ADMT — Automated Decision-Making Transparency
v1.20.10 “Proof” — a read-only assembly for the question “why did this become memory, by what path, from what source?” Each decision that turns a proposal into memory is human-approved (v1.14 Gate); this kit surfaces the decision’s own recorded trail.
The record
scripts/admt-kit.sh <chunk-id> [--out DIR]
requires a running server + a read-token (default ~/.config/brain-server/auth-token,
override BRAIN_TOKEN_FILE). It calls existing, already-audited endpoints
and assembles them verbatim — it fabricates nothing:
| Field | Source | Meaning |
|---|---|---|
decision_evidence | GET /get/{id} | the chunk’s origin (v1.18.2 provenance), owner, title, evidence span |
decision_path | GET /audit?kind=reconcile | the proposal-gate trail — proposal:{id} approve/reject rows |
The audit rows come from the tamper-evident hash chain (verified by
/audit/verify); the /health integrity.chain_ok posture (v1.20.10) says
whether that chain currently verifies. Together: who approved it, from what
source, against an unbroken chain.
Why this is trustworthy (not a re-derivation)
- No new computation. Every field is copied from an already-served JSON response; the record can be diffed against the live endpoints at any time.
- No new authority. It inherits the server’s existing integrity posture — it cannot vouch for a chain the server itself reports as broken.
- PII-safe. Proposals were PII-redacted at write time (v1.20.1);
/get/{id}revealsowneronly through the operator’s own read token. The record carries provenance + gate rows, never secret content.
Honest ceiling
- The audit rows are records of the decision, not a causal/score model of why the reviewer approved. Explainability beyond the gate trail (e.g. the exact scoring signals that ranked a proposal) is a separate, future surface.
chain_okreflects the integrity watcher’s last full verify (default 60s), not a live per-request scan.