External-doors acquisition checklist — the parity gates no software can move

Drafted 2026-08-19, keyed to docs/MODEL-STATE-SNAPSHOT-2026-08-19.html. Planning artifact, uncommitted, for operator review. It relabels no gate and manufactures no pass. Every acceptance criterion below is quoted verbatim from the frozen pre-registration (experiments/cross-study-preregistration.v1.json) or the machine gate reports.

Why this document exists

The model sits at P3 (held-out predictive, as an executed activity) on the P0–P8 parity ladder. The next rungs — P4 transfer, P5 intervention, P7 independent replication — and eight BLOCKED_EXTERNAL gates cannot be closed by any amount of modelling, code, or analysis in this repository (docs/CROSS-STUDY-PARITY.md:251). "Fully replicating the behaviour in inference" is therefore a target world defined by external receipts we do not yet hold, not a software finish line. This checklist is the acquisition contract for those receipts.

A door is credited only when real external evidence lifts its status out of BLOCKED_EXTERNAL / NOT_ESTABLISHED — never by relabeling, and never by averaging a missing domain away (X16 stays FAIL while any required gate is non-PASS).

Door P4 — TRANSFER (first unsatisfied rung)

Gates: X10_CROSS_STUDY_PARAMETER_TRANSFER NOT_ESTABLISHED · H-AIF-G8 NOT_LOCATED · G08_LOAD_TORQUE_TRANSFER BLOCKED_EXTERNAL.

Acceptance criterion (verbatim, cross-study-preregistration.v1.json:107-109):

"At least one mechanistic parameterization frozen on one laboratory/study predicts a second laboratory's commensurate raw observations with a predeclared advantage over baselines. Source-paper predictions or unit-incompatible mappings do not count."

What it needs. A second, independent held-out cohort of single-motor stator/dwell series that is commensurate with the current lab across all six axes, then a freeze-on-A / score-B protocol:

1. Strain — same organism and genotype (E. coli behavioural evidence stays separate from Salmonella/Bacillus structural; a structural dataset does not transfer a behavioural claim).
2. Load — matched viscous-load / bead-size regime (G08 needs a full multi-load surface, not one post-electrorotation regime).
3. Perturbation — same intervention family; for G08/P5, recorded onset with paired pre/post on the same motors.
4. Calibration — instrument calibration released with the raw data, so OBSERVED is earned.
5. Units — the same observable (dwell duration, state, censor flag), so the frozen scoring rule transfers without a units change. Forcing a common coefficient across incompatible assays is forbidden — this is the exact axis on which the current corpus fails (cross-study-parity-report.json:653).
6. Observation operator — same measurement model mapping latent state to recorded signal.

Protocol. Freeze the mechanistic parameterization entirely on lab A → predeclare and commit the split + scoring rule + baselines before any lab-B held-out field is read (prospectivity is decided by the commit graph: the prediction commit must be a proven strict ancestor of the result commit) → score untouched lab B once. The cohort must carry enough independent motors: 19 holdout motors is nesting-blind, and the motor count at which a 0.042-nat contrast resolves at ≥80% is not yet measured on real data — derive it from a real-data power atlas, never choose N after seeing an interval width.

Falsifier. Transfer is refuted / H_PARITY fatally weakened if, on the independent cohort with everything predeclared: (a) a simple adversary (M2 lognormal) beats the motor-stack candidate by a CI-bound material margin above the 0.042-nat floor; or (b) parameter recovery fails on genuine out-of-distribution motors; or (c) the frozen lab-A parameterization simply fails to beat baselines on lab B's raw observations. Each is a legitimate reportable negative; none may be tuned away.

Candidate sources. Published single-motor bead-assay / fluorescent-switch stator-dwell datasets carrying dwell state + duration + transition target with motor identities — not Wadhwa-2022 (that holdout is spent, D5). Named load/torque anchors already in-repo: 10.1038/s41467-021-25774-2, 10.1038/s41467-019-13030-1. Natural targets to make commensurate: Nord 2017, Perez-Carrasco 2022 (load), Antani 2021 (switching), Ito 2021 (assembly; its 4.09 GB Class-A archive is integrity-verified but is one lab, not a transfer pair).

Blocking dependencies. The raw data/remodeling_data.mat archive is absent (blocks P2 re-derivation and Class-A transfer); no independent second stator-remodelling dataset exists in the repo (the single hard external blocker); D5 (burned mark channel) and D6 (nextStateN range defect) constrain any mark-process claim.

