@agenticprimitives/witness

Evidentiary dispute attestation — the neutral function that answers, from artifacts alone: was this action inside the delegated relationship at the time it happened — and if not, whose check failed?

Spec: specs/307-witness.md · Consumes: verification receipts (spec 303) + relationship records (spec 306).

The walk

import { evaluateDispute, mintWitnessAttestation } from '@agenticprimitives/witness';

const evaluation = await evaluateDispute({
  record,            // the spec-306 hash-linked evidence chain
  receipt,           // the disputed spec-303 public receipt
  detail,            // optional — released to the witness via releaseMode 'witness_escrow'
  verifyReceiptSignature,
});
// → { outcome: 'in_scope' | 'out_of_scope' | 'revoked_before_execution'
//              | 'stale_status_used' | 'insufficient_evidence', ... }

const attestation = await mintWitnessAttestation(
  { witness: witnessSa, relationshipId, evaluation, receipt, checkedDetail: !!detail, attestedAt },
  { sign: witnessSigner },
);

The walk is total and ordered (evidence integrity → record membership → revocation timeline → scope), so two witnesses given the same artifacts produce the same outcome and the same attestation hash. Anything unverifiable — a broken chain, an invalid receipt, a receipt that isn't an entry of the record, a detail that doesn't open — is insufficient_evidence, never a guess.

stale_status_used is the load-bearing outcome: a revocation entry precedes the receipt in the relationship record, yet the verifier allowed on revocation: 'active' — the verifier relied on stale status, which is exactly the liability trigger the article's procurement checklist assigns.

Status

w1-evaluate-attest — the walk + attestation mint/verify, 8-test suite (one per outcome, determinism, tamper, no-identity-leak). W2: attestation anchoring + the witness-escrow DecryptGrant release path. W3: the G-12 liability schema over these outcomes.