Wiki · Hierarchical Active Inference
World-Where-Full-Bio-Parity-Is-True — Receipt Tree
Hierarchical Active Inference · hierarchical-aif/docs/WORLD-WHERE-FULL-BIO-PARITY-IS-TRUE-RECEIPT-TREE.md @ b909801f3db4 (hierarchical-aif/motor-stack) — opens the published snapshot 8b4b5935bcba
How to read this page
A Plain and a Clear version of this page have not been written yet. What follows is the document itself.
Precise — the source document
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.
Full biological parity is not a current status. It is the target world defined by these receipts.
Prospectivity caveat that travels with every receipt below (D9). Two artifacts this tree leans on are pinned by defect D9 as
NOT_SATISFIEDon the commit graph: B4C10 — its prediction record and its result entered the repository in ONE commit (b9b5670), so strict ancestry is structurally unattainable and no future commit can repair it — and the F-side scoring result, whose protocol was untracked while the result already existed. Both remain valid measurements; what they cannot carry is the wordPROSPECTIVE. Where a node below cites either, read it as a retrospectively-graded receipt. By contrast B4C02 is prospective (with a recorded mid-run-commit caveat), and B4C11 (897c8ab) and B4C01 (28ce738) had their prediction records committed ahead of their results.
Type: BUILDER-SUPPORT ARTIFACT. This document moves no P-level, changes no frozen verdict, and creates no claim. It answers one question per rung: what receipt would make this rung true, and what would kill it?
Authority: the P0..P8 definitions live in CLAUDE.md and are authoritative and unchanged.
This file does not redefine them, does not create a v2, and does not run a parallel ladder. It
is consistent with, and subordinate to,
hierarchical-aif/ledgers/HIERARCHICAL-AIF-GATE-TO-EXISTING-P-LADDER-MAP.md and
hierarchical-aif/docs/BIOLOGICAL-PARITY-RECEIPT-MAP.md. Where this file and those disagree, they
win.
Two cells are executing as this file is written. B4C11_CORRECTED_FULL_RESULT.json and
B4C01_CORRECTED_FULL_RESULT.json do not exist (verified by directory listing, this session).
Their results are therefore not in evidence, and this document does not anticipate their
outcomes in either direction. Their live progress counters are telemetry and are not cited here.
Reading key
| field | meaning |
|---|---|
| current receipt | a real artifact that exists now, with hash and status — or NONE |
| missing receipt | the artifact that does not yet exist |
| next action | the smallest next step that produces the missing receipt |
| falsifier | the observation that would show the rung is not on track |
| what would make this rung true | the sufficient condition, stated so it can be checked |
| what would kill it | the result that would close this route rather than delay it |
In-repo = achievable with code, compute and analysis in this repository. Irreducibly external = requires data this project does not have; no amount of modelling closes it.
P0 — computational integrity
| field | content |
|---|---|
| current receipt | hierarchical-aif/reports/H-AIF-G1-REPO-AND-FROZEN-EVIDENCE-INTEGRITY.md — gate H-AIF-G1, status ESTABLISHED. Frozen-evidence hash baseline hierarchical-aif/reports/frozen-evidence-baseline.sha256 (250 files) re-verified this session against hierarchical-aif/reports/frozen-evidence-recheck-active-flow.sha256 (250 files): diff is empty — no drift. Determinism proven byte-identical over two full executions of the F-side scoring pipeline (F_SIDE_MOTOR_STACK_SCORING_RESULT.json, sha256 b3b12720f32c0aee3bfa456f52ae0901976e59e3b43c0f2690fa7a17386ab297). Every landed result carries a hash sidecar: B4C10_CORRECTED_FULL.sha256 = 959a00e974641eca1c0d6f3c2f7322b8c6c8411f68ca953c7e169492c4a53dde; B4C02_CORRECTED_FULL.sha256 = 0633988dbfd690c0c0d12075dba4e0d8c25ddd178125064bd01fbdaf4629e398; M4_M6_M7_PER_MOTOR_CONTRASTS.sha256 = 751a59ef45c8aecd1bdbe5fb5ef645423572a24d1691fb199d1b3b33fb8d4dbb. Ladder-map status: holds. |
| missing receipt | (a) the same hash-sidecar + determinism treatment applied to B4C11_CORRECTED_FULL_RESULT.json and B4C01_CORRECTED_FULL_RESULT.json once they exist; (b) a clean-clone execution of the CLAUDE.md required-validation block for any release claim — NOT_RUN in this working tree. |