In-repo prep that is agent-authorable now (prepares the door, does not pass it): commit the transfer prediction record (split, baselines, scoring rule, motor-count target, falsifier), the <TASK>-DATA-ACCESS-PROTOCOL.md, and the freeze-A/score-B harness. The range-check repair, quarantine policy, and one-step TRAIN-only conditional are already built and tested.

Door P5 — INTERVENTION · "does a real bacterium implement active inference?"

Gates: G10 / X12_ACTIVE_INFERENCE_CAUSAL_IDENTITY NOT_ESTABLISHED (the gate proper) · G09_SWITCH_COOPERATIVITY BLOCKED_EXTERNAL (assay substrate / hard co-requisite — current data contain no switching trajectories). discriminatingInterventions: 0.

Acceptance criterion (verbatim, cross-study-preregistration.v1.json:119):

"A preregistered intervention must distinguish an Active-Inference-specific causal prediction from matched kinetic, control, and non-equilibrium statistical-mechanics alternatives."

The discriminating experiment (the crown of this whole programme). A 2×2 factorial live single-motor assay in E. coli:

• Factor 1 — stimulus gradient present vs absent (pragmatic / reward value on/off).
• Factor 2 — sensory ambiguity high vs low (epistemic value on/off).
• Held matched across ambiguity levels: mean ligand concentration, viscous load, proton-motive force.

The Active-Inference-specific signature is the epistemic term −information_gain in the expected-free-energy functional G(π) = risk + ambiguity − information_gain + effort (SCIENCE.md:81-89): AIF uniquely predicts that raising sensory ambiguity at zero gradient shifts the CW/CCW switching policy — an uncertainty-driven action taken when there is nothing to climb — with a pre-committed sign and magnitude. The three matched alternatives structurally cannot produce that effect, and each emits a committed contradicting prediction:

• (a) integral-feedback chemotaxis control → motor bias unchanged at matched mean, zero gradient;
• (b) non-equilibrium global mechanical coupling (Mattingly–Tu; in-repo M_GMC, X07 PASS) / catch-bond → switching set by torque/load/PMF, unchanged by ambiguity alone at matched drive;
• (c) kinetic hidden-state dwell (Wadhwa DLT; in-repo M_DLT, G02/G04 PASS) → dwell durations set by fitted intrinsic rates, insensitive to ambiguity at fixed mean stimulus.

This is expressly not Bayesian curve-fitting (the forbidden evidence): it is a prospective sign-and-magnitude prediction of a behavioural response to an intervention on sensory uncertainty, with mean stimulus and physical drive held constant. Secondary arm — an optogenetic / microfluidic CheY-P clamp that opens the feedback loop; AIF predicts residual uncertainty-modulated switching a static CheY-P→bias map cannot.

Apparatus. Single-motor readout at motor resolution (tethered-cell rotation or bead assay with back-focal-plane interferometry / high-speed switch-time video); a programmable microfluidic device with independently controllable mean, zero-mean fluctuation spectrum (ambiguity) and gradient; an in-vivo CheY-P FRET reporter (CheY-YFP / CheZ-CFP); matched-load beads / defined-PMF medium. Controls (pre-registered): dCheY/dCheA (loop broken — epistemic effect must vanish), dCheR dCheB (adaptation disabled), smooth/tumbly-bias. Unit = the individual motor/cell; switching events and frames are not independent replicates.

Falsifier. AIF-as-mechanism is falsified if the ambiguity factor produces no switching change at zero gradient beyond the pre-committed noise floor (data inside the "no-epistemic-effect" prediction, outside the AIF interval). The gate fails to discriminate (stays NOT_ESTABLISHED, never silently a pass) if any pre-enumerated serious mimic — e.g. a noise-adaptive-gain controller — matches the shift as well or better; such mimics must be frozen in advance or the discrimination is declared collapsed.

Estimated effort. ESTIMATE-class: ~12–36 months in one specialised single-motor bacterial-biophysics lab, then a second independent lab for G12/X14. Not on the software critical path.

In-repo prep that is agent-authorable now: freeze the four competitor models with calibration-only parameters (AIF EFE with frozen γ / C / A / B; an integral-feedback chemotaxis controller — NOT yet in-repo, must be added and frozen; M_GMC; M_DLT); write the prediction-record templates (point + interval, explicit sign + magnitude); draft the pre-registration for operator sign-off (S4). This prepares the door; only a live organism, an intervention on uncertainty, and predictions committed before outcome can pass it.

Door P7 — REPLICATION + LIVE + PHYSICAL

Three separable external acquisitions; each is individually required for X16 full parity.

