RFC-032: Post-Hoc Evaluation Loop — 16. Quality Moderator — Consumer Tier Deliverable

AIGP Specification › RFC-032: Post-Hoc Evaluation Loop › 16. Quality Moderator — Consumer Tier Deliverable

← 15. Implementation Priority · Section index

16. Quality Moderator — Consumer Tier Deliverable

16.1 Position in the AIGP Ecosystem

The ANTICIPATE/VERIFY protocol mechanism produces quantitative evaluation signals: pass rates, weighted scores, trajectory quality metrics, verdict distributions. These signals are risk indicators — and risk is governance’s native language.

However, translating raw eval signals into governance-consumable risk posture requires a dedicated product layer. This is the Quality Moderator — the fourth consumer tier deliverable alongside the Evidence Analyzer, Model Moderator, and Prompt Moderator:

Deliverable	Temporal Position	Input	Output
Prompt Moderator	Before (pre-execution)	Prompt content	Safe/unsafe classification
Model Moderator	During (runtime)	Model behavior signals	Compliance/violation classification
Evidence Analyzer	After (post-hoc forensics)	RECORD/TRACE evidence	Findings with dimension attribution
Quality Moderator	After (post-hoc evaluation)	VERIFY verdicts + eval scores	Governance-consumable risk posture

graph LR
    subgraph "Evaluation Signals"
        VS[VERIFY Scores]
        PR[Pass Rates]
        TQ[Trajectory Quality]
        DR[Drift Metrics]
        FB[FEEDBACK Signals]
    end
    subgraph "Quality Moderator"
        QM[Score → Risk Translation]
    end
    subgraph "Governance Engine"
        GE[ALLOW / DENY / ESCALATE]
        SE[Scope Envelope Adaptation]
        CB[Circuit Breaker]
    end
    VS --> QM
    PR --> QM
    TQ --> QM
    DR --> QM
    FB --> QM
    QM -->|risk posture| GE
    QM -->|adaptation signal| SE
    QM -->|halt signal| CB

16.2 What the Quality Moderator Does

The Quality Moderator does NOT:

Enforce (that’s the Governance Engine’s job)
Record evidence (that’s the Evidence Analyzer’s job)
Grade individual executions (that’s the VERIFY grader’s job)

The Quality Moderator DOES:

Aggregate VERIFY verdicts across time, agents, and workflows into quality posture scores
Translate quantitative eval signals (pass rates, trajectory scores, drift metrics) into governance-compatible risk classifications
Emit risk posture changes that the Governance Engine consumes to adapt scope, budgets, and enforcement
Correlate quality signals with FEEDBACK signals to calibrate what “good” means empirically
Detect systemic quality degradation across agent populations (not just individual drift)

16.3 Risk Posture Model

The Quality Moderator translates eval scores into a 5-tier risk posture:

Risk Posture	Quality Signal	Governance Implication
`EXCELLENT`	match_rate > 0.90, zero violations, trajectory scores > 0.85	Eligible for scope expansion, higher autonomy
`ACCEPTABLE`	match_rate > 0.70, violation_rate < 0.02	Standard governance posture — no change
`DEGRADED`	match_rate 0.50–0.70 OR trajectory declining	Tighten monitoring, reduce budget headroom
`FAILING`	match_rate < 0.50 OR violation_rate > 0.05	Narrow scope, escalation required for high-risk actions
`CRITICAL`	violation_rate > 0.10 OR systemic mismatch pattern	Circuit break, revoke scope, human intervention required

16.4 Quality Moderator Output: QUALITY_POSTURE Message

{
  "protocol_version": "4.1",
  "message_type": "QUALITY_POSTURE",
  "agent_id": "agent-gandalf-spell-executor-001",
  "app_id": "APP_PRJ",
  "posture": "DEGRADED",
  "computed_at": "2026-06-24T12:00:00Z",
  "window": {
    "executions_evaluated": 30,
    "period_days": 7
  },
  "signals": {
    "match_rate": 0.63,
    "partial_match_rate": 0.27,
    "mismatch_rate": 0.07,
    "violation_rate": 0.03,
    "avg_trajectory_score": 0.58,
    "avg_outcome_score": 0.79,
    "feedback_correlation": 0.72,
    "drift_detected": true
  },
  "governance_recommendation": {
    "action": "NARROW_SCOPE",
    "reason": "Trajectory quality declining (0.58 < baseline 0.72); match_rate below ACCEPTABLE threshold",
    "suggested_constraints": {
      "reduce_max_actions_by": 0.20,
      "reduce_max_tokens_by": 0.15,
      "add_approval_gate_pattern": "iam_*"
    }
  }
}

16.5 Relationship to Existing Moderators

graph TB
    subgraph "Before Execution"
        PM[Prompt Moderator<br/>Content safety check]
    end
    subgraph "During Execution"
        MM[Model Moderator<br/>Runtime compliance]
    end
    subgraph "After Execution"
        EA[Evidence Analyzer<br/>Forensic findings]
        QMod[Quality Moderator<br/>Eval → Risk translation]
    end
    subgraph "Governance Engine"
        GE[Scope + Budget + Circuit Breaker]
    end
    PM -->|safe/unsafe| GE
    MM -->|compliant/violation| GE
    EA -->|findings + severity| GE
    QMod -->|risk posture| GE
    EA -->|findings| QMod
    QMod -->|calibration| EA

The Quality Moderator consumes findings from the Evidence Analyzer (to detect contradictions where quality criteria missed a safety issue) and feeds calibration back (to evolve criteria via the second-order loop).

16.6 Why This Is a Separate Deliverable

The Quality Moderator is not part of the open-source AIGP protocol. It is a consumer tier deliverable — a product that organizations license to build governance platforms that consume AIGP data.

The distinction:

Protocol (open, emitter tier): ANTICIPATE, VERIFY, QUALITY_POSTURE messages — anyone can emit and consume
Quality Moderator (licensed, consumer tier): The algorithms, scoring models, calibration infrastructure, and adaptation logic that translate raw eval signals into actionable governance risk posture

This follows the existing AIGP business model: the protocol is free to emit; the intelligence that consumes and interprets protocol data is the licensed value.

← 15. Implementation Priority · Section index