Motor-Stack Markov Blanket — Justification, Not Assertion

Status: BUILDER_SUPPORT_PROBE · Probe: 10 · Date: 2026-07-22 Gate mapping: informs H-AIF-G6 (scope ruling). Moves no P-level. Changes no frozen verdict. Creates no claim. Split boundary declared: TRAIN_ONLY for every new read performed for this document. Holdout motor/event counts quoted below are re-echoed already-published summary values, not new reads. Held-out mark channel: nextStateN / direction / jump were NOT read, printed, or reasoned from at any point in producing this document.

---

0. Why this document exists

A Markov blanket is a modelling choice about where to cut the world, not a fact discovered in the data. Any partition can be written down; only some partitions carry the conditional independence that makes them worth anything. This document therefore does two things the codebase did not previously do in one place:

1. it names, for every variable in the motor stack, which side of the cut it is on and why;
2. it states, for every one of those assignments, what breaks if the assignment is wrong.

The second column is the load-bearing one. A partition with no stated failure consequence is a diagram, not a model.

0.1 Two blankets, never to be conflated

The single largest source of category error in this lane is that there are two candidate blankets, and only one of them is instantiated.

	Blanket A — the ANALYST's (instantiated)	Blanket B — the MOTOR's (not instantiated)
internal states	mu, tau, and the per-motor belief q(eta_m) from hierarchy.posterior_motor_shape	motor chemistry / mechanics / switch complex
sensory states	recorded fields of events[]	ligand binding, PMF sensing, torque load sensing
active states	choice of next experiment — exercised at ORCHESTRATE level, outside the dataset	stator engagement / disengagement, switch
external states	the physical motor, the cell, and the apparatus	everything outside the motor
status	scored on the frozen split	entirely unobserved in this dataset

Everything scored in hierarchical-aif/ is Blanket A. free_energy.py:7-10 says this in the module docstring, and docs/MOTOR-STACK-AIF-SCOPE-RULING.md §2 says it again: computing F_motor is the analyst's fitting objective and it is not evidence that the organism performs inference. Blanket B is DESIGN_ONLY. Where this document says "internal state" without qualification, it means internal to Blanket A.

0.2 The formal condition being asserted

The standard partition is external psi, sensory s, active a, internal mu, with blanket b = s ∪ a, and the substantive requirement

mu  ⫫  psi  |  b

That is the whole content of the choice. Naming boxes costs nothing; the conditional independence is what would have to hold. §5 examines whether it does, and §6 states the reason it currently cannot be checked.

---

1. Data actually available (verified, this session)

Source: experiments/data/wadhwa-2022-events.json, schema uni.flagellum.observed-events/1.0.0, protocol UNI-FLAGELLUM-OBS-001. Loaded via motor_stack_aif._bridge.b3().load_events() and _bridge.frozen_cohort().

quantity	value	how obtained
event records, all partitions	1349	len(b3().load_events())
event record fields	12: direction, durationS, enteredAtS, eventAtS, eventId, jump, motorId, nextStateN, partition, rightCensored, splitRemainder, stateN	key names only
motor records	129	len(d['motors'])
motor record fields	8: analysisSamples, motorId, nominalElectrorotationSpeed, partition, runCount, sampleIntervalS, sourceSamples, splitRemainder	key names only
frozen cohort derived_eligible_1_to_8	80 train motors / 793 train events; 19 holdout motors / 233 holdout events; states (1..8)	frozen_cohort() (already-published values)
TRAIN-only nominalElectrorotationSpeed levels	7 levels over exactly 80 train motors: {50: 5, 100: 6, 150: 6, 200: 17, 250: 20, 272: 9, 300: 17}	new TRAIN_ONLY read, this session
TRAIN-only sampleIntervalS	one distinct value, 0.01999999999998181 s, for all 80 train motors	new TRAIN_ONLY read, this session
TRAIN-only runCount	26 distinct values spanning 3 to 83 recordings per motor	new TRAIN_ONLY read, this session
holdout-side values of the three fields above	NOT_READ — deliberate channel-spend avoidance	—

The motor record contains no load, no PMF, no stator number, no temperature, no viscosity, and no CheY-P field. That is a schema fact, verified above, and it drives §4 and §6.

