How do you safely deploy a new model version?

Use progressive rollout: shadow mode (run the new model alongside the old without serving its output) to compare, then a canary or A/B test on a small traffic slice with guardrail metrics, then full rollout — with instant rollback ready. Version the model, data, and config together for reproducibility, and keep an audit trail. For LLMs, pin model snapshots and run your eval suite before promoting. Never swap a production model based on offline metrics alone.