3.4.2. The GenAIOps Lifecycle and CI/CD for AI
💡 First Principle: GenAIOps treats prompts, models, and agents as versioned, tested, deployable artifacts — the same discipline as DevOps, plus evaluation gates that block a release if quality regresses. The lifecycle loops: develop → evaluate → deploy → monitor → feed findings back into the next iteration.
Where traditional CI/CD tests code, GenAIOps adds quality evaluation as a release gate: a prompt or model change must pass groundedness/relevance/safety thresholds on an evaluation dataset before promotion. This makes AI changes reproducible and safe to ship, and ties monitoring (3.4.1) back into development — production quality signals become the next iteration's test cases.
⚠️ Exam Trap: A CI/CD pipeline that deploys a prompt/model change after only functional tests (does the code run?) misses the AI-specific gate. The correct GenAIOps answer inserts an evaluation step that checks output quality against thresholds before promotion — code-runs is necessary but not sufficient.
Reflection Question: Why does promoting a "working" prompt change straight to production after unit tests pass violate GenAIOps, and what gate does the lifecycle insert that ordinary CI/CD lacks?