---

2. The partition

Legend: INT internal (Blanket A) · SENS sensory · ACT active · EXT external · OBS-BLANKET an element of the recorded blanket surface · META not a state at all.

2.1 durationS — SENS (observed blanket, Lmotor-0)

Justification. It is the channel — the only channel — from which mu, tau and q(eta_m) are updated. It is causally downstream of the motor's hidden dynamics and upstream of nothing in the model. hierarchy.motor_log_marginal consumes exactly the normalised y derived from it, and the frozen NLPD is published in seconds as normalised-y NLPD + log scale_N[state].

Consequence if wrong. If durationS is not conditionally sufficient — that is, if it also carries apparatus and analysis-pipeline state — then the sensory channel is contaminated and eta_m absorbs instrument variation along with motor variation. One contamination route is closed on the train side and one is open:

• closed: sampleIntervalS is a single constant 0.02 s across all 80 train motors, so dwell-time resolution does not vary between train motors.
• open: runCount spans 3 to 83 per motor. Recording effort is not constant, it is nested inside motorId, and it therefore travels with eta_m. Whether that matters is NOT_MEASURED; it is a live confound route, not a demonstrated one.

2.2 rightCensored — SENS, but excluded rather than conditioned in the frozen cohort

Justification. It is a recorded fact about the observation process, produced jointly by the motor and the recording window. hazard_survival.log_event_density takes a censoring vector and hierarchy.motor_log_marginal threads censored_motor through, so the F-side implementation is built to condition on it.

Consequence if wrong. The frozen cohort excludes right-censored events entirely. Exclusion is not conditioning. Dropping the longest, truncated dwells is a length-biased selection on the sensory channel, which shifts the tail the shape parameter is estimated from. This is already on record as Lmotor-1 ABSENT in B3 in docs/GAP-AUDIT-FULL-HIERARCHICAL-MOTOR-MODEL.md:40. The magnitude of the induced bias is NOT_MEASURED in this document. If the classification is wrong — if censoring is really an external nuisance independent of shape — nothing breaks; if it is right, every shape estimate on this cohort carries an unquantified selection bias.

2.3 stateN — SENS / conditioning index (Lmotor-3)

Justification. Recorded per event; selects the frozen per-state normalisation scale_N. It is the observed index of motor occupancy that makes dwell times from different stator numbers commensurable.

Consequence if wrong. stateN is arguably itself a reconstruction — a stator number inferred by a step-detection procedure from a bead trace — rather than a directly recorded measurement. Under the truth contract a reconstruction may not be relabelled observed, and whether this field earns the OBSERVED label depends on source pinning that this document did not verify (NOT_CHECKED — would require the raw archive, which is BLOCKED_EXTERNAL). If a stateN assignment is wrong, y = durationS / scale_N[stateN] is mis-normalised and the published seconds-scale NLPD is shifted by log scale_N[state] for that event — the error enters twice, once in the likelihood and once in the reporting scale.

2.4 motorId — META. Not a state.

Justification. docs/DATA-CHANNEL-SPEND-LEDGER.md:44 rules it "Structural, not an observable", permits it for "Defining the split; motor-cluster resampling", and forbids "Using motor identity as a predictor". Its two legitimate roles are (a) the frozen split sha256_mod5(motorId)==0 => holdout and (b) marking the exchangeable unit — events.split_by_motor carries the comment "Group events by motor - the experimental unit. Never group by event."

Consequence if wrong — and this one is not hypothetical. D1 is exactly the failure mode. Per ledgers/HIERARCHICAL-AIF-DEFECT-CLOSURE-LEDGER.md:30, the C11 bootstrap "regrouped resampled motors by motorId, collapsing K draws of a motor into one K-fold group. 80 draws → 46 groups at the declared seed." The objects being grouped there were bootstrap draws, which are exchangeable by construction; grouping them by the identity label destroyed that exchangeability. The lesson generalises: a grouping key is safe only where the grouped objects are the units it indexes. P6 for C11 U4 is withdrawn until the corrected run lands; P6 for duration-only B3/B4 is unchanged.

