RFC-010: Autonomous Intelligence Governance Protocol (AIGP) — 16. Post-Hoc Evaluation Loop (RFC-032)

AIGP Specification › RFC-010: Autonomous Intelligence Governance Protocol (AIGP) › 16. Post-Hoc Evaluation Loop (RFC-032)

← 15. Agentic Governance Extension · Section index

16. Post-Hoc Evaluation Loop (RFC-032)

This section introduces the post-hoc evaluation loop — the mechanism that bridges the non-causal gap between governance and quality. Governance is a necessary precondition for quality, but governance alone does not produce quality. The post-hoc loop closes this gap by comparing outcomes against pre-declared expectations.

16.1 New Protocol Messages

Message	Temporal Position	Purpose
ANTICIPATE	Pre-execution (after PLAN_SUBMIT, before first TOOL_REQUEST)	Declare measurable expectations for what governed execution should produce
VERIFY	Post-execution (after final RECORD)	Compare actual outcome against declared anticipation; emit verdict
SESSION_COMPLETE	Post-execution	Signal execution is done; trigger VERIFY pipeline
QUALITY_POSTURE	Periodic (consumer tier)	Translate aggregated eval signals into governance-consumable risk posture

16.2 Message Lifecycle

PLAN_SUBMIT → APPROVED → ANTICIPATE → [seal on first execution event]
→ TOOL_REQUEST × N → STEP_COMPLETE × N → RECORD
→ SESSION_COMPLETE → VERIFY → QUALITY_POSTURE → Governance Adaptation

16.3 ANTICIPATE Message

{
  "protocol_version": "2.0",
  "message_type": "ANTICIPATE",
  "agent_id": "agent-gandalf-spell-executor-001",
  "app_id": "APP_PRJ",
  "anticipation_id": "ant-20260624-001",
  "plan_id": "plan-abc123",
  "session_id": "ses-xyz789",
  "expectations": {
    "outcome_criteria": [...],
    "trajectory_criteria": [...],
    "anti_criteria": [...]
  },
  "grading_policy": {
    "grader_type": "CODE",
    "passing_threshold": 0.7,
    "weighted_scoring": true,
    "fail_on_any_anti_criteria": true
  }
}

ANTICIPATE is sealed (immutable) once the first execution event arrives. This prevents retroactive redefinition of success criteria.

16.4 VERIFY Verdicts

Verdict	Meaning	Governance Action
MATCH	All criteria passed	LOG, optionally REWARD
PARTIAL_MATCH	Score above threshold, some criteria failed	LOG, optionally ALERT
MISMATCH	Score below threshold	ALERT, NARROW_SCOPE
VIOLATION	Anti-criterion triggered	CIRCUIT_BREAK, REVOKE_SCOPE
INDETERMINATE	Grader could not evaluate	ALERT, RETRY_VERIFY

16.5 Quality Moderator (Consumer Tier)

The Quality Moderator is the consumer tier deliverable that translates VERIFY verdicts into governance-compatible risk posture:

Risk Posture	Signal	Governance Effect
EXCELLENT	match_rate > 0.90	Scope expansion eligible
ACCEPTABLE	match_rate > 0.70	No change
DEGRADED	match_rate 0.50–0.70	Tighten monitoring
FAILING	match_rate < 0.50	Narrow scope
CRITICAL	violation_rate > 0.10	Circuit break

16.6 Governance Stage 27

stage_id	stage_name	AIGP Role
27	post_hoc_verification	Compare outcome to anticipation, emit verdict, trigger governance action

16.7 Reference

Full specification: RFC-032: Post-Hoc Evaluation Loop

---

© 2024-2026 Kanjani AI Research & Causum. All rights reserved. This specification is proprietary and subject to the terms in LICENSE.md. Unauthorized modification, redistribution, or commercial resale is prohibited.

---

← 15. Agentic Governance Extension · Section index