UNI Universal Natural Intelligence

Wiki · Hierarchical Active Inference

Gap Audit — Stacked AIF Spec vs Actual Code

Hierarchical Active Inference · hierarchical-aif/docs/GAP-AUDIT-FULL-HIERARCHICAL-MOTOR-MODEL.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.

Gate: H-AIF-G6 · Method: 8-track parallel read-only audit, 134 findings · Date: 2026-07-21 Status counts: PRESENT 33 · PARTIAL 47 · ABSENT 41 · WRONG 13

Do not re-run this audit. It is recorded here so a new agent inherits it. Individual findings were spot-verified by the builder where consequential; unverified rows are marked.

Headline

The active-inference machinery is essentially absent from the science pipeline. All nine B3 models (M0–M8) are maximum-likelihood density fits. No model computes a variational or an expected free energy. The stack that exists is a hierarchy of statistical models, not a stack of inference loops.

Core components

component expected status reality
Variational free energy F E_q[ln q − ln p], belief update only ABSENT (science pipeline) No F anywhere in audits/phase-b/*.py or scripts/*.py. Runtime lib/uni-motor.js computes a scalar labelled F from an exact categorical Bayes step where KL ≡ 0 by construction — a readout, not a minimised objective (unverified by builder)
Expected free energy G risk + ambiguity, policy selection only ABSENT / WRONG Absent from the science pipeline. Runtime version reportedly adds ambiguity twice plus an undeclared effort term; docs/SCIENCE.md:83 freezes the same non-standard formula (unverified by builder)
Policy layer π action sequences with horizon ABSENT No action field exists in the dataset. The action set is structurally empty
Policy priors E, policy precision γ habits, softmax temperature ABSENT —
Sensory precision Π prediction-error weighting ABSENT "precision" in-repo means numeric tolerance
Preferences C, state priors D preferred/prior outcome distributions PARTIAL Only as implicit MLE priors
Up/down messages U = {q, ε, Π, F, …} / D = {C, D, E, γ, Π} ABSENT No inter-level message structure
Markov blanket b = {s, a} p(μ,η∣b) = p(μ∣b)p(η∣b) PARTIAL Sensory half exists as an event table; no active state a
Asynchronous per-level clocks levels update at own rates ABSENT —

Motor-stack levels

level status reality
Lmotor-5 population prior Θ PARTIAL M7 has 2 marginal parameters; no population posterior
Lmotor-4 motor identity η_m PARTIAL M7 integrates a per-motor latent but reports point estimates only — no q(η_m)
Lmotor-3 occupancy N_i PARTIAL State used for cohort membership and per-state scale normalisation; no transition kernel
Lmotor-2 kinetic/policy z_i, π_i ABSENT —
Lmotor-1 hazard/survival ABSENT in B3 / PRESENT elsewhere B3 uses plain densities and excludes censored events. A correct competing-risks hazard/survival implementation exists in lib/source-first-passage.js:59-65 and scripts/run-science-gates.py:104-116
Lmotor-0 observed blanket PARTIAL All 12 fields recorded; nextStateN/direction/jump never read by B3

The 13 WRONG findings — the consequential ones

Builder-verified:

  1. M6_SEMI_MARKOV_STATE_DEPENDENT is not semi-Markov. It is 8 independent per-state mean-one Weibull fits with no transition kernel. Blocks any reading of B3 as evidence about transition structure, and means the most obvious mechanistic alternative — that transitions carry information beyond dwell duration — was never entered into the competition.
  2. C11 U4 cluster-collapse bootstrap → defect D1.
  3. Resource costs overstated 17–29× → defect D2.
  4. hash() seeding non-deterministic → defect D3.
  5. C01 reason cites models the cell skips → defect D4.
  6. nextStateN = −1 in 2 holdout events; ingest never range-checks next_state → defect D6.
  7. width field is the percentile companion, not the BCa/intervalUsed width, in 48/48 entries → defect D7.

Reported, NOT builder-verified (carry as REPORTED_UNVERIFIED; do not repeat as fact):

  1. Runtime G(π) double-counts ambiguity and adds an undeclared effort term.
  2. Runtime F is a readout of exact Bayes; tests/model.test.mjs "free-energy identity is exact" is tautological (KL ≡ 0 by construction).
  3. Runtime agent re-executes the world's exact torque-speed constants → shared-implementation oracle on the speed/stator channels.
  4. UI declares q(o∣π)=Σ P(o∣s)q(s∣π) but implements ad-hoc linear ligand extrapolation (constants 0.08 / 0.25).
  5. Runtime models stators as a continuous ODE relaxation, not a discrete occupancy jump process.
  6. ingest-wadhwa-data.py counts 109 right-censored dwells under exclusions but does not exclude them (no continue); all 109 are in the 1349-event artifact.

What this implies for the build

  • Building the full stack would mean building F, G, policies, precision, and message passing from scratch — none is inherited.
  • G cannot be tested here: the action set is empty. See MOTOR-STACK-AIF-SCOPE-RULING.md.
  • The hazard/survival and per-motor posterior pieces are genuinely new contributions of the F-side build; the frozen pipeline has neither in its competition path.
  • Findings 8–13 concern the shipped runtime, not the science pipeline. They do not affect B3/B4 results. They remain unverified and must be checked before being acted on.

Provenance

Raw findings: workflow wf_455a1ab3-222. Verification of the Track D/E/F design documents: reports/ULTRACODE-TRACK-{D,E,F}-VERIFICATION.md — note that verification contradicted six of Track D's supporting receipts while its headline (full stack not identifiable) survived.

sha256 aa3b8b7ad54062da — of the original file, so what was ingested stays checkable.