RFC-033: Quantitative Outcome Evaluation Model — 2. Motivation
AIGP Specification › RFC-033: Quantitative Outcome Evaluation Model › 2. Motivation
← 1. Abstract · Section index · 3. Outcome Score Model →
2. Motivation
2.1 The Gap
AIGP currently proves:
- Authorization — the action was permitted (scope envelope)
- Compliance — jurisdictional rules were satisfied (IHL)
- Observability — the action was recorded (TRACE, RECORD, evidence)
- Containment — budgets and boundaries were respected
AIGP does not currently produce a number that answers: “How good was this outcome?”
2.2 Why Quantitative?
Qualitative verdicts (MATCH/MISMATCH/VIOLATION from RFC-032) enable binary governance actions. But production systems require:
- Granular optimization — a 0.72 vs 0.88 distinction informs model selection
- Trend detection — score degradation over time signals model drift
- Autonomy calibration — earned autonomy scales with demonstrated quality
- Cost-quality tradeoffs — cheaper models acceptable when score > threshold
- SLA enforcement — contractual quality guarantees need measurable proof
2.3 Design Principle
The score must be:
- Computable — derivable from data already in RECORD/TRACE (no new instrumentation)
- Decomposable — each dimension scored independently, composite aggregated
- Comparable — scores across apps, agents, and time are on the same scale
- Actionable — governance rules can consume the score directly