RFC-032: Post-Hoc Evaluation Loop — 4. VERIFY Message

AIGP Specification › RFC-032: Post-Hoc Evaluation Loop › 4. VERIFY Message

← 3. ANTICIPATE Message · Section index · 5. Grader Types →

4. VERIFY Message

4.1 Structure

{
  "protocol_version": "4.1",
  "message_type": "VERIFY",
  "app_id": "APP_PRJ",
  "agent_id": "agent-gandalf-spell-executor-001",
  "verification_id": "ver-20260624-001",
  "anticipation_id": "ant-20260624-001",
  "plan_id": "plan-abc123",
  "session_id": "ses-xyz789",
  "timestamp": "2026-06-24T10:02:30Z",
  "inputs": {
    "anticipation_ref": "ant-20260624-001",
    "record_refs": ["rec-001", "rec-002"],
    "trace_ref": "trc-xyz",
    "step_complete_refs": ["step-1", "step-2"]
  },
  "results": {
    "verdict": "PARTIAL_MATCH",
    "overall_score": 0.78,
    "criterion_results": [
      {
        "criterion_id": "oc-001",
        "result": "PASS",
        "score": 1.0,
        "evidence": "User 'jane.doe' exists in IAM, member of 'engineering' group"
      },
      {
        "criterion_id": "oc-002",
        "result": "PASS",
        "score": 1.0,
        "evidence": "User has 2 policies; group baseline is 3"
      },
      {
        "criterion_id": "tc-001",
        "result": "FAIL",
        "score": 0.0,
        "evidence": "Execution used 5 tool invocations (expected <= 3)"
      },
      {
        "criterion_id": "tc-002",
        "result": "PASS",
        "score": 1.0,
        "evidence": "Zero escalations recorded"
      },
      {
        "criterion_id": "ac-001",
        "result": "PASS",
        "score": 1.0,
        "evidence": "iam_remove_user not invoked"
      }
    ]
  },
  "grader_metadata": {
    "grader_type": "CODE",
    "grader_version": "1.0.0",
    "grader_id": "aigp-builtin-deterministic",
    "execution_duration_ms": 12
  },
  "governance_action": {
    "action": "LOG",
    "reason": "Score 0.78 above passing threshold 0.7; trajectory criterion tc-001 failed but weighted score sufficient"
  },
  "d_dna_signature": "sha256:..."
}

4.2 Field Definitions

Field	Type	Required	Description
message_type	string	Yes	Always "VERIFY"
app_id	string	Yes	Application that owns the execution
agent_id	string	Yes	Agent that performed the work
verification_id	string	Yes	Unique identifier for this verification
anticipation_id	string	Yes	The ANTICIPATE declaration being verified against
plan_id	string	Conditional	Plan that was executed
session_id	string	Yes	Session within which execution occurred
timestamp	string	Yes	ISO 8601 UTC when verification was performed
inputs	object	Yes	References to evidence consumed by the grader
results	object	Yes	Verdict, scores, and per-criterion outcomes
grader_metadata	object	Yes	Identity and version of the grading system
governance_action	object	Yes	What governance does in response to the verdict
d_dna_signature	string	Yes	Cryptographic signature for tamper evidence

4.3 Verdicts

Verdict	Meaning
MATCH	All criteria passed. What happened matches what was anticipated.
PARTIAL_MATCH	Weighted score above threshold, but one or more criteria failed. Outcome is acceptable but process deviated.
MISMATCH	Weighted score below threshold. What happened does NOT match what was anticipated.
VIOLATION	One or more anti-criteria triggered. Hard governance failure regardless of score.
INDETERMINATE	Grader could not evaluate (missing evidence, grader error, timeout).

4.4 Governance Actions

The VERIFY verdict triggers a governance response. The governance server selects an action based on the verdict and the application’s configured response policy:

Verdict	Default Action	Configurable Alternatives
MATCH	LOG	LOG, REWARD (extend budget/scope for next session)
PARTIAL_MATCH	LOG	LOG, ALERT, NARROW_SCOPE
MISMATCH	ALERT	ALERT, NARROW_SCOPE, CIRCUIT_BREAK, ESCALATE
VIOLATION	CIRCUIT_BREAK	CIRCUIT_BREAK, REVOKE_SCOPE, ESCALATE
INDETERMINATE	ALERT	ALERT, RETRY_VERIFY, ESCALATE

4.5 Governance Action Definitions

Action	Effect
LOG	Record the verdict in the audit trail. No enforcement change.
ALERT	Emit a governance alert to configured notification channels.
REWARD	Increase budget allocation or extend scope for subsequent sessions.
NARROW_SCOPE	Reduce the agent’s scope envelope for subsequent sessions (fewer tools, lower budget, shorter TTL).
CIRCUIT_BREAK	Trip the agent’s circuit breaker immediately. Agent is halted.
REVOKE_SCOPE	Invalidate the agent’s current scope envelope. Agent must re-register.
ESCALATE	Route the verdict to a human reviewer for decision.
RETRY_VERIFY	Re-run verification with a different grader type (e.g., escalate from CODE to MODEL to HUMAN).

---

---

← 3. ANTICIPATE Message · Section index · 5. Grader Types →