A library summarizer fine-tune is ready, but production still serves the prior model. What promotion sequence is correct?

Registering the fine-tuned model, passing evaluation gates, then swapping it into the production endpoint is the correct promotion sequence when a library summarizer fine-tune is ready but production still serves the prior model. Registration creates lineage; evaluation gates prove fitness; controlled endpoint swap completes cutover without silent overwrite. That ordered lifecycle is how GenAIOps moves a library summarizer from ready artifact to resident-facing service.

Passing evaluation gates on the fine-tuned model before registering it and then swapping it into the production endpoint inverts lineage and accountability; gates must run against a registered version so scores attach to immutable lineage before any cutover. Registering the fine-tuned model, collecting patron usage metrics, then swapping it into the production endpoint skips required evaluation gates; patron usage signals cannot prove fitness before cutover, so an unvalidated summarizer would reach residents without a controlled quality clearance. Registering the fine-tuned model, passing evaluation gates, then deploying it as a new production endpoint creates a parallel serving path instead of completing cutover; a controlled swap into the existing production endpoint preserves routing, version history, and instant rollback.

Exam caveat: promotion sequence questions want register → evaluate → swap into the existing endpoint (often with canary), not evaluate-before-register, metrics-for-gates, or spinning up a separate production endpoint. Version history remains after cutover for rollback.

Operational check: register the summarizer checkpoint in the model catalog, record eval gate results, swap the production endpoint through a controlled deployment, verify hold-summary quality on a smoke set, and retain the prior endpoint version for instant rollback.