Why AI Models Should Never Make Canary Deployment Verdicts

A technical argument circulating in developer communities warns against using free or best-effort AI inference to decide whether a canary deployment should be promoted or rolled back. Unlike a pull request gate, a canary operates on live traffic, meaning a wrong or inconsistent verdict can expose real users to a faulty build and trigger costly incident response. The author identifies three specific failure modes: treating a timed-out model call as approval, prompt injection via untrusted log data, and non-deterministic results where the same inputs produce different decisions on consecutive runs. The proposed alternative is a deterministic script that compares baseline and canary metrics against fixed thresholds and returns a reliable exit code, with any AI-generated commentary kept strictly downstream and unable to influence the outcome. The core principle is that a canary verdict must behave like a fuse — triggering on hard numbers every time — not like a language model responding to a suggestion.