not use. Not fabricated — mis-scoped, and it inflates the apparent feasibility of the kernel.

C5 — "19 holdout motors spread over 7 load levels (6/6/3/2/2/2/2)" sums to 23, not 19. CONTRADICTED / NO_DATA_ACCESS_NEEDED §3.1 row (ii) and §5.4. 6+6+3+2+2+2+2 = 23. experiments/data/wadhwa-2022-events.json → ingestion.holdoutMotorCount == 23; the B3 cohort has 19 because 4 holdout motors contribute no eligible uncensored 1–8 event. The distribution and the motor count are from two different populations, presented as one. The qualitative conclusion (≤ 3 motors per level at the thin end) is unaffected and is if anything worse at n=19.

C6 — the S0 median-censored-duration figure does not reproduce. CONTRADICTED / TRAIN_ONLY §4-S0 and §5.2 claim "censored dwells have median 83.4 s" and "median censored duration 36× the median uncensored". Recomputed on train: median censored duration over the eligible 1..8 states = 85.46 s (n=43); over all train censored events = 94.34 s (n=89). Neither is 83.4. The 36× ratio survives (85.46 / 2.32 = 36.8×). I did not test whether 83.4 is the median over the pooled train+holdout censored set — that would require computing over holdout rows. NOT_CHECKED — would require holdout access; requires prospective record first. If it turns out to be the pooled figure, that is an additional undisclosed holdout read beyond the three disclosed in §5.1.

0.3 Internal self-contradiction

X1 — §5.1's own disclosure invalidates §5.1's own exemption list. CHECKED_AGAINST_CODE (the document is the artifact) / HOLDOUT_MARK_CHANNEL_BURNED_RETROSPECTIVE_ONLY §5.1 discloses three holdout computations, the third being "the identities of the 5 zero-support holdout mark events (§4, S2)". Two sentences later it concludes: "S0, S2, S3 and S4 are unaffected — their motivating numbers above were computed on training data only." S2's entire blocker — the five eventIds, the claim that holdout jump support reaches +5, and the −∞/HALT argument — is a holdout mark-channel derivation, disclosed in the same section that declares S2 unaffected. S2 must be listed alongside S1 and S1b as RETROSPECTIVE / EXPLORATORY on this holdout. I did not verify the five eventIds or the "+5" support claim: NOT_CHECKED — would require holdout access; requires prospective record first. (Train max jump is +4, verified — so the "+5" claim is only derivable from holdout.)

Q1 — Is the full motor-stack AIF model identifiable on current Wadhwa data?

Status: NOT_IDENTIFIABLE_CURRENT_DATA. CHECKED_AGAINST_CODE + CHECKED_AGAINST_RESULTS / TRAIN_ONLY for the training-side counts, HOLDOUT_ALREADY_SPENT_DURATION_ONLY for the resolution ceiling.

I agree with Track D's verdict and reach it with a stricter count.

Explicit parameter count vs data count

Instantiation S=8 occupancy states, K=2 hidden kinetic states, η_m 1-D lognormal random effect, |Π|=4, 6 levels — Track D's own spec, with C2 and C3 repaired.

Level	Block	Formula	dir-mark	jump geom.	jump sat.
Lmotor-0	per-state normalization scale `s_N` (omitted by track, C3)	`S`	8	8	8
Lmotor-1	hazard shape `k_{N,z}`	`S·K`	16	16	16
Lmotor-1	hazard scale (mean-one removes `S`)	`S·K − S`	8	8	8
Lmotor-2	kernel `q(z′\|z,N)` (corrected, C2)	`S·K·(K−1)`	16	16	16
Lmotor-2	initial `q(z₁\|N)`	`S·(K−1)`	8	8	8
Lmotor-0	direction mark `p(dir\|N,z)`	`S·K`	16	16	16
Lmotor-0	jump magnitude	`S·K·2` / `S·K·7`	—	32	112
Lmotor-4	`η_m` scale `τ` + loading	`1+d`	2	2	2
Lmotor-5	`Θ` hyperprior	`≥2`	2	2	2
policy	`ln E(π)`	`\|Π\|−1`	3	3	3
policy	precision `γ`	1	1	1	1
Lmotor-0	sensory precision `Π` per channel	3	3	3	3
all	per-level clocks	`6−1`	5	5	5
	TOTAL		88	120	200

Data count.