2.5 eta_m, the per-motor log-shape — INT (of Blanket A)

Justification. eta_m ~ Normal(mu, tau^2), k_m = exp(eta_m), integrated out by 33-node Gauss-Hermite quadrature and never estimated as a free parameter. The whole model carries 2 free parameters (mu, tau), so parameter count does not grow with motor count — hierarchy.py:14-17 records that this is the point of the constraint. posterior_motor_shape returns q(eta_m) as a mean/sd pair; that is a belief held by the analyst.

Consequence if wrong — two distinct failures.

• If eta_m is read as an internal state of the motor, the document has asserted that the organism carries the representation. free_energy.py:7-10 and the scope ruling both disclaim that reading, and the B3 result's own notEstablished[0] disclaims it as well.
• If eta_m is in fact absorbing external variation, then it sits on both sides of the cut and the partition is invalid as a blanket. This is the strongest objection and it gets §6 to itself.

Unit note. tau here is the population SD of the log-shape and is dimensionless. It is not a torque. The thermodynamic tau * delta_theta is a different quantity with different units and is never mixed with variational free energy in this stack.

2.6 Latent kinetic mode z_i (Lmotor-2) — EXT, declared and NOT INSTANTIATED

Justification. hierarchy.py:11 records Lmotor-2 kinetic mode : NOT INSTANTIATED - no identifiable capacity, and hierarchy.py:19 records the reason: "Adding Lmotor-2 policies would add unidentifiable capacity, not testability." With 793 training events over 80 motors against a corrected motor-equal resolution floor of ≈0.042 nats (BCa half-width of the narrowest frozen B3 contrast, M4_MIXTURE_K3: BCa width 0.08414086126525253), a per-event hidden mode cannot be constrained. Declaring the layer external and empty is a choice recorded in code, which is the correct way to hold a boundary.

Consequence if wrong. If a real kinetic mode drives dwell statistics, the mean-one Weibull with a per-motor shape is a marginalisation over it, and the fitted shape is a mixture artefact rather than a motor property. Note what the evidence says about detecting that: B4C02 (landed, 0633988d…, full frozen N=200/generator, 600 sims, 0 failures) found M2-over-M3 at frac 0.9400 for three_timescale_heavy_tail but 0.0050 for both weibull_gamma_blend and per_motor_heterogeneous_weibull — 1 of 3 generators — so the frozen GENERATOR-ROBUST_ADVERSE label was REFUTED. Heavy-tailed dwell shape does not uniquely identify a hidden-mode structure. The Lmotor-2 omission is therefore not currently detectable from shape alone.

2.7 External biological conditions — EXT, UNOBSERVED IN THIS DATASET

Load, PMF, stator state, temperature, viscosity, CheY-P condition.

Justification. §1 verified that the 8-field motor record contains none of them. The nearest recorded quantity is nominalElectrorotationSpeed, which reports/ULTRACODE-TRACK-D-VERIFICATION.md:195 calls "the only manipulated quantity in the repository", and which is "a per-motor scalar label with no onset". This session independently confirmed 7 levels over the 80 train motors: {50: 5, 100: 6, 150: 6, 200: 17, 250: 20, 272: 9, 300: 17}.

The distinction that matters: UNOBSERVED is not ABSENT. These variables act on every motor in the recording. They are missing from the record, not from the world. The consequences run in both directions:

• Treating them as absent would license reading the fitted model as a description of motors in general, when it is a description of motors under whatever unrecorded conditions happened to obtain. Every unsampled condition is extrapolation-only, never supported.
• Treating them as unobserved external states is the honest classification, and it immediately implies the model cannot license any condition-conditional statement at all.

Consequence if wrong. If any of these varies within a motor during recording and materially moves dwell statistics, then even the per-motor eta_m is a time-average over a varying external state, and the exchangeability of events within a motor — the assumption motor_log_marginal relies on when it integrates a motor's events jointly against one eta_m — fails. That is NOT_CHECKED: no within-recording condition timeseries exists in this dataset.