| next action | On each corrected run's completion, hash the result, record the command/env sidecars already staged (*_COMMAND.txt, *_ENV.txt), and re-run the 250-file frozen-evidence diff. In-repo. |
| falsifier | Any diff against the 250-file frozen baseline; any non-byte-identical re-execution of a pipeline declared deterministic; any result JSON without a matching hash sidecar. |
| what would make this rung true | It is already the strongest rung: pristine remote-verified repository identity, a zero-drift frozen baseline, proven determinism, and hashed artifacts. It stays true only by continuing to hold. |
| what would kill it | A discovered edit under audits/**; a seed-dependent result that does not reproduce byte-for-byte; or a rewritten history that breaks the prediction-before-result ancestry the prospectivity checks rely on. |
P1 — equation / implementation
| field | content |
|---|---|
| current receipt | Six of the nine frozen B3 models re-derived at exact-zero oracle residual. M3, M0, M8 in F_SIDE_MOTOR_STACK_SCORING_RESULT.json (b3b12720…); M4, M6, M7 in M4_M6_M7_PER_MOTOR_CONTRASTS_RESULT.json (751a59ef…, report §3 "Oracle gate — PASS, residual exactly 0.0 on all three"). The oracle shares no code with the implementation under test. Suite state at the time of writing: 470 passed, 2 skipped, 1 xfailed (re-measured; an earlier draft carried a stale 458 from its brief without re-running). D1 (bootstrap cluster collapse) and D3 (hash-seed nondeterminism) repaired and covered by tests under hierarchical-aif/src/motor_stack_aif/. Ladder-map status: repaired in new namespace. |
| missing receipt | (a) exact-zero oracle re-derivation for the remaining three frozen models — M1_WEIBULL, M2_LOGNORMAL, M5_GAMMA — which would take oracle coverage from 6/9 to 9/9; (b) D7 closure: publish the BCa width beside the percentile width in the reporting layer, with the honest note that BCa intervals exist only for the 48 frozen B3 contrasts — for the 6 F-side contrasts and the 3 M4/M6/M7 contrasts the BCa width is NOT_COMPUTED; (c) D1 closure evidence, which is blocked on the corrected C11 run. |
| next action | Extend the existing independent-oracle harness to M1/M2/M5 using the frozen fitted parameters (M1 k=0.6250888335850175; M2 sigma=1.5783076101407152; M5 shape=0.5115799798433341) and the frozen per-state scale_N, and assert residual against the published motor-equal NLPD. In-repo, cheap. |
| falsifier | Any of M1/M2/M5 failing to re-derive within the declared tolerance; any oracle found to import the code under test (circular); any test whose success is vacuous. |
| what would make this rung true | 9/9 frozen models re-derived at declared tolerance by a code-disjoint oracle, D7 width reporting corrected without re-thresholding any verdict, and D1 closed by the corrected C11 result. |
| what would kill it | An oracle disagreement that traces to the frozen implementation rather than to the oracle — that would put the published leaderboard, not the new namespace, in question. |
P2 — observational
| field | content |
|---|---|
| current receipt | The derived cohort derived_eligible_1_to_8 (Wadhwa 2022, E. coli): 80 train motors / 793 train events; 19 holdout motors / 233 holdout events, split sha256_mod5(motorId)==0 => holdout, FROZEN; right-censored events excluded from the frozen cohort. For the duration channel this is a source-pinned recorded measurement. For the mark channel the receipt is NONE: hierarchical-aif/reports/D6-INGEST-NEXTSTATE-RANGE-CHECK-DEFECT.md records 2 holdout events with nextStateN = -1, 5 holdout events with zero training support, and 15–17% of marks leaving {1..8} — QUARANTINED, because the raw archive needed to adjudicate them is absent. Ladder-map status: limited. |
| missing receipt | The raw MAT archive re-derivation that would resolve the 2 impossible marks and re-establish the mark channel's provenance. Status BLOCKED_EXTERNAL — raw archive NOT_LOCATED (gate H-AIF-G8). |
| next action | Obtain the raw Wadhwa 2022 archive from its source and re-derive the mark fields under a predeclared quarantine/smoothing policy written before the archive is opened. Irreducibly external to obtain; in-repo once obtained. |
