G-SIDE INTERVENTION DISCRIMINATOR — design protocol

Status: DESIGN_ONLY_UNTIL_INTERVENTION_OR_TRANSFER. This document is a builder-support probe. It moves no P-level, changes no frozen verdict, and creates no claim.

0. FENCE (binding; stated first, never violated below)

• The Wadhwa-2022 derived_eligible_1_to_8 dataset is passive. The action set is empty. That is STRUCTURAL, not sample-size-limited — no quantity of additional passive dwell-time observation substitutes for an intervention. This is the reason recorded in hierarchical-aif/docs/MOTOR-STACK-AIF-SCOPE-RULING.md §3 and enforced verbatim by hierarchical-aif/tests/motor_stack_aif/test_no_G_biological_claim_in_passive_data.py.
• expected_free_energy does not exist in motor_stack_aif and must not be added. Its absence is the fence, not an oversight. Two test modules enforce it (test_no_G_biological_claim_in_passive_data.py, test_fside_motor_stack.py:: test_no_expected_free_energy_function_exists). This protocol does not propose adding it, and no section below may be read as authorising it. The symbol may be introduced only after intervention data with a recorded onset time exists in the repository.
• F is not G. F_motor is a variational free energy over beliefs given recorded observations — an analyst's fitting objective, evaluated after the observation exists. G_motor(pi) is an expected free energy over policies, scored before the action is taken, and it presupposes an action to score. Computing F_motor asserts nothing about whether the organism performs inference. Nothing in this document merges the two.
• Underpowered is not equivalence. Every verdict contemplated here is CI-bound. An interval crossing threshold is NOT_ESTABLISHED — never "no effect", never "equivalent".
• This document proposes wet-lab data acquisition, which this repository cannot perform. P5 intervention therefore remains NOT_ESTABLISHED and cannot be closed by any modelling in this repository.

1. Units and symbol discipline

Symbol collision, explicit: the tau fitted by fit.fit_motor_stack (0.18372082607308418, source F_SIDE_MOTOR_STACK_SCORING_RESULT.json, sha256 b3b12720…) is the dimensionless population SD of the log-shape. It is not a torque and must never be multiplied by an angle. Thermodynamic work stays in its own units and its own paragraph.

2. The three requirements already on record

From hierarchical-aif/docs/BIOLOGICAL-PARITY-RECEIPT-MAP.md §"The G-side requirement, stated concretely" (lines 51–57), restated identically in MOTOR-STACK-AIF-SCOPE-RULING.md §4. All three are required; no one of them is sufficient:

1. A manipulated variable with recorded onset time — a load step, PMF step, stator-availability step, or electrorotation change, timestamped against dwell boundaries.
2. Paired pre/post observation on the same motors, so the response is measured within unit.
3. Enough independent motors under each condition to give a CI-bound verdict on the response.

The receipt map's own note that the exact motor count is "to be derived, not guessed" is honoured in §4 below: it is reported as NOT_MEASURED, with the measurement machinery supplied and the one measured variance anchor that exists labelled for what it actually is.

3. Candidate perturbations

The experimental unit is the MOTOR throughout. Dwell events within a motor are not independent replicates; every design below aggregates to a per-motor statistic first, then resamples motors. t_onset is mandatory in every row — without it, pre and post cannot be separated and the manipulated variable is not a manipulated variable.

3.1 Load perturbation (viscous load / bead size)

3.2 Stator availability perturbation

3.3 Energy / ion-motive-force perturbation

3.4 Recovery after perturbation (time course back to baseline)

3.5 State-transition response (does dwell-state structure change under perturbation)

4. Minimum independent motors per condition — what is measured and what is not

4.1 The honest headline

NOT_MEASURED for the intervention design — and the premise below is now STALE. A power atlas was produced in the same probe batch as this protocol (hierarchical-aif/results/motor_stack_aif/power_atlas.json, reports/POWER-ATLAS-MOTOR-EQUAL-SCORING.md). It is a synthetic motor-equal resolve-rate atlas and does not supply a minimum-N for an intervention design, which needs a within-motor pre/post contrast the atlas never simulates. The required minimum N therefore remains NOT_MEASURED — for that reason, not for absence. Historical note — when this section was drafted, a search of hierarchical-aif/docs/, hierarchical-aif/ledgers/, and the whole hierarchical-aif/ tree returns no power-atlas artifact (command and output in the return record). Deriving a required N needs a minimum effect size of interest and a variance of the intervention response. Neither exists, because no intervention has been performed. Any number stated for these would be invented, and this document does not state one.

