UNI Universal Natural Intelligence

Wiki · Hierarchical Active Inference

Data-Channel Spend Ledger

Hierarchical Active Inference · hierarchical-aif/docs/DATA-CHANNEL-SPEND-LEDGER.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.

Opened: 2026-07-21 (gate H-AIF-G2, after D5) · Append-only. Dataset: experiments/data/wadhwa-2022-events.json — 1349 events, 99 motors, single study (Wadhwa 2022). Split: holdout iff sha256_mod5(motorId) == 0, else train.

A channel is a field × split. Once a held-out channel is read by any analysis, it is spent: no later claim on it can be PROSPECTIVE. Reading is irreversible. This ledger records what has been spent, by whom, and what claims each channel can still support.

Why this exists: D5. A read-only audit track was asked for "empirical marginals of direction/jump per state" and, in answering, read the held-out mark channel. Nothing was written and no file was modified — and the channel was destroyed anyway. Read-only is not consequence-free.

1. Split-boundary vocabulary

Every new analysis must declare exactly one:

TRAIN_ONLY
HOLDOUT_ALREADY_SPENT_DURATION_ONLY
HOLDOUT_ALREADY_SPENT_DIRECTION
HOLDOUT_MARK_CHANNEL_BURNED_RETROSPECTIVE_ONLY
INDEPENDENT_TRANSFER_REQUIRED
PROSPECTIVE_NEW_DATA_ONLY
NO_DATA_ACCESS_NEEDED

2. Ledger

field split first_read_by first_read_artifact prospective_status claim_allowed claim_forbidden notes
durationS holdout B2/B3 held-out scoring (pre-existing, authorised) audits/phase-b/b3-model-competition-result.json SPENT — legitimately, under a committed prospective record (e5b4969 prediction-only commit preceded 9f24848 result) Held-out predictive scoring of duration models; the retained adverse M2-over-M3 result New "prospective" duration claims on this holdout without a fresh prediction commit This spend was correct: prediction was committed before the result. Duration is the channel B3/B4 are built on.
durationS train B2/B3 fitting b3-model-competition-runner.py n/a (training) Model fitting Using training fit as held-out evidence —
stateN holdout cohort eligibility + per-state scale normalisation b3-model-competition-result.json (summary.scale_N, holdoutMotorEventCounts) SPENT — used to define cohorts [1..8] / [0..8], to normalise _y, and in B4C07 eligibility reproduction Cohort membership, state-stratified scoring, eligibility statements Treating a new state-conditional analysis on this holdout as prospective Per-state holdout counts are published in the B4 result (B4C07.perStateHoldout).
rightCensored holdout cohort construction (exclusion flag) b3-model-competition-result.json; B4C05 censoring sensitivity SPENT Censoring-treatment sensitivity; the 18-uncensored-events-at-state-0 exclusion Prospective censoring claims on this holdout B3 competition EXCLUDES censored holdout events by frozen rule.
direction holdout competing-risks first-passage likelihood (pre-existing) lib/source-first-passage.js:59-65; scripts/run-science-gates.py:104-115,181-197 SPENT — before D5, by a committed gate Cause-specific (on/off) first-passage scoring Claiming the binary direction channel is unspent I initially and wrongly assumed the whole mark was unused. direction was already consumed.
nextStateN holdout UltraCode Track C, 2026-07-21 (D5) hierarchical-aif/ledgers/HIERARCHICAL-AIF-DEFECT-LEDGER.md (D5); Track C body BURNED — improperly, with no prospective record Retrospective / exploratory analysis only Any PROSPECTIVE mark-process claim on Wadhwa-2022 Cannot be repaired in this dataset: one study, no second holdout.
jump holdout UltraCode Track C, 2026-07-21 (D5) same as above BURNED — improperly Retrospective / exploratory only Any PROSPECTIVE mark-magnitude claim on Wadhwa-2022 Joint (N, N') transition structure also exposed.
nextStateN, jump train Track C; also D6 verification Track C body; defect ledger D6 Training data — free to use Fitting a mark model on training Presenting a training fit as held-out evidence Training-side use remains fully available.
motorId both split derivation sha256_mod5(motorId) b3-model-competition-runner.py Structural, not an observable Defining the split; motor-cluster resampling Using motor identity as a predictor The split itself must never change.
eventId, enteredAtS, eventAtS, splitRemainder, partition both bookkeeping — Not scientific observables Provenance/ordering Use as covariates —

3. Consequences now in force

  1. Mark-process models on Wadhwa-2022 are RETROSPECTIVE_EXPLORATORY_ON_THIS_DATASET. They may still be built and reported — labelled as exploratory — but they cannot move P3 as prospective held-out evidence.
  2. Prospective mark-process mechanism evidence now requires INDEPENDENT_TRANSFER_REQUIRED — a new dataset with a new prospective split. This folds the mark question into the P4/P7 requirement the project already carries.
  3. durationS on holdout remains legitimately spent, so B3/B4 duration work is unaffected and the corrected C02/C10/C11/C01 runs are unaffected — none of them reads the mark.

4. Access protocol required before any new data-touching task

Any task that might touch held-out fields must first write hierarchical-aif/protocols/<TASK>-DATA-ACCESS-PROTOCOL.md declaring: fields to read, split to read, fields forbidden, claim boundary, prospective/exploratory status, falsifier, stop condition.

The Track C brief did none of this. That omission is the root cause of D5, and the requirement exists so it cannot recur silently.

sha256 fbf9dba7ac46a891 — of the original file, so what was ingested stays checkable.