| falsifier | The impossible marks persist after raw re-derivation (i.e. they are in the source, not the ingest); or the smoothing constant, not the data, controls the outcome — it currently flips the sign. |
| what would make this rung true | Raw-archive provenance for every field actually used in a claim, with the mark channel either repaired or explicitly excluded from claim scope. |
| what would kill it | The raw archive proves unobtainable and the derived mark fields prove unreconstructible — which would permanently confine every mark-process statement to retrospective-only, independent of any modelling improvement. |
P3 — held-out predictive
| field | content |
|---|---|
| current receipt | Frozen B3 result audits/phase-b/b3-model-competition-result.json, sha256 5d7a0589e94de6b10f425f2d483e1e2a8f899d336aa59c335990209795e6b2bd — 9 models × 2 cohorts × 2 rules, held-out motor-equal NLPD on the frozen split, 44/44 independent oracle, determinism byte-identical. Leaderboard leader M2_LOGNORMAL 3.4093141566; reference CONTROL_CURRENT M3_TWO_TIMESCALE 3.4343333331. Corrected robustness cells landed: B4C10 (959a00e9…, M4 IDENTIFIED_ON_THIS_COHORT, U2/U3/U4 all OK) and B4C02 (0633988d…, frozen prediction GENERATOR-ROBUST_ADVERSE REFUTED; the adverse M2-over-M3 pattern is generator-specific). F-side candidate scored at held-out motor-equal 3.4326923382675303, NOT_ESTABLISHED against CONTROL_CURRENT and against every serious adversary (M2, M8, M1), RESOLVED_ABOVE only against M0 and M5; it joined the leaderboard at 5th of 10 combined and did not move the level. Ladder-map status: stands, duration-only. |
| missing receipt | A candidate whose motor-equal contrast interval lies entirely above 0 against CONTROL_CURRENT (M3) and against the strongest adversaries (M2_LOGNORMAL, M8_EMPIRICAL_KDE), with an effect above the ~0.042-nat resolution floor — the corrected motor-equal half-width, BCa half-width 0.04207043063262626 from the narrowest frozen B3 contrast (M4_MIXTURE_K3). Nothing in evidence meets this. |
| next action | Stop extending coverage of models that re-derive the incumbent and design a candidate that can beat its own tau → 0 limit. The F-side hierarchy reproduces frozen M7_HIERARCHICAL_MOTOR to 2.5e-7 nats (exp(mu)=0.659632669755436 vs M7 k=0.6596322379287862; tau=0.18372082607308418 vs 0.18372185667134974) and buys +0.000615 [−0.010881, +0.011304] over M1_WEIBULL — NOT_ESTABLISHED. In-repo to attempt; the binding limit is 19 holdout motors, which is not in-repo. |
| falsifier | An adversary's contrast interval lies entirely below 0 against the candidate. Not observed — so the F-side duration hypothesis is not falsified, and equally not supported, against the serious competitors. |
| what would make this rung true | A resolved, above-floor, CI-bound win over the control and the strongest simple baselines, on the frozen split, scored by the frozen rule, with the win surviving the misspecified-world discriminator. |
| what would kill it | The 19-motor holdout is shown to be structurally unable to resolve any effect of scientifically material size — i.e. the resolution floor exceeds the largest plausible model difference. That is a power limit; NOT_ESTABLISHED is not equivalence, and an underpowered contrast is never read as "no difference". D10 records the complementary hazard: the frozen CI rule has no minimum-effect-size guard, so a paired motor-cluster bootstrap can resolve a difference of any magnitude if its sign is consistent across motors — M7 resolved at +2.506984e-07 nats, ~168 000× below the floor, and is reported RESOLVED_ABOVE but SCIENTIFICALLY_NULL. D10 was repaired by added interpretation, never by re-thresholding. |
P4 — transfer (first unsatisfied level)
| field | content |
|---|---|
| current receipt | NONE. Status NOT_ESTABLISHED. One study, one cohort, one apparatus, one species (E. coli, Wadhwa 2022). Gate H-AIF-G8 maps here and is NOT_LOCATED / NOT_ESTABLISHED. |
| missing receipt | A second, independent dataset — different lab, different apparatus — with a split declared before the data are opened, scored by the same frozen motor-equal rule, and a CI-bound verdict on the transferred model. |
| next action | Locate and license an independent flagellar-motor dwell dataset; write the data-access protocol and the predeclared split first; commit the prediction record before any observation exists. Irreducibly external. |