Live signal chain — G11 / X13 (BLOCKED_EXTERNAL)

Acceptance (verbatim, cross-study-preregistration.v1.json:124):

"A calibrated instrument provides timestamped raw signal, uncertainty, prediction-before-outcome, and immutable audit linkage during the same live run."

The Web Serial aperture is already built (app/uni-flagellum-lab.tsx:539-575; serialAdapterImplemented: true). Missing: a real single-motor rig (tethered-cell / bead assay) driven by an MCU emitting the frozen serial contract — UTF-8 NDJSON at 115200 baud with t_ms, ligand_uM, motor_rpm, rotation, load_pNnm, pmf_mV (docs/HARDWARE.md:4-19) — plus a calibration record, an uncertainty band, the model version, and a prediction committed before the outcome (device time kept independent from browser receivedAtMs). Falsifier: the prediction commit is not a strict git ancestor of the outcome commit (retrospective); or uncalibrated; or the two clocks collapse; or required rotation fields missing/non-finite. Effort ~weeks — a bench-encoder shakedown can precede a wet motor.

Independent replication — G12 / X14 (BLOCKED_EXTERNAL, critical path)

Acceptance (verbatim, cross-study-preregistration.v1.json:129):

"A laboratory independent of the model authors reproduces the predeclared effects and releases raw data, calibration, exclusions, and analysis provenance."

A lab independent of the four corpus author-groups (Wadhwa 2022, Ito 2021, Antani 2021, Lisevich 2025) and of this software. Transfer the frozen protocol unchanged — including exclusion criteria and failure rules — and require publication of all raw files, calibration, exclusions, and provenance. independentLaboratories: 0. Repository replay and independent re-implementation do NOT count. Falsifier: shared personnel / shared code; or not all artifacts released; or the predeclared effects not reproduced under the protocol's own rules. Effort: months–quarters, fully external — this dominates the road to P7/P8.

Physical model validation — G13 / X15 (BLOCKED_EXTERNAL)

Acceptance (verbatim, cross-study-preregistration.v1.json:134):

"The exported UNI physical model is fabricated, measured against dimensional tolerances, and its sensor-to-screen math trace is tested with real inputs."

Fabricate cad/uni-flagellum-educational-model.scad (parametricCadExportImplemented: true), then run the 10 fabrication gates (docs/HARDWARE.md:64-73): tolerance coupon, caliper metrology, recorded fit/backlash, regenerate clearance from corrected CAD (never sand evidence away), per-gear print, free-rotation check, and confirm the Markov-boundary plate has only its two declared apertures with no undeclared shaft crossing. Then install/calibrate rotary encoders and bind the measured gear trajectory to the on-screen identity θ_posterior = θ_prior + θ_likelihood, reporting encoder error, backlash, and calibration residual. physicalPrintRuns: 0, measuredBacklashRuns: 0. Falsifier: measured angles/backlash miss the identity beyond the declared band; or an undeclared shaft crosses the boundary plate; or clearance was sanded rather than regenerated. Effort ~weeks, bench-scale, largely in-house (requires adult supervision for small rotating parts).

What must not be done

• No relabeling. A BLOCKED_EXTERNAL / NOT_ESTABLISHED gate leaves that status only on real external evidence. Turning a green software gate into a biological-parity claim is the contract violation this whole ladder exists to prevent.
• No averaging a missing external domain into a pass. X16 is conjunctive.
• No claim-wording drift. No unqualified "full/exact parity with nature", no "digital life" (the claim fence, CLAUDE.md). Verdicts are operator-signed (S4).
• Species discipline. E. coli behavioural evidence never conflated with Salmonella/Bacillus structural evidence, nor a live/printed/replicated OBSERVED measurement with model output.

Provenance

Every acceptance criterion above is quoted from experiments/cross-study-preregistration.v1.json (X10:107-109, X12:117-119, X13:124, X14:129, X15:134, X16:139) and the machine reports experiments/results/cross-study-parity-report.json and experiments/results/science-gates-report.json. Requirement detail traces to docs/CROSS-STUDY-PARITY.md:224-251, docs/SCIENCE.md:81-136, docs/HARDWARE.md:4-73, hierarchical-aif/protocols/NEXT-GATE-TRANSFER-AND-INTERVENTION-PLAN.md, and hierarchical-aif/protocols/MARK-PROCESS-TRANSFER-RESCUE-PROTOCOL.md. Assembled from the 3-agent external-doors mapping, 2026-08-19.