A city GenAIOps runbook is cutting over to a new fine-tuned checkpoint that cleared eval and canary. What should happen to the old fine-tuned deployment?

Retiring the prior fine-tuned deployment after the new checkpoint clears eval and canary cutover is what should happen to the old fine-tuned deployment in a city GenAIOps runbook. Once the new checkpoint is proven and serving, retaining the prior fine-tuned deployment in service with 5% traffic for extended canary comparison wastes cost and creates confusion about which version is authoritative, since comparison belongs in canary and eval rather than in live production. Controlled retirement completes the lifecycle after successful cutover.

Retaining the prior fine-tuned deployment in service with 5% traffic for extended canary comparison never finishes cutover and splits traffic that should go entirely to the proven checkpoint. Retaining the prior fine-tuned deployment endpoint live with routing for post-cutover audit review keeps a superseded version serving when audit review should use the retained registered artifact and logs, wasting cost and delaying clean retirement. Reassigning the prior fine-tuned deployment as the staging endpoint for the next fine-tune cycle bypasses proper evaluation and risks inserting an unverified model into the pipeline after a successful canary.

Exam caveat: retirement follows successful eval and canary—not keeping the prior endpoint live with routing for post-cutover audit review or reassigning the prior fine-tuned deployment as the next staging endpoint. Retain artifacts for rollback window even after traffic moves off the old deployment.

Operational check: after canary metrics clear, shift 100% traffic to the new checkpoint, drain and retire the prior deployment per the runbook, keep the prior registered artifact available for emergency rollback for the defined retention window, and update the civic model inventory.