| falsifier | The model's held-out ranking inverts on the second dataset, or its intervals widen to NOT_ESTABLISHED everywhere off the calibration cohort. |
| what would make this rung true | A model fitted on Wadhwa 2022 predicting a foreign cohort at CI-bound advantage over the same adversaries, without refitting. |
| what would kill it | Nothing in this repository can kill or close it. P4 cannot be closed by any amount of modelling, code, or analysis here — it requires data the project does not have. That is precisely why the movable rungs (P0, P1, P3 in-repo work, P6 per scope) are worth moving: they are what a transfer dataset would arrive to test. |
P5 — intervention
| field | content |
|---|---|
| current receipt | NONE. Status NOT_ESTABLISHED. The dataset is passive: the action set is EMPTY, and this is STRUCTURAL, not sample-size-limited. The G-side is fenced by a test asserting expected_free_energy does not exist; it is DESIGN_ONLY_UNTIL_INTERVENTION_OR_TRANSFER. Note the type boundary: F (observational free energy over beliefs) is not G (expected free energy over policies). |
| missing receipt | A wet-lab perturbation dataset satisfying all three conditions already stated in BIOLOGICAL-PARITY-RECEIPT-MAP.md: (1) a manipulated variable with recorded onset time, (2) paired pre/post observation on the same motors, (3) enough independent motors per condition for a CI-bound verdict. |
| next action | Specify the intervention protocol as DESIGN_ONLY and use G only at ORCHESTRATE level to choose the next experiment. Do not implement expected_free_energy. Irreducibly external. |
| falsifier | The intervention produces no measurable change in the modelled quantity, or the pre/post pairing is broken so the motor cannot serve as its own control. |
| what would make this rung true | A predeclared, motor-paired perturbation whose observed response falls inside the prospectively committed prediction interval. |
| what would kill it | Nothing here can close it. P5 cannot be closed by any amount of modelling, code, or analysis in this repository. Adding a policy layer to a passive dataset would manufacture unfalsifiable scaffolding, which is why the G-side fence is a test and not a convention. |
P6 — structural / mechanistic (carried per scope; no single verdict exists)
There is no single P6 verdict and none may be written. An unscoped weakening statement is a
contract violation. P6 is carried as separate scoped statements, each with its own receipt and
its own falsifier.
P6 scope |
current receipt | missing receipt | next action | falsifier | what would make it true | what would kill it |
|---|---|---|---|---|---|---|
| C11 U4 (M7 dispersion stability) | NONE — the prior receipt is WITHDRAWN. D1 cluster collapse invalidated U4_OK. Prediction record committed ahead of observation at 897c8ab. B4C11_CORRECTED_FULL_RESULT.json does not exist (verified by listing). |
the corrected full-N C11 result, hashed | let the in-flight run finish undisturbed, then hash it and write the old-vs-new status table | the corrected cell fails, or remains inconclusive at full N | a corrected full-N result meeting the frozen U4 criterion | the corrected run shows the original U4_OK was an artifact of the collapse — the honest and useful outcome, already prepaid by the withdrawal |
| Wadhwa mark-process mechanism | RETROSPECTIVE-ONLY, TRANSFER-REQUIRED — D5 (holdout mark channel burned, hierarchical-aif/reports/D5-HOLDOUT-MARK-CHANNEL-BURNED.md) and D6 (nextStateN unchecked) |
an independent prospective mark dataset with a predeclared smoothing/quarantine policy | write the policy first, obtain the dataset second | the smoothing constant controls the sign of the result (it currently does); transfer fails | prospective mark-channel prediction on foreign data | the mark fields prove unreconstructible and unobtainable — permanent retrospective-only |
| duration-only B3/B4 | UNCHANGED. B3 (5d7a0589e94…) stands; the adverse M2-over-M3 headline is retained. D1 does not touch the B3 leaderboard |
a mechanism discriminator that makes the competing models disagree out-of-sample | build on B4C02's finding — the adverse pattern is generator-specific (weibull_gamma_blend frac 0.0050; three_timescale_heavy_tail frac 0.9400; per_motor_heterogeneous_weibull frac 0.0050; gensWithM2overM3 = 1 of 3) — and design the discriminator around the generator axis that separates them |