4.2 The one variance anchor that does exist — and what it is not

Computed from hierarchical-aif/results/motor_stack_aif/F_SIDE_MOTOR_STACK_SCORING_RESULT.json (sha256 b3b12720…), field perMotorNLPD, across the 19 frozen holdout motors:

Two consistency checks, both passed, both against artifacts not used to produce the numbers:

• the mean paired difference 0.0016409948078282934 reproduces the frozen F-side M3 contrast point estimate +0.001641 (established fact);
• the normal-approximation half-width at n=19, 1.96 × 0.18831385847627796 / sqrt(19) = 0.0846762376, sits close to the published M3 percentile half-width (0.080849 − (−0.077147)) / 2 = 0.0789980000.

What this anchor is NOT. It is the per-motor SD of a model-versus-model NLPD difference on passive duration data. It is not the variance of a motor's response to a manipulated variable. There is no evidence that an intervention response has this variance, and assuming it does is an assumption with no support. It is reported here because it is the only measured motor-level paired variance in this repository, and because the pairing gain of 7.30× is the one transferable design lesson: requirement (2), paired pre/post on the same motors, is not a nicety — on the only motor-level paired quantity ever measured here it reduced the SD of the difference by a factor of 7.3, i.e. it cut the required motor count by roughly its square.

4.3 The sizing relation (supply the machinery, not a fabricated N)

For a per-motor paired response with SD s_d (nats or whatever the response unit is) and a target CI half-width d in the same unit:

n_motors  ≈  ( 1.96 * s_d / d )^2          [normal approximation; the frozen path uses a
                                            paired motor-cluster bootstrap, which is what
                                            any real design must be sized against]

s_d for any intervention response is NOT_MEASURED. A pilot arm exists precisely to measure it, and the pilot's own N is not derivable in advance either.

DESIGN_ONLY illustration — NOT EVIDENTIAL, and not a prediction about any intervention. Substituting the §4.2 NLPD paired SD 0.18831385847627796 purely to show the shape of the curve:

The 0.042 row uses the corrected motor-equal resolution floor (BCa half-width 0.04207043063262626, narrowest frozen B3 contrast, M4_MIXTURE_K3) as d. These rows are DESIGN_ONLY. They describe an NLPD contrast, not an intervention response, and they must not be quoted as a required motor count for any perturbation in §3.

4.4 The minimum-effect-size guard is mandatory here (D10)

D10 recorded that the frozen CI rule has no minimum-effect-size guard: a paired motor-cluster bootstrap resamples MOTORS, so it resolves a difference of any magnitude whose sign is consistent across motors. M7 exposed this at +2.506984e-07 nats — roughly 168 000× below the floor — and was repaired by added interpretation (scientificReading, reportableAsAWin), never by re-thresholding.

An intervention design is exactly where this bites hardest, because a systematic instrument shift between the pre and post arms (focus drift, stage settling, illumination change) is perfectly sign-consistent across motors and will resolve. Therefore every arm in §3 must pre-register:

• a minimum effect size of interest, declared before unblinding, in the response's own unit — NOT_MEASURED at present and to be set by domain judgement, not by the observed data;
• a sham / time-matched control arm whose measured shift bounds the instrument-drift floor;
• reporting of both the interval and the scientific reading, exactly as the D10 repair does.

5. THE CORE QUESTION — what does policy selection predict that M2 / M4 / M7 do not?

5.1 The answer, stated without softening

Against M2 / M4 / M7 as they actually exist: nothing testable, because those models make no prediction about an intervention at all.

M2_LOGNORMAL, M4_MIXTURE_K3, and M7_HIERARCHICAL_MOTOR are unconditional dwell-duration densities. None of them takes a manipulated variable as an argument. Under a load step, a stator step, or a PMF step, they are silent, not wrong. "Policy selection predicts a response and M2 does not" is therefore not a discriminator — it is a comparison against an undefined baseline, and treating it as evidence would be the emptiest possible way to appear to close P5.

Any honest discriminator must first extend the incumbents to stimulus-conditioned form: M2/M4/M7 refitted with the manipulated variable and t - t_onset as covariates, on identical splits and the identical scoring rule. Only then does the comparison have two live sides. That extension is a prerequisite of the experiment, not a result of it.