Fitting: 793 uncensored eligible training events (836 records if S0's censoring-complete likelihood is adopted), on 80 training motors. Note that 106 train motors exist in motors[] — 26 contribute no eligible event, so 26 η_m would be prior-only.
Discrimination: 19 holdout motors / 233 holdout events. TRAIN_ONLY for the first, published-artifact for the second.
Events per free parameter: 9.0 (direction), 6.6 (jump geometric), 4.0 (jump saturated).
Latents on top: η_m × 80 + z_i × 793 + π_i × 793 = 1666 latents against 793 observed events (Track D's arithmetic here is correct).

Information-content bound, which is the argument the track does not make. Each event carries: one positive scalar dt; one bit of direction; and jump over 8 observed train levels ≤ 3 bits. Even the saturated stack's 200 parameters are being asked to absorb ≈ 793 × (1 continuous + ≤ 4 bits). But the binding constraint is not fitting — it is discrimination, and the experimental unit there is the motor, of which there are 19 held out. A model with 200 parameters compared on 19 units is not underpowered by a factor; it is categorically outside the regime where a CI-bound verdict other than NOT_ESTABLISHED can be produced.

The decisive empirical bound. Nine models spanning 0 to ~8 effective free parameters plus a nonparametric KDE produce a total motor-equal spread of 0.1387 nats, and all 8 M3-contrasts are INCONCLUSIVE with the narrowest interval 0.0835 nats wide (C1). executionFindings[2] states it directly: "with 19 holdout motors every motor-equal M3-contrast interval contains zero." Adding 80–200 parameters cannot be resolved on a holdout that cannot separate a 1-parameter Weibull from a 3-component mixture from a KDE.

Q2 — Which parameters are identifiable, weakly identifiable, or not identifiable?

Per-block, with the evidence that decides each.

IDENTIFIABLE_WITH_CURRENT_DATA

Parameter	Why	Label / split
`s_N`, the 8 per-state normalization scales	Sample means over 40–150 train events each; `B4C05.fitted.M0.trainNLL == 793` exactly shows they are pinned	CHECKED_AGAINST_RESULTS / TRAIN_ONLY
Single-state marginal hazard shape (`k_N` in a `K=1` model)	M1's shape 0.6250888 is stable to 1e-4 across the frozen-exclusion and naive-inclusion treatments' structure; my independent grid refit reproduces 0.624971	CHECKED_AGAINST_CODE / TRAIN_ONLY
`p(direction \| N)` marginals, states 1–8	Every state has both directions in train (`on` 544 / `off` 249); the thinnest per-state cell is well populated	CHECKED_AGAINST_CODE / TRAIN_ONLY

WEAKLY_IDENTIFIABLE

Parameter	Why	Label / split
`k_{N}` for high states (7, 8)	Per-state `s_N` estimated on the fewest events; `scale_N[8] = 24.52 s` is a mean over a heavily right-truncated sample carrying most of the discarded 26.1 % exposure	CHECKED_AGAINST_CODE / TRAIN_ONLY
`p(z′\|z,N)` kernel entries for rare transitions	46 distinct `(state,next)` cells in the eligible cohort, only 23 with ≥ 10 observations (C4). The other 23 cells rest on ≤ 9 events each	CHECKED_AGAINST_CODE / TRAIN_ONLY
`η_m` scale `τ`	Fittable but flat. `B4C11.U2_profile.flatSetTauRange = [0.1133, 0.2331]`, `flatLogspan_raw_natural_log = 0.7213` — a 2-log-likelihood-unit flat set spanning a factor of ~2 in `τ`. And D1 withdrew the U4 bootstrap evidence; the ladder map records P6 as weakened	CHECKED_AGAINST_RESULTS / HOLDOUT_ALREADY_SPENT_DURATION_ONLY
`jump` magnitude law, low-`	jump	` cells