a discriminator that cannot separate the candidates on any realisable dataset size | a design under which M2 and M3 make opposite, checkable predictions, and the observation picks one | predictive superiority is promoted to mechanism — a category error the contract forbids; predictive rank is never mechanism |
| B4C10 / M4 identifiability | LANDED (959a00e9…, full frozen N_boot=2000, 1994/2000 completed): U2 collapseFrac 0.0050150451 (fires ≥0.25), U3 span 0.4332708748 decades (fires ≥2.0), U4 omega_3 95% CI [0.030746372918754247, 0.28838105457744573]; all OK, M4 IDENTIFIED_ON_THIS_COHORT |
none for this scope; it is complete as scoped | do not transfer it — different model, different likelihood structure from C11 | a re-run at the frozen criteria fires any of U2/U3/U4 | it is already true as scoped | identifiability is not correctness and is not mechanism; reading it as either would kill its value, not its status |
| full motor-stack AIF | PENDING. Built and scored (b3b12720…, 751a59ef…) but NOT_ESTABLISHED against control and every serious adversary; at this resolution the candidate is a re-derivation of the incumbent M7, not an advance on it |
a mechanism receipt distinct from a fit receipt | see P3 next action; a model that beats its own tau → 0 limit |
the hierarchy continues to buy nothing measurable over M1_WEIBULL |
a structural component that earns an above-floor, CI-bound, discriminating win | the constrained stack is shown to be observationally indistinguishable from the incumbent on any dataset this cohort can supply |
P7 — independent replication
| field | content |
|---|---|
| current receipt | NONE. Status NOT_ESTABLISHED. External review is in progress and is not complete; review is not replication. |
| missing receipt | A second lab or an independent team re-executing the frozen pipeline from the recorded artifacts and reproducing the leaderboard and the verdicts. |
| next action | Keep the reproduction path executable from the hashed artifacts alone (command + env sidecars, frozen baseline, clean-clone validation block) so a replicator needs nothing from this session. In-repo preparation; irreducibly external execution. |
| falsifier | An independent execution produces different numbers from the same recorded artifacts, or cannot execute at all from what is published. |
| what would make this rung true | Independent re-derivation of the headline results, including the adverse M2-over-M3 result, by a party with no access to this working tree. |
| what would kill it | Nothing here can close it. P7 cannot be closed by any amount of modelling, code, or analysis in this repository. |
P8 — full verdict
| field | content |
|---|---|
| current receipt | FULL_PARITY = false. P8 is conjunctive: any single required FAIL, CONTRADICTED, NOT_ESTABLISHED, BLOCKED_EXTERNAL, NOT_RUN, or INVALID_PROVENANCE makes it false. The first unsatisfied level is P4 transfer. No P-level has been raised by any work this session; one P6 scope (C11 U4) was lowered by withdrawing defective evidence. |
| missing receipt | P0–P7 conjunctively supported, each by a named artifact, a named scope, and a carried falsifier. |
| next action | Work the movable rungs (P0 maintenance, P1 to 9/9 oracle coverage, P3 discriminating design, the C11/C01 P6 scopes) while treating P4/P5/P7 as procurement problems, not modelling problems. |
| falsifier | Any required level remaining NOT_ESTABLISHED — which is the current state. |
| what would make this rung true | Every rung above holding simultaneously, on receipts, with the adverse results still retained and still visible. P8 is reached by supplying the evidence each level requires, never by relabelling a gate. |
| what would kill it | A transfer or intervention dataset arriving and contradicting the model. That is a legitimate scientific outcome and would be reported as such; a manufactured pass would not be. |
Critical path — ordered by expected information gain
Ordered by how much uncertainty the receipt removes and by its ability to make competing models
disagree. IN-REPO and EXTERNAL are marked so the list cannot be misread as a work queue that
this repository could finish alone.
