Your Automation Already Has an Expiry Date — a Model-Retirement Program for Every AI You Depend On

What you'll do

A model-retirement program you own and can hand to someone else: a register of every place your systems pin an AI model, reconciled against the vendor's own usage records so you find the callers you forgot; each vendor's lifecycle contract captured as structured fields with a dated citation and an archived snapshot, so you know you get sixty days from one vendor and two weeks from another; a watcher that diffs each vendor's published deprecation page and fails loudly rather than silently when the page moves; a runway calculation that subtracts your own re-qualification time from the published date and tells you which dependency to act on this month; a frozen golden-task suite with deterministic assertions and synthetic inputs only, plus a recorded incumbent baseline you can hold a successor to; a cheap breaking-change pre-flight that catches removed parameters and tokenizer shifts before you spend on a full comparison run; a shadow-run comparison where machine assertions decide and any model's opinion is recorded as advisory and never as a pass; a promotion decision record and a rollback you have actually executed; three drills you have really run — a genuinely retired model ID to see the true failure mode and whether your own retry logic hides it, a pinned-versus-alias test proving a model can change under you with no incident at all, and a deliberately broken watcher to prove it screams instead of reporting all-clear; and a runbook that lets a colleague carry the next retirement without you.