A clerk's office requires clear promotion states before any endpoint deploy of registered models. Which practice should their MLOps runbook enforce?

Requiring stage labels such as Development, Staging, and Production on registered models before endpoint deploy is the correct MLOps runbook practice for a clerk’s office. Stage metadata makes promotion state explicit in the Azure ML registry so only appropriately labeled versions can be deployed to resident-facing endpoints. Clear stages separate experimentation from controlled release and support audit questions about what was live when. Requiring detailed lineage to versioned training pipelines on registered models before endpoint deploy fails because lineage tracks provenance, not promotion state; it does not enforce that only appropriately labeled versions reach resident-facing endpoints and leaves the core stage-gate requirement unmet. Requiring monitoring thresholds such as latency, errors, and drift on registered models before endpoint deploy fails because thresholds set operational health guardrails, not promotion state; they do not enforce that only appropriately labeled versions reach resident-facing endpoints and leave the core stage-gate requirement unmet. Requiring traffic policies such as canary, blue-green, and A/B on registered models before endpoint deploy fails because it concerns traffic management on live endpoints rather than the promotion state of registered models; it does not establish the Development, Staging, and Production labels required before any endpoint deploy. Exam caveat: stage labels must be backed by process—moving a tag to Production without tests is theater. Automate checks that block endpoint updates unless the model stage matches policy. Operational check: enforce Development → Staging → Production labels in the runbook, require smoke tests at Staging, allow endpoint deploy only from Production-labeled versions, and log stage transitions with approver identity for the clerk’s office.