NOT_IDENTIFIABLE_CURRENT_DATA

Parameter	Why
Per-motor `η_m` (all 80, and prior-only for the other 26 train motors)	Median 7 events/motor, 22 motors with < 5, 1 motor with a single event. A 1-D random effect on a heavy-tailed hazard cannot be recovered from 1–7 dwells. `CHECKED_AGAINST_CODE / TRAIN_ONLY`
Per-event `z_i` and any `K ≥ 2` kernel, jointly with the hazard shapes	Classical mixture label-switching plus scale/shape trade-off. The empirical proof is already in the artifact: M6 (8 shapes) and M3 (2-parameter mixture) and M4 (3-component mixture, 5 free) are mutually `INCONCLUSIVE` on holdout. `CHECKED_AGAINST_RESULTS`
Per-event `π_i`	See Q3 — structurally unidentified, not merely underpowered
`γ` (policy precision)	Structurally unidentified from passive data; see Q3
Sensory precision `Π` per channel	Fully confounded with the hazard scale and the mark-model concentration. There is no repeated-measurement / replicate structure at the event level from which an observation-noise precision could be separated from process variance — every dwell is observed once
Per-level clocks (5)	The data have one clock: 0.02 s sampling (`sampleIntervalS = 0.01999999999998181`, uniform across train motors). Nothing in the derived dataset varies at a second timescale that could inform a per-level rate. `CHECKED_AGAINST_CODE / TRAIN_ONLY`
`Θ` hyperprior on `τ`	80 motors informing a hyperprior whose first-level parameter `τ` is itself flat over a factor of 2
Load-dependence of any of the above	7 levels × ≤ 3 holdout motors per level, and no within-trace timing (F11, C5)

Q3 — Does adding a policy π improve testability, or only add unidentifiable capacity?

Status: NOT_IDENTIFIABLE_CURRENT_DATA. Adding π adds unidentifiable capacity and, worse, adds a labelling risk with no compensating falsifier. CHECKED_AGAINST_CODE / NO_DATA_ACCESS_NEEDED

Testability requires that some setting of π change the predicted distribution of something recorded, in a way another setting does not. Grounded in the schema (verified: the event record has exactly 12 fields — direction, durationS, enteredAtS, eventAtS, eventId, jump, motorId, nextStateN, partition, rightCensored, splitRemainder, stateN), every non-identifier field is an outcome. There is no chosen quantity anywhere in events[]. In motors[] the only non-outcome is nominalElectrorotationSpeed, a per-motor constant with no onset time.

So p(o_future | π) cannot depend on π, and three things follow:

π_i per event adds 793 latents that the likelihood is flat in. If G is constant across policies, q(π) = softmax(ln E) and the data drop out of the policy posterior entirely — the marginal likelihood is unchanged by any π_i. This is not weak identification; the profile is exactly flat.
γ is not identified even under the "policies index predictive hypotheses" reading. Only the combination ln E(π) − γG(π) enters, and only up to an additive constant. Track D states this correctly in §3.3. See the API defect V3 below — §1.6 test 5 states the wrong equivalence and would not prove it.
The one thing π genuinely buys is negative. Under reading (iii), q(π) = softmax(ln E − γG) is Bayesian model averaging under a nonstandard weight. Calling that a policy makes a model-averaging weight look like a behavioural choice. Track D identifies this correctly and policies.PolicyKind.PREDICTIVE_HYPOTHESIS with is_evidence_bearing() == False is the right structural response.

Would π ever improve testability? Only if it changed the predictive factorization on a channel that is scored. It does not: score.py's per-event term is −log p(o_i) + log s_{N_i} over {dt} (± mark). Marginalizing a flat π out of that returns the same number. A stage that reported an improvement after adding π would be reporting an optimizer artifact.

