Subagent Brief Template

Paste §1 verbatim into every subagent brief that touches this repository. Defect D5 happened because a brief omitted it.

1. D5 HELD-OUT DATA FIREWALL — paste verbatim

=========================  D5 HELD-OUT DATA FIREWALL  =========================
A prior read-only audit track was asked for "empirical marginals of direction/jump per state" and,
in answering, READ THE HELD-OUT MARK CHANNEL. That permanently burned it: no future mark-process
claim on Wadhwa-2022 can ever be labelled PROSPECTIVE. Reading held-out data is IRREVERSIBLE.
Read-only is NOT consequence-free.

BINDING RULES FOR YOU:
1. You may read CODE freely (any .py/.js/.ts/.md/.json under audits/, scripts/, lib/, app/,
   hierarchical-aif/).
2. You may read RECORDED RESULT ARTIFACTS freely (audits/phase-b/*result*.json) - those are
   already-spent, published numbers.
3. You MUST NOT compute any new statistic, table, marginal, correlation, plot, or model fit over
   experiments/data/wadhwa-2022-events.json rows whose partition == "holdout".
4. If a verification would require touching holdout rows, DO NOT DO IT. Instead report:
   "NOT_CHECKED - would require holdout access; requires prospective record first."
   That is a CORRECT and VALUABLE answer here. It is never a failure.
5. TRAIN-partition reads are permitted where genuinely necessary, but say so explicitly and
   declare the split boundary of anything you compute.
Every claim you emit must carry a split-boundary label from:
   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. Contract block — paste verbatim

CONTRACT:
- audits/phase-c/**, audits/phase-d/**, audits/phase-b/** are FROZEN read-only. Never edit.
- Existing parity ladder P0..P8 in CLAUDE.md is authoritative and must NOT be redefined:
  P0 computational integrity, P1 equation/implementation, P2 observational, P3 held-out predictive,
  P4 transfer, P5 intervention, P6 structural/mechanistic, P7 independent replication, P8 full verdict.
  P8 is conjunctive and currently false.
- FORBIDDEN language: "biological parity achieved", "active inference demonstrated", "flagellum
  solved", "general intelligence", "awareness achieved", "M2 is the UNI model", "G proves motor
  agency". A design document is NOT evidence and raises NO P-level.
- M2_LOGNORMAL is an ADVERSARIAL BASELINE. It currently out-predicts the two-timescale mixture M3
  on held-out data - a retained adverse result.
- Experimental unit is the MOTOR: 80 training / 19 holdout motors. Bootstraps resample MOTORS.
  Underpowered is NOT equivalence. Point estimates are never verdicts; CI-bound only.
- No floor anywhere: a non-finite log density HALTS.

3. Verification status vocabulary — require it

CHECKED_AGAINST_CODE      - you opened the file and confirmed it
CHECKED_AGAINST_RESULTS   - confirmed against a recorded result artifact
REPORTED_BY_TRACK         - asserted by a prior track; you did not verify
NOT_CHECKED               - not verified (say why; holdout-blocked counts here)
CONTRADICTED              - you checked and the prior claim is WRONG (say what is actually true)

State explicitly: "Being adversarial is the job. A CONTRADICTED finding is the most valuable output you can produce." This is what surfaced D7 and the Track D ranking error.

4. Established facts to hand the agent

- F, G, policy layer, policy priors E, sensory precision Pi, up/down messages are ABSENT from the
  science pipeline. All 9 B3 models are MLE density fits; none computes F or G.
- M6_SEMI_MARKOV is NOT semi-Markov: 8 independent per-state Weibull fits, no transition kernel.
  nextStateN/direction/jump exist in the data and are never read by B3.
- Verified defects D1-D7 - see ledgers/HIERARCHICAL-AIF-DEFECT-CLOSURE-LEDGER.md.
- The recorded `width` field is the PERCENTILE companion width, not the BCa/intervalUsed width (D7).
  Narrowest motor-equal contrast is M4 at 0.083461; resolution floor is ~0.042 nats.
- One study only (Wadhwa 2022). No transfer dataset. Raw MAT archive ABSENT.

5. Builder-side rules when spawning agents

1. Never ask for "marginals", "distributions", "summary statistics", or "what does the data look like" without naming the partition. That phrasing is what caused D5.
2. If a task might touch held-out fields, write protocols/<TASK>-DATA-ACCESS-PROTOCOL.md first: fields to read, split, forbidden fields, claim boundary, prospective status, falsifier, stop condition.
3. Verify a subagent's consequential claims yourself before repeating them. Track D's headline survived; six of its receipts did not.
4. When two independent computations disagree, investigate the disagreement — do not pick one. That is how D7 was found.
5. Keep spawn counts low and briefs precise. A vague brief is a live hazard, not merely inefficient.