RFC-032: Post-Hoc Evaluation Loop — 3. ANTICIPATE Message

AIGP Specification › RFC-032: Post-Hoc Evaluation Loop › 3. ANTICIPATE Message

← 2. Proposed Extension: Post-Hoc Evaluation Loop · Section index · 4. VERIFY Message →

3. ANTICIPATE Message

3.1 Structure

{
  "protocol_version": "4.1",
  "message_type": "ANTICIPATE",
  "app_id": "APP_PRJ",
  "agent_id": "agent-gandalf-spell-executor-001",
  "anticipation_id": "ant-20260624-001",
  "plan_id": "plan-abc123",
  "scope_envelope_id": "scope-env-001",
  "session_id": "ses-xyz789",
  "timestamp": "2026-06-24T10:00:00Z",
  "expectations": {
    "outcome_criteria": [
      {
        "criterion_id": "oc-001",
        "description": "IAM user created successfully with correct group membership",
        "measurement": "DETERMINISTIC",
        "assertion": "user_exists(target_username) AND group_member(target_username, 'engineering')",
        "weight": 1.0
      },
      {
        "criterion_id": "oc-002",
        "description": "No permissions beyond engineering group baseline",
        "measurement": "DETERMINISTIC",
        "assertion": "policy_count(target_username) <= group_policy_count('engineering')",
        "weight": 0.8
      }
    ],
    "trajectory_criteria": [
      {
        "criterion_id": "tc-001",
        "description": "Complete within 3 tool invocations",
        "measurement": "DETERMINISTIC",
        "assertion": "step_count <= 3",
        "weight": 0.5
      },
      {
        "criterion_id": "tc-002",
        "description": "No escalation required for standard onboarding",
        "measurement": "DETERMINISTIC",
        "assertion": "escalation_count == 0",
        "weight": 0.3
      }
    ],
    "anti_criteria": [
      {
        "criterion_id": "ac-001",
        "description": "Must not invoke iam_remove_user during onboarding",
        "measurement": "DETERMINISTIC",
        "assertion": "tool_invoked('iam_remove_user') == false",
        "weight": 1.0
      }
    ]
  },
  "grading_policy": {
    "grader_type": "CODE",
    "passing_threshold": 0.7,
    "weighted_scoring": true,
    "fail_on_any_anti_criteria": true
  },
  "d_dna_signature": "sha256:..."
}

3.2 Field Definitions

Field	Type	Required	Description
message_type	string	Yes	Always "ANTICIPATE"
app_id	string	Yes	Application that owns the execution
agent_id	string	Yes	Agent that will perform the work
anticipation_id	string	Yes	Unique identifier for this anticipation declaration
plan_id	string	Conditional	Plan being anticipated (required if plan-based execution)
scope_envelope_id	string	Yes	Active scope envelope at time of anticipation
session_id	string	Yes	Session within which execution occurs (RFC-010b)
timestamp	string	Yes	ISO 8601 UTC when anticipation was declared
expectations	object	Yes	Structured expectations (see §3.3)
grading_policy	object	Yes	How expectations will be evaluated (see §3.4)
d_dna_signature	string	Yes	Cryptographic signature for tamper evidence (RFC-024)

3.3 Expectations Object

Expectations are organized into three categories:

Category	Semantics	Failure Implication
outcome_criteria	What the final state SHOULD be after execution	Outcome did not match anticipation
trajectory_criteria	How the agent SHOULD get there (process quality)	Process was suboptimal or anomalous
anti_criteria	What MUST NOT happen during execution	Hard violation — governance failure

Each criterion contains:

Field	Type	Required	Description
criterion_id	string	Yes	Unique identifier within this anticipation
description	string	Yes	Human-readable statement of the expectation
measurement	enum	Yes	DETERMINISTIC (code-verifiable), HEURISTIC (model-judged), HUMAN (requires reviewer)
assertion	string	Yes	Machine-evaluable expression or natural language description for model/human graders
weight	float	Yes	Relative importance (0.0–1.0) for weighted scoring

3.4 Grading Policy

Field	Type	Required	Description
grader_type	enum	Yes	CODE (deterministic), MODEL (LLM-as-judge), HUMAN (manual review), COMPOSITE (mixed)
passing_threshold	float	Yes	Minimum weighted score (0.0–1.0) for a PASS verdict
weighted_scoring	boolean	Yes	Whether to use criterion weights in score computation
fail_on_any_anti_criteria	boolean	Yes	Whether any anti-criterion failure produces automatic FAIL regardless of score

3.5 Immutability Requirement

Once execution begins (first TOOL_REQUEST or STEP_COMPLETE for the associated plan_id), the ANTICIPATE message is sealed. The governance server MUST:

1. Reject any attempt to modify a sealed ANTICIPATE
2. Preserve the D-DNA signature chain linking ANTICIPATE → execution events → VERIFY
3. Expose the sealed ANTICIPATE to the VERIFY grader as the authoritative standard of comparison

This prevents the retroactive redefinition of success — the most common form of evaluation circularity.

3.6 Who Declares Anticipation

Emitter	When	Use Case
Application	At plan submission time	Developer-defined quality bar for their agent’s task
Governance Authority	At scope envelope issuance	Organization-wide minimum expectations for agent type
External Evaluator	Before execution (via API)	Third-party audit, regulatory compliance testing

Multiple ANTICIPATE declarations may exist for the same execution. Each is evaluated independently in VERIFY.

---

---

← 2. Proposed Extension: Post-Hoc Evaluation Loop · Section index · 4. VERIFY Message →