| # | receipt | why it ranks here | achievable |
|---|---|---|---|
| 1 | B4C11_CORRECTED_FULL_RESULT.json lands and is hashed |
Restores or refutes the one P6 scope that is currently withdrawn. It is the only rung movement already paid for and in flight; its prediction was committed ahead of its observation (897c8ab). Highest information per unit of remaining work. |
IN-REPO (running — do not disturb) |
| 2 | B4C01_CORRECTED_FULL_RESULT.json lands and is hashed |
Second in-flight corrected cell; prediction committed at 28ce738 with zero observations in existence, so its prospectivity is the cleanest in the program. |
IN-REPO (running — do not disturb) |
| 3 | Oracle re-derivation of M1, M2, M5 at exact-zero residual | Takes P1 code-disjoint verification from 6/9 to 9/9 frozen models. Cheap, bounded, and it audits the leader of the leaderboard (M2) — the model carrying the retained adverse headline. Auditing the leader is worth more than auditing another also-ran. |
IN-REPO |
| 4 | D7 width-reporting correction | Publishing BCa beside percentile removes a documented reporting defect that touches every verdict (intervalUsed == bca in 48/48). Must carry that BCa is NOT_COMPUTED for the 6 F-side and 3 M4/M6/M7 contrasts. Low cost, removes ambiguity from the whole evidence base. |
IN-REPO |
| 5 | A mechanism discriminator built on the B4C02 generator axis | B4C02 already showed the adverse M2-over-M3 result is generator-specific, so "it is just heavy-tailed shape" is no longer an adequate account. This is the single highest-value design item: it is the receipt most able to make competing models disagree. Any threshold introduced for the design is DESIGN_ONLY and not evidential. |
IN-REPO to design; the discriminating observation is EXTERNAL |
| 6 | Raw MAT archive re-derivation | Closes D6 (2 impossible marks, 5 zero-support holdout cells, 15–17% out-of-range marks) and is the only route to a mark-channel P2 receipt. Blocked on obtaining the archive. |
EXTERNAL to obtain |
| 7 | An independent dataset with a predeclared split | Closes the first unsatisfied level, P4. Ranked below items 1–5 only because none of this repository's work can produce it; ranked above P5/P7 because it is the binding constraint on P8 and because it also unblocks the mark-process P6 scope. |
IRREDUCIBLY EXTERNAL |
| 8 | A motor-paired intervention dataset | The only route to P5, and the only thing that converts the G-side from DESIGN_ONLY to scoreable. The action set is empty for structural reasons, so no in-repo work substitutes. |
IRREDUCIBLY EXTERNAL |
| 9 | Independent replication by a second party | The only route to P7. Preparation is in-repo (reproduction must run from hashed artifacts alone); execution is not. |
IRREDUCIBLY EXTERNAL |
The honest shape of this path: items 1–4 are finishable here and move receipts, not levels.
Item 5 is where in-repo effort has the highest scientific leverage, and it still terminates in an
observation this project does not own. Items 7–9 are the levels that decide P8, and they cannot
be closed by any amount of modelling, code, or analysis in this repository. Saying so plainly is
what makes items 1–5 worth doing: they are the state in which an external dataset would arrive and
find a model already worth testing.
Nature supplies the architecture candidate. The gate supplies the status.
Full biological parity is not a current status. It is the target world defined by these receipts.
NEXT_ACT = Run python hierarchical-aif/src/motor_stack_aif/claim_guard.py hierarchical-aif/docs/WORLD-WHERE-FULL-BIO-PARITY-IS-TRUE-RECEIPT-TREE.md, then leave the two in-flight corrected cells undisturbed and take critical-path item 3 — extend the code-disjoint oracle to M1_WEIBULL, M2_LOGNORMAL and M5_GAMMA to bring P1 oracle coverage from 6/9 to 9/9 frozen models.
sha256 5de094e06f976043 — of the original file, so what was ingested stays checkable.