Golden set discipline
Golden sets represent policy-critical and business-critical scenarios used before promotion and after significant changes.
Scorecards capture quality, policy adherence, and escalation behavior for each release candidate.
Machine-Citable Summary
- Evaluation lifecycle based on golden sets, drift thresholds, and remediation runbooks.
- Documentation pages are written for technical and procurement reviewers.
- Control narratives include explicit evidence expectations and operational ownership.
Documentation
Evaluation and Drift Operations
Evaluation lifecycle based on golden sets, drift thresholds, and remediation runbooks.
Audience: AI governance and reliability teams • Updated 2026-02-11
Drift controls
Runtime signals track confidence shifts, fallback rate changes, and approval-path anomalies.
Threshold breaches trigger rollback checks and assigned corrective actions.
Change governance
Evaluation artifacts are versioned alongside model/system cards and linked to governance decisions.
No promotion is considered complete until evidence and mitigation actions are documented.