Net: π adds ≥ 4 free parameters (ln E 3 + γ 1) and 793 flat latents, and buys zero discriminative power on a holdout whose entire resolvable range is 0.0417 nats. It should be built (as a typed refusal) and never fitted.

Q4 — What does G mean for a passive observational dwell dataset?

Status: ORCHESTRATE_LEVEL_ONLY. CHECKED_AGAINST_CODE / NO_DATA_ACCESS_NEEDED

I read audits/phase-b/b3-model-competition-runner.py in full-field terms and the event schema directly. What I found:

The runner's entire interface to the world is {motorId, stateN, durationS, rightCensored, partition} (verified by exhaustive field extraction — see F3 above). There is no action, no control input, no time-varying covariate, no intervention flag, no experimenter variable of any kind reaching the model.
The schema's remaining fields (nextStateN, direction, jump, enteredAtS, eventAtS) are also outcomes. Adding them (Track D's events.py) widens the observation space; it does not create an action space.
motors[].nominalElectrorotationSpeed is the only manipulated quantity in the repository, it is a per-motor scalar label with no onset, and re-deriving its timing requires the raw MAT, which is BLOCKED_EXTERNAL (B4C06.analysisStartIndex.3400, rawSha256 c14de12c… absent).

So, precisely: G(π) has a well-defined mathematical value here and a null empirical one. G(π) = KL[q(o_f|π) ‖ p(o_f|C)] + E_q[H[p(o_f|z_f,π)]]. Both terms are computable. Both are constant in π, because there is no π on which q(o_f|·) conditions. A G number emitted for wadhwa-2022-events would be a real number with no falsifier attached to it.

What a "policy" could range over, exhaustively, given the actual schema:

Candidate π	Well-defined?	What it would range over	Verdict
Agent action sequence	Yes	Nothing — no field is chosen	Not instantiable. The set is empty, not small
Experimenter load protocol	Yes	7 values of `nominalElectrorotationSpeed`	Not testable: constant per motor, no onset time, raw archive `BLOCKED_EXTERNAL`, ≤ 3 holdout motors at the thin levels (C5)
The motor's own stator-exchange trajectory `z_{1:n}`	Yes	Hidden-state paths	Category error and a truth-contract hazard. Scoring `G` over `z` asserts the organism performs the inference — precisely what `notEstablished[0]` disclaims. This is the reading that would produce "G proves motor agency"
Index over predictive hypotheses (models)	Yes	The 9 B3 models, or a `StackSpec` set	Bayesian model averaging under a nonstandard weight. Legitimate as inference, illegitimate as active inference
Analyst's next experiment	Yes	Experiment designs	This is the only live reading. See Q5

Track D's §3 reaches the same conclusion. I confirm it independently and add that the failure is structural (empty action set), not sample-size-limited — so no amount of additional Wadhwa-2022 data would fix it.

Q5 — Is G testable with current data, or ORCHESTRATE-level only?

Status: ORCHESTRATE_LEVEL_ONLY for the analyst-facing use; DESIGN_ONLY_UNTIL_INTERVENTION for any biological claim. CHECKED_AGAINST_CODE + CHECKED_AGAINST_RESULTS / NO_DATA_ACCESS_NEEDED

Not testable as a motor claim. The falsifier does not exist. status.verdict_from_ci refuses a point estimate and demands an interval; there is no contrast that could produce one, because there is no π-varying quantity to contrast.

Genuinely usable at the ORCHESTRATE level. G over experiment designs is well-posed and the ingredients are already in the repository:

Ambiguity E_q[H[p(o|z,π_design)]] — the expected residual entropy of a design, computable from the frozen fits' posterior predictive.
Risk KL[q(o|π_design) ‖ p(o|C)] where C encodes "designs on which competing models disagree" — the repository already quantifies disagreement: eight motor-equal contrasts all straddling zero, with widths ranging 0.083 → 0.329 nats. Expected information gain across designs is exactly the quantity that would rank "acquire the raw MAT for within-trace load timing" against "acquire a second study" against "add holdout motors."

