A courts summarization service has a new foundation-model version ready. What production deployment strategy fits?

Promoting the new foundation-model version through staging before switching production traffic is the correct production deployment strategy for a courts summarization service. Staged promotion lets municipal GenAIOps validate summarization quality, safety filters, latency, and integration against court-like workloads in staging, then cut production traffic only after evidence passes. That protects docket-facing users from untested model behavior while still allowing the service to adopt improved FM versions deliberately. Promoting the prior foundation-model version through staging before switching production traffic fails because it re-validates the old model and never deploys the ready new version, so production sees no improvement. Promoting the new foundation-model version through staging before switching development traffic fails because development traffic does not serve docket-facing users, so the production service never receives the validated new version. Promoting the new foundation-model version through staging after switching production traffic fails because it reverses the safe order and exposes production users to unvalidated behavior before staging evidence passes. Exam caveat: staged promotion is the default safe production strategy; canary/dual-run is a related pattern when policy requires side-by-side comparison for a period. When the stem simply says a new FM version is ready and asks for production strategy, choose promote-through-staging first. Operational check: deploy the new FM version to staging, run courts summarization evals and integration checks, retain the prior production version until cutover succeeds, then switch production traffic and watch quality/latency dashboards.