2.8 Action / policy status — ABSENT. The active set is empty.

Treated in §3 as a required ruling.

2.9 Remaining recorded fields

field	class	note
sampleIntervalS, analysisSamples, sourceSamples, runCount	EXT — observation-process state	Constant 0.02 s interval on train motors; runCount 3–83, nested in motorId, an open confound route (§2.1)
enteredAtS, eventAtS, eventId, splitRemainder, partition	META — bookkeeping	Ledger line 45: "Not scientific observables"; forbidden use is "Use as covariates"
nextStateN, direction, jump	OBS-BLANKET, quarantined	Train-side available; holdout side is D5-burned, RETROSPECTIVE_EXPLORATORY_ON_THIS_DATASET, and events.load_events raises HoldoutMarkAccessError unless the caller acknowledges it in writing. Not used here.

---

3. Required rulings, stated plainly

R1 — The active channel is ABSENT, so the blanket is incomplete in the formal sense

There is no action field in the event record and no action field in the motor record. The set is empty, not small. docs/MOTOR-STACK-AIF-SCOPE-RULING.md:51 records this, and it is structural: no additional quantity of Wadhwa-2022 passive recording would create the channel.

Therefore, in the standard partition b = s ∪ a, this stack has a = ∅ and b = s. A partition without active states is not a full Markov blanket. What is instantiated is a sensory-only boundary — a conditional-independence screen over a passively observed system. The document says this rather than presenting the partition as complete, because the missing half is precisely the half that would carry any agency reading. The blanket is incomplete, and the incompleteness is declared, not repaired by wording.

R2 — F-side blanket is testable; G-side active policy is design-only until intervention

F_motor = E_q[ ln q(Theta, eta, z) - ln p(o, z, eta, Theta) ] = complexity - accuracy is a belief-update objective over recorded observations. It needs a likelihood, a hierarchy over the experimental unit, and an approximate posterior — all three exist here, and the F-side model was scored on the frozen split (held-out motor-equal NLPD 3.4326923382675303, 2 free params mu = -0.41607215987582913, tau = 0.18372082607308418, train NLL 575.6701064153622, determinism proven byte-identical over two full executions, independent-oracle residual exactly 0.0).

G_motor(pi) ranges over policies over actions. With a = ∅ there is nothing for pi to range over. G-side is DESIGN_ONLY_UNTIL_INTERVENTION_OR_TRANSFER, and the fence is enforced in code: free_energy.py:12-14 records that there is deliberately no expected-free-energy function, and three separate tests assert the absence (test_fside_motor_stack.py::test_no_expected_free_energy_function_exists, test_F_motor_observational_objective.py, test_no_G_biological_claim_in_passive_data.py). F is an objective over beliefs; G would be an objective over policies. They are not the same quantity and are never summed.

R3 — motorId is metadata and a grouping key, never a state

See §2.4. The forbidden use is naming motor identity as a predictor. D1 is the recorded cost of using the grouping key where the objects being grouped were not the units it indexes.

R4 — External biological conditions are UNOBSERVED, not absent from the world

See §2.7. The blanket may therefore license statements about the recorded observable, and may not license statements about how motors behave under conditions the record does not carry.

R5 — This partition is a modelling choice, and it is currently justified by design coherence

rather than by a resolved predictive gain

Stated bluntly because it is the weakest point in the lane, and it is adverse:

The F-side hierarchy is numerically the same object as the frozen M7 baseline. exp(mu) = 0.659632669755436 against M7 k = 0.6596322379287862; tau = 0.18372082607308418 against M7 tau = 0.18372185667134974. The M7 contrast came out at +2.506984e-07 nats [+1.604451e-07, +3.374688e-07] — RESOLVED_ABOVE by the frozen CI rule, and SCIENTIFICALLY_NULL, roughly 168 000× below the ≈0.042 nats floor. That pairing is D10: a paired motor-cluster bootstrap resamples motors, so it resolves a difference of any magnitude whose sign is consistent across motors. The floor states what is scientifically material; the bootstrap states only what it will call.