This use is model-selection bookkeeping for the analyst, not a claim about the motor. It moves no P-level. It must never be reported in the same artifact scope as a motor claim, or the two readings of π will be conflated.

Concrete recommendation, sharper than the track's. expected_free_energy.py should not merely raise on NO_ACTION_CHANNEL; the ActionChannel dataclass should carry a channel_role field with exactly two admissible values, MOTOR_ACTION and EXPERIMENT_DESIGN, and claim_guard should refuse any artifact in which a G computed under EXPERIMENT_DESIGN appears alongside a motor-level P-level receipt. Track D's design leaves the ORCHESTRATE use undeclared, so an ORCHESTRATE-level G would flow through the guard unlabelled.

Q6 — Minimum intervention/transfer data to make G a biological motor claim

Status: DESIGN_ONLY_UNTIL_INTERVENTION. CHECKED_AGAINST_RESULTS for the power arithmetic / PROSPECTIVE_NEW_DATA_ONLY

Two things must be supplied. Neither is optional; the first without the second gives an untestable design, the second without the first gives an underpowered one.

(a) A recorded action channel with timing — the perturbation

Perturbation. A within-trace step change in external load, applied at an operator-chosen instant and recorded to the same 0.02 s clock as the stator trace. Concretely, for the Wadhwa electrorotation bead assay: hold a motor at nominalElectrorotationSpeed = s₁ for a pre-step window, step to s₂ at recorded time t_step, hold for a post-step window. Randomize (s₁, s₂, t_step) per motor from a pre-registered schedule so the intervention is exogenous to the motor's state.

Why this and not something else: it is the one manipulation the existing apparatus already performs (the 7 speed levels prove the actuator exists), and it acts on the mechanical variable the two-state remodeling hypothesis says the motor is tracking. It converts nominalElectrorotationSpeed from a per-motor label into a time-stamped action channel — which is the single missing schema element (Q4). Minimum new schema fields: actionAtS, actionFrom, actionTo, actionScheduleId, preRegisteredScheduleSha256.

Second, weaker option if new wet-lab work is impossible: recover t_step for existing traces from the raw MAT (c14de12c…). This lifts B4C06 out of BLOCKED_EXTERNAL and may yield within-trace load changes retrospectively. It cannot support a randomized-intervention claim (no pre-registration, no exogeneity guarantee) and it would be RETROSPECTIVE, not PROSPECTIVE.

(b) Enough motors — the power requirement, derived from this repository's own numbers

Back out the per-motor score-difference SD from the frozen contrasts. Narrowest interval (M4): half-width 0.0417 ⇒ SE ≈ 0.0417/1.96 = 0.0213 ⇒ σ_motor ≈ 0.0213·√19 = 0.093 nats. Widest (M0): σ_motor ≈ 0.366 nats. Required n = 7.85·σ²/δ² (two-sided α=0.05, 80 % power):

target effect δ (motor-equal nats)	n at σ = 0.093	n at σ = 0.366
0.10	7	105
0.05	27	421
0.025 (the observed M3→M2 gap)	108	1682

Minimum concrete specification:

≥ 60 motors held out for the intervention contrast, and a training set of comparable size — i.e. ≥ 120 motors total, roughly 6× the current holdout and 1.5× the current 129-motor study. Sixty holdout motors resolves δ ≈ 0.034 nats at σ = 0.093 and δ ≈ 0.13 at σ = 0.366; anything smaller cannot separate a G-driven prediction from M2_LOGNORMAL.
Paired within-motor design, pre-step vs post-step on the same motor. This is the leverage: it cancels the between-motor variance that drives σ up to 0.366, and it is the only way to reach the 60-motor figure rather than the 400-motor one. It also makes η_m a nuisance rather than a target.
≥ 2 step directions (load up and load down) per arm, so a monotone-drift confound is a testable alternative rather than an assumption.
Both load levels represented at ≥ 20 motors each. The current 6/6/3/2/2/2/2 (C5) is unusable at the thin end regardless of total n.

