A city auditor wants a checkpoint before production inference goes live after validation. Which lifecycle stage provides that gate?

The city auditor’s checkpoint before production inference goes live after validation is the validation and testing stage: evaluate the model against held-out sets or scenario tests, meet documented acceptance criteria, then promote the inference service. That gate sits between successful training or experimentation and city-wide production serving. Monitoring dashboard setup before production deployment of the inference service is preparatory observability tooling, not a pre-production gate; it builds dashboards and alerts to use later but does not evaluate the model against acceptance criteria to approve go-live. Retraining trigger design before production deployment of the inference service plans future maintenance thresholds, not the validation gate that must pass before go-live; it defines when to retrain later but does not test the current model for promotion. Data collection planning before production deployment of the inference service arranges inputs and collection steps, not the pass/fail checkpoint immediately before production inference goes live after validation; it prepares data rather than evaluating the model for release. Exam caveat: validation does not eliminate the need for monitoring after go-live; it is the pre-production control, while monitoring watches drift, incidents, and SLO breaches afterward. Operational check: require documented test results, approval sign-off, and a pinned model artifact before the inference endpoint is exposed—and keep notebooks from being the sole promotion path into production.