Feature Flags for Model Rollouts

A model swap is a product change. Treat it like one: percentage rollout, kill switch, and a documented owner.

Flag model identity, prompt version, and retrieval config separately when you can. Bundled releases hide the real culprit.

Keep a manual override for critical paths — support tooling, billing-adjacent copy, or anything customer-facing under time pressure.

Pair flags with eval diffs. Rolling out a worse model to 10% is still a choice; blind is not.

Delete stale flags. The graveyard of experimental prompt variants is how teams lose institutional memory.