5.2 Once the incumbents are stimulus-conditioned, what is left?

Three candidate signatures were examined. Two collapse.

5.3 Why the one real axis is not instantiable at the motor

Testing the epistemic term requires an observable action channel in which information gain and setpoint attainment dissociate — a situation where the system can pay a measurable pragmatic cost to obtain a measurable reduction in uncertainty, and where the analyst can observe the action separately from the state it produces.

At the single flagellar motor, with the observables this project has or could plausibly acquire:

• there is no recorded action channel at all (ObservedEvent carries event_id, motor_id, partition, state_n, duration_s, right_censored, next_state_n, direction, jump, meta — and a test asserts exactly that field set);
• the nearest candidate, the motor's own stator-exchange or switching trajectory, is scored as a category error by MOTOR-STACK-AIF-SCOPE-RULING.md §3: treating a hidden-state trajectory as a policy asserts that the organism performs the inference, which is the very thing at issue;
• the experimenter's protocol is an experimenter policy, not the motor's. Scoring it says something about the laboratory, not about the motor.

5.4 Verdict of this probe

A discriminating prediction that separates biological policy selection from the stimulus-conditioned incumbents cannot currently be named at the single-motor level. The only axis that would separate them — the epistemic term — has no instantiable action channel in this system with the recorded or plausibly recordable observables.

Therefore, and stated plainly because it is the most valuable output of this probe:

G is currently unfalsifiable in this system and must stay fenced. G_motor biological policy selection remains DESIGN_ONLY_UNTIL_INTERVENTION_OR_TRANSFER, and the intervention arms in §3, even if all five were executed perfectly, would not by themselves make it falsifiable. They would make responsiveness and stimulus-conditioned prediction testable — which is worth doing, and is P5- and P6-relevant — but responsiveness is not policy selection.

No discriminator was manufactured to fill this section. The absence is the finding.

5.5 What would change this verdict

Recorded here so a future agent can check whether the fence is still correct, not as a plan:

1. an observable, separately recorded action channel at the motor or cell level in which an uncertainty-reducing action is pragmatically costly, so the two expected-free-energy terms dissociate;
2. or transfer of the question to a level where an action channel already exists and is recorded with onset times (chemotactic run-and-tumble decisions of a whole cell in a controlled gradient are the obvious candidate) — noting that this changes the experimental unit from motor to cell, which invalidates every motor-level variance figure in §4 and requires its own split, its own frozen cohort, and its own pre-registration;
3. and, in either case, a pre-registered stimulus-conditioned M2/M4/M7 family as the adversary, scored on the identical split with the identical rule, before the policy account is scored.

Until at least (1) or (2) exists in the repository as data with recorded onset times, adding expected_free_energy to the package would create unfalsifiable scaffolding. The tests that prevent it are correct and must not be weakened.

6. What this document does NOT do

• It does not add, propose adding, or design an expected_free_energy symbol, a policy posterior, a policy prior, or an action-set field. The package fence stands unchanged.
• It does not move P5, P6, or any other level. P5 intervention stays NOT_ESTABLISHED; P4 remains the first unsatisfied level; P8 FULL_PARITY = false is unchanged.
• It does not license reading held-out nextStateN, direction, or jump. §3.5 explicitly routes the mark-channel requirement to a new dataset because D5/D6 quarantined the current one.
• It does not report a required motor count. Every such cell reads NOT_MEASURED, and the §4.3 table is labelled DESIGN_ONLY and non-evidential.
• It does not claim that any perturbation in §3 would establish that the motor performs inference. On the evidence and the argument in §5, none of them would.

7. Provenance of every number in this document

Everything not in this table that would have been a number is written NOT_MEASURED, NOT_COMPUTED, or DESIGN_ONLY.

NEXT_ACT = Run python hierarchical-aif/src/motor_stack_aif/claim_guard.py hierarchical-aif/protocols/G-SIDE-INTERVENTION-DISCRIMINATOR.md, then — without touching the two runs in flight — record in hierarchical-aif/docs/BIOLOGICAL-PARITY-RECEIPT-MAP.md LANE D that the G-side discriminator search returned NO instantiable discriminator at the single-motor level, so the fence is retained on argument and not only on missing data.