The honest reading: the blanket partition as scored has not been shown to buy resolved predictive content over a model carrying the same two numbers. Its justification today is coherence of the partition and identifiability discipline, not a resolved win. The candidate ranks 5th of 10 combined.

---

4. What the partition would have to satisfy, and what is checkable

required property	status	why
mu ⫫ psi | b (internal independent of external given the blanket)	NOT_CHECKED	psi — load, PMF, stator number, temperature, viscosity, CheY-P — is unobserved (§2.7). An independence claim cannot be tested against a variable that was never recorded.
Events exchangeable within a motor given eta_m	NOT_CHECKED	Requires a within-recording condition timeseries; none exists.
Motors exchangeable given (mu, tau)	partially supported	population_log_likelihood assumes it. nominalElectrorotationSpeed varies between motors at 7 levels, which is a known between-motor inhomogeneity the model does not condition on.
Sensory channel free of instrument variation	partially supported	sampleIntervalS constant at 0.02 s on train motors; runCount 3–83 remains an open route.
Active states well defined	FAILS by construction	a = ∅ (R1).

---

5. The strongest objection

Objection. If the external states that actually drive dwell statistics — load, PMF, stator number — are unobserved, and if they differ between motors, then a per-motor latent fitted to per-motor dwell data will soak up that between-motor external variation. Under that reading eta_m is not an internal property of a motor at all; it is a condition-absorbing nuisance term wearing an internal-state label. Both readings produce the same likelihood, the same fitted (mu, tau), and the same held-out NLPD.

This objection is not speculative here — the design makes it unfalsifiable on the current data. The one recorded condition proxy, nominalElectrorotationSpeed, is constant within each motor. It is therefore perfectly nested inside motorId, and eta_m is indexed by motorId. There is zero within-motor contrast in the only external covariate the dataset carries. A design with no within-unit variation in a covariate cannot separate that covariate from a unit-level latent — the two are collinear by construction, not by bad luck.

The current data cannot distinguish the two readings. That is a statement about identifiability, not about effect size. It is not a finding that eta_m is internal, and it is not a finding that it is external. Both remain live.

5.1 Discriminators, ranked by strength

id	design	what separates the readings	availability
D-1	Within-motor load step. Hold a motor at level s1, step to s2 at a recorded onset t_step timestamped against dwell boundaries, hold a post-step window. Randomise (s1, s2, t_step) per motor from a pre-registered schedule. Refit q(eta_m) on the pre and post windows separately.	Internal-state reading predicts q(eta_m) is stable across the step relative to its own posterior sd. Absorber reading predicts it shifts with the step. Within-unit, so it does not depend on between-motor exchangeability.	PROSPECTIVE_NEW_DATA_ONLY. Blocked twice: nominalElectrorotationSpeed has no onset time, and the raw archive is BLOCKED_EXTERNAL (B4C06.analysisStartIndex.3400, rawSha256 c14de12c… absent).
D-2	Test–retest on the same motor. Two recording sessions separated in time under nominally identical conditions; independent q(eta_m) from each.	Internal reading predicts across-session agreement beyond what the motor's condition assignment explains. Absorber reading predicts agreement no better than the condition label predicts.	NOT_CHECKED. runCount exists per motor but the frozen cohort does not separate sessions into independent estimates; re-derivation needs the raw archive (BLOCKED_EXTERNAL).
D-3	Between-motor covariate association, train-only. Regress the posterior eta_m mean from hierarchy.posterior_motor_shape on nominalElectrorotationSpeed across the 80 train motors, with a motor-level bootstrap.	A resolved association is evidence eta_m carries external variation.	Runnable today on train data with no new spend — and it is the weakest of the three.

