Audit & Replay

Blueprint · ← Step-up · Audit & replay · Policy scenarios →

If you can answer examiner questions from PEP/PDP logs alone, you have PGAR. If you need the chat transcript, you do not.

THE CLAIM

Audit is a verdict chain with policy version, immutable, written before side effects, redacted for retention.

What examiners ask

• Which policy version decided? Every log carries pgar.payments.wire/v3, not "the system prompt from Tuesday."
• Was execution blocked until attestation? Proof STEP_UP fired and ALLOW came only after four-eyes.
• Can you replay without model logs? Subject, action, resource, context, verdict, before side effects.

Minimum audit record

Minimum audit record (JSON)

{
  "audit_id": "aud-7c3e",
  "ts": "2026-07-10T14:22:01Z",
  "subject_sub": "officer-123",
  "action": "initiate_wire",
  "resource": { "beneficiary_id": "bene-acme-441" },
  "context_redacted": { "amount": 47500, "sanctions_status": "clear" },
  "policy_version": "pgar.payments.wire/v3",
  "verdict": "STEP_UP",
  "reason_code": "wire_above_auto_approved",
  "downstream_called": false
}

RAG extension: context pack log

For retrieval paths, also log the assembled context pack (sources, scores, policy version). Answers "what did the model see on March 12?"

See RAG retrieval.

Worked example: ev-refund-agent-2026-03

Examiner opens the evidence pack and these three rows. Chat is optional color. It is not the source of truth.

Refund audit rows matching the evidence pack (JSON)

[
  {
    "audit_id": "v-10041",
    "use_case_id": "uc-refund-agent",
    "action": "issue_refund",
    "context_redacted": { "amount_aud": 80, "purpose": "duplicate_charge" },
    "policy_version": "pdp-refund-write-v4",
    "verdict": "ALLOW",
    "downstream_called": true
  },
  {
    "audit_id": "v-10058",
    "use_case_id": "uc-refund-agent",
    "action": "issue_refund",
    "context_redacted": { "amount_aud": 750, "purpose": "duplicate_charge" },
    "policy_version": "pdp-refund-write-v4",
    "verdict": "STEP_UP",
    "reason_code": "refund_above_auto_approved",
    "downstream_called": false
  },
  {
    "audit_id": "v-10062",
    "use_case_id": "uc-refund-agent",
    "action": "issue_refund",
    "context_redacted": { "amount_aud": 80, "purpose": "goodwill_unmapped" },
    "policy_version": "pdp-refund-write-v4",
    "verdict": "DENY",
    "reason_code": "purpose_mismatch",
    "downstream_called": false
  }
]

Replay v-10058 against pdp-refund-write-v4 without the attestation: still STEP_UP. After attaching att-2026-03-11-1842: ALLOW. If replay disagrees with the log, the pack is broken.

Failure classes

• Chat as audit: only conversation history, no PEP record
• Post-hoc narrative: log written after downstream success
• Missing policy version: cannot prove which rules applied
• PII in immutable log: retention violation

Replay tooling

1. Ingest audit id or (subject, time range, action)
2. Reconstruct SARAC from stored fields
3. Re-run PDP against same policy version (regression) or current (what-if)
4. Compare verdict and downstream_called flag

Release gate

• Audit completeness on ACTION cases: 100%
• Policy version present: 100%
• Verdict-before-downstream ordering: 100% on trace replay

Trace fields

audit_id, policy_version, verdict, reason_code, downstream_called, context_pack_id (RAG)

See: PEP enforcement · Eval plane Action