Skip to content
AIGP

RFC-033: Quantitative Outcome Evaluation Model — 4. Temporal Model

AIGP Specification › RFC-033: Quantitative Outcome Evaluation Model › 4. Temporal Model

← 3. Outcome Score Model · Section index · 5. Protocol Integration →

4. Temporal Model

4.1 Session Score

A session (multiple invocations with shared context) produces a session-level score:

OS_session = mean(OS_i)  for all invocations i in session

4.2 Rolling Baseline

Each (app_id, use_case) pair maintains a rolling baseline:

Baseline(t) = EMA(OS, α=0.1)  // Exponential moving average, 10% weight to new

4.3 Drift Detection

Score drift is flagged when:

|OS_current - Baseline| > 2σ  (where σ = rolling standard deviation)

This triggers a GOVERNANCE_ALERT with drift details.

4.4 Regression Rule

IF OS_week < OS_prev_week × 0.85 THEN
    emit GOVERNANCE_ALERT(type=QUALITY_REGRESSION, severity=HIGH)
    action: restrict_to_verified_prompts

← 3. Outcome Score Model · Section index · 5. Protocol Integration →