5.2 Why the only runnable discriminator is the weakest — stated as the honest headline

D-3 is the one test available now, and three separate limits cap what it could deliver:

1. It cannot establish direction. Even a resolved association is confounded unless level assignment was exogenous to the motor. Whether assignment was randomised is NOT_CHECKED and verifying it requires the raw archive (BLOCKED_EXTERNAL).
2. A null result would establish nothing. 7 levels with cells of 5, 6, 6, 17, 20, 9, 17 motors is thin. An interval crossing zero is NOT_ESTABLISHED, never "no association" and never "independent". Underpowered is not equivalence.
3. A positive result would not rescue the partition either. It would show eta_m is partly external, which invalidates the internal-state label without telling us what the internal state would have been.

D-3 was NOT_RUN in this probe — deliberately. It would produce a new analysis number, and a builder-support probe may not create a claim. It is specified here precisely enough to be pre-registered and run as its own bounded change, with its prediction committed first.

---

6. What would kill this lane

kill condition	effect
D-1 shows q(eta_m) moves with a within-motor load step by more than its posterior sd (DESIGN_ONLY: the 1x-posterior-sd cut is introduced by this document, is NOT a frozen criterion, and must be pre-registered before any such experiment)	The internal-state reading of Lmotor-4 is contradicted. eta_m must be relabelled a condition-absorbing term. The F-side held-out NLPD survives as a predictive statement; its blanket interpretation is withdrawn.
D-3 shows a resolved association with the load proxy and assignment is shown exogenous	Same relabelling, on weaker evidence, retrospectively only.
A within-recording condition timeseries shows the external state varies during a recording	Within-motor event exchangeability fails; motor_log_marginal's joint integration over one eta_m per motor is misspecified.
stateN is shown to be a reconstruction whose error is state-dependent	scale_N normalisation and the published seconds-scale NLPD are both biased; the leaderboard ordering would need rechecking.
Censoring exclusion is shown to bias the shape estimate materially	Every shape on this cohort carries a selection bias; the ≈0.042 nats floor would need re-derivation.

What would NOT kill it, and what would not save it: any additional quantity of passive Wadhwa-2022 recording. The blocking limitation is the absence of a within-unit manipulated variable with a recorded onset. That is structural. Transfer to an independent dataset carrying condition covariates, or intervention, are the two routes — P4 transfer is the first unsatisfied level on the ladder, and P4/P5/P7 cannot be closed by any modelling in this repository.

---

7. Adverse findings and limitations

1. The blanket partition has not been shown to buy resolved predictive content. §3 R5. The F-side hierarchy equals frozen M7 to ~7 significant figures in both parameters, and the contrast against it is SCIENTIFICALLY_NULL.
2. The central conditional independence mu ⫫ psi | b is untestable on this dataset, because psi was never recorded. §4.
3. eta_m and the sole recorded condition proxy are collinear by design. Zero within-motor contrast. §5.
4. The blanket has no active half. a = ∅, structurally. §3 R1.
5. nominalElectrorotationSpeed is absent from docs/DATA-CHANNEL-SPEND-LEDGER.md. It is a real per-motor field, it is the only manipulated quantity in the repository, and it has no ledger row assigning it a split status or a permitted claim. This document read it TRAIN_ONLY and declares that spend here. The ledger gap is a defect-shaped finding and should be routed, not just noted.
6. runCount varies 3–83 per motor and is nested in motorId, so recording effort travels with eta_m. Impact NOT_MEASURED.
7. Censored events are excluded, not conditioned, in the frozen cohort. Bias magnitude NOT_MEASURED.
8. No FLOW card was appended to reports/FLOW-JOURNAL.jsonl by this probe, to avoid a concurrent write while two long runs are in flight. The parent orchestrator should append it.

---

8. Standing frame

Nature supplies the architecture candidate.
The gate supplies the status.
A partition is a choice; only the conditional independence it asserts can be earned.

The partition in §2 is defensible as a design. It is not established as the correct cut of the world, and the one measurement that would test its central claim does not exist in this dataset.

NEXT_ACT = Open a defect-ledger row for the unledgered nominalElectrorotationSpeed channel (assign it a split status, permitted claim, and forbidden claim in docs/DATA-CHANNEL-SPEND-LEDGER.md), then pre-register discriminator D-3 as its own bounded TRAIN_ONLY change — prediction and falsifier committed before the regression of posterior eta_m on the 7 train load levels is ever run.