Measured response — the falsifier, committed before data. G becomes a motor claim only if it makes a risky prospective prediction that a non-AIF hazard model does not make. The pre-registered contrast:

Motor-equal joint NLPD on the post-step window, [G-driven policy model] − [best B3-class hazard model refit on pre-step data + a step-indicator covariate], paired motor-cluster bootstrap over the ≥ 60 held-out motors, 95 % BCa. Pre-registered predicted effect and direction committed and hash-sealed before the post-step windows are unblinded. An interval containing 0 ⇒ NOT_ESTABLISHED.

The discriminating observable that makes this genuinely risky, rather than a curve-fitting exercise: G predicts an anticipatory / ambiguity-reducing signature in the post-step dwell distribution that a reactive hazard model does not — e.g. a transient over-representation of exploratory z-transitions in the first post-step window that decays as the ambiguity term falls, with the decay rate tied to the same γ that the pre-step data fixed. If γ must be refitted per step to keep the prediction alive, G has been falsified and the correct record is CONTRADICTED, not a retuned γ.

Transfer (P4) is a separate, additional requirement. One study (Wadhwa 2022), no independent dataset, ladder map line 76 P4 NOT_ESTABLISHED. Even a perfect intervention result on this apparatus establishes P5 within one study, not P4. INDEPENDENT_TRANSFER_REQUIRED.

Module / API contract scan

Track D's design is unusually disciplined — the refusal-shaped API (assert_no_latent_in_blanket, NotAScoreError, assert_action_channel_present, assert_point_estimate_is_not_a_verdict) is the right architecture and I would keep almost all of it. The following would silently violate the contract if built as written.

V1 — CRITICAL. `score.py`'s scale term is not frozen, and S0 as specified can move every NLPD by up to 0.45 nats while looking like a model improvement.

CHECKED_AGAINST_CODE / TRAIN_ONLY

§1.10 states the per-event score as nlpd_i = −log p(o_i) + log scale_{N_i}, and cites b3-model-competition-runner.py:1046-1050. But scale_N is defined at :121-126 as the mean over train uncensored events per state. §1.4 (hazard_survival.py), §1.9 (fit.py) and §4-S0 all describe including the 43 censored training events in the likelihood, and never mention scale_N. There is no assert_scale_frozen anywhere in the proposed API.

If an implementer extends the training cohort to include censored events and lets scale_N follow — the natural reading of "refit with the censored events contributing log S" — here is the measured consequence (train-only recomputation, per-state):

state  scale_N (B3, train-uncensored)  scale_N (incl. censored)   Δ log scale  [nats added to every NLPD in that state]
  1              4.749153                    4.749153               +0.00000
  2              3.482609                    3.482609               +0.00000
  3              6.350886                    6.350886               +0.00000
  4              5.062716                    5.538293               +0.08978
  5              7.948454                    9.700400               +0.19919
  6             14.389919                   19.642520               +0.31117
  7             18.245270                   28.527607               +0.44697
  8             24.520584                   35.243567               +0.36277

Up to +0.447 nats per event, purely from renormalization. For scale: the entire nine-model leaderboard spans 0.139 nats, and the resolution floor is 0.042 nats. A silent scale_N shift is 3.2× the whole leaderboard and 10× the noise floor. §4-S0's stage table marks S0 "Comparable to B3 numbers? yes (same 233 events)" — same events is not sufficient; same normalization is the actual precondition, and it is nowhere stated.

Required fix: ScoredChannelSet must hash scale_N alongside the channel names, score.leaderboard must refuse to merge results with differing scale_N hashes, and hazard_survival.event_loglik must take scale_N as an explicit frozen input rather than reading it off a cohort. Without this, S0 — the one stage Track D recommends actually running — is the stage most likely to manufacture a false positive.