Scope & risk mapping

• Map risks to tests: toxicity, hallucination, privacy leakage, bias, jailbreak, tool abuse, data exfiltration.

Safety metrics

• Refusal accuracy, harmful content rate, prompt-injection success, PII leakage rate, groundedness error.

Robustness testing

• Stress tests; adversarial prompts; distribution shift; ablation of guardrails; fail-safe behaviours.

Fairness & cohort tests

• Metrics per cohort; minimum viable sample sizes; statistical significance; remediation gates.

RAG/groundedness tests

• Retrieval precision/recall; citation correctness; hallucination with/without context; source freshness.

Tool-use & function safety

• Action approval flows; rate limiting; argument validation; guard policies for destructive tools.

EvalOps & repeatability

• Seed control; dataset versioning; environment capture; CI gates; golden sets; flaky test detection.

Thresholds & waivers

• Approval matrix; waiver policy with expiry; residual risk statements; compensating controls.

Sign-off & packaging

• Evaluation report; model card links; reviewer signatures; release artifact checksums.

Post-release validation

• Canary rollout; shadow tests; rollback criteria; telemetry validation; incident hooks.

Records & dashboards

• Scorecards over time; cohort drift; guardrail health; links to incidents/CAPA and PMM.

Evaluation checklist

• Coverage mapped; thresholds set; repeatable runs; sign-off complete; dashboards live.

© Zen AI Governance UK Ltd • Regulatory Knowledge • v1 05 Nov 2025 • This page is general guidance, not legal advice.