5.1.5 · Lesson 5 of 14
Gate Deployments with Quality Thresholds
Enforce automated quality gates so prompt or model changes only promote when evaluation thresholds pass.
WHAT
Enforce automated quality gates in deployment pipelines for GenAI changes.
HOW
- Run regression evals on new prompt/model commits.
- Fail the gate on low faithfulness or
stopReason: max_tokenstruncation spikes. - Promote only when scores meet thresholds; otherwise rollback.
EXAM TRAP
Model Evaluation is an assessment tool — it does not automatically edit failing prompts.
Key points
- Quality gates in CI/CD
- Faithfulness thresholds (e.g., ≥ 0.85)
- Fail build on truncation
- Canary before full rollout
- Rollback on gate failure
Exam focus
Bedrock Model Evaluation assesses quality — it does not automatically fix prompts. Gates must block bad changes.