---
name: ml-release-gate
description: Evaluate and promote a model, prompt or retrieval change through reproducible quality, safety, cost and operations gates.
---

1. Identify the candidate and baseline versions of data, code, model, prompt, tools and retrieval configuration.
2. Confirm the evaluation dataset, slices, leakage controls and expected outcome metrics.
3. Run deterministic checks, model-quality evaluations, safety tests and representative human review.
4. Compare latency, cost, reliability and operational limits with the baseline.
5. Record regressions and uncertainty; do not promote on aggregate score alone when critical slices fail.
6. Require approval for promotion, then use a staged rollout with monitoring and rollback thresholds.
7. Preserve the evaluation evidence and exact artifact identities.