A plant that treats ML like a static appliance discovers the failure mode on a Monday after a “minor” recipe tweak. Camera angle unchanged. Same model ID in the registry. False rejects climb or soft-sensor bias walks. The MoC closed for the process. Nobody froze, canaried, or re-validated the model against the new world.
This is not look-ahead joins, not active-learning easy-case bias, and not scrap-code taxonomy. It is change control that includes the scoring artifact.
What actually couples
Features are not only tags. They are lighting spectra, recipe segments, alarm limits that gate labels, and the physical setup the camera or soft sensor was calibrated against. A POR revision that moves a rinse time or a weld schedule can shift the input distribution without touching a weight file. Scoring the old champion into that shift is silent domain change.
The gate that belongs on the MoC form
Require a model disposition on every process MoC that touches measured or imaged quality: freeze and re-score a challenge set, run a time-boxed canary with kill criteria, or explicitly accept risk with a named owner and review date. “Model unchanged” is not a disposition—it is a hope.
Log recipe revision ID next to model version on every scored lot. When escapes cluster after a MoC, join those two columns first.
What quality meetings should stop doing
Stop opening retrain tickets before asking which recipe, fixture, or lighting MoC landed last week. Stop promoting champions on data collected under a POR that no longer runs.
The short rule: if the process may change under a controlled form, the model must appear on that form.