After launch
The equipment changed. Here is what it does.
A new machine generation, a new housing or a new supplier's part, and the model has never seen the thing it is now looking at.
From the floor
Procurement changed a supplier or a site upgraded a fleet. The asset does the same job and looks meaningfully different: a different casing, a different colour, a fastener in a new place. The model was never wrong about the old part and has no opinion at all about the new one.
Which failure is it
Covariate shift
The definition of a defect is unchanged; the object carrying the defect is new. Inputs moved, the mapping did not.
Four modes, and they need different responses: covariate, concept, prior, and a pipeline event wearing drift’s coat.
What moves first
Accuracy is the last thing to tell you.
Detections collapse on the specific sites or lines that received the change, and hold everywhere else. The boundary is an asset population rather than a date or a camera, which is what distinguishes it from a repositioning.
Measuring accuracy directly needs fresh labels, so teams lean on proxies between audits. By the time a quarterly audit shows a drop, the system has been wrong for a quarter. The cheap signals are confidence per class, detection rate per camera against its own history, and how often a reviewer overrides the model.
What to do
Label the new equipment and retrain. Worth doing before the rollout rather than after, since the changeover date is usually known months ahead.
What it costs to ignore
Blind spots that map exactly onto your newest and most expensive assets, which are the ones you can least afford to stop watching.
Where this shows up
What it gets first