After launch
A camera was replaced. Here is what it does.
New hardware, different colour response and sharpness, and the model treats a familiar scene as unfamiliar.
From the floor
A failed unit was swapped for whatever was in stock. Same position, same view, different sensor. To an operator the picture looks fine, possibly better. To the model the colour balance, noise floor and edge sharpness have all shifted at once.
Which failure is it
Covariate shift
Inputs changed at the sensor rather than in the world. It behaves like a repositioning but the boundary is the hardware swap, so the diagnosis is different even though the mode is the same.
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.
A single camera changes character on the exact day it was serviced, while its neighbours are unchanged. Reads as an abrupt step rather than a slope, which is the clue that this is hardware and not the world.
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
Re-label a short window from the new sensor. If the fleet is being replaced in phases, do the first unit properly and the rest of the rollout inherits it.
What it costs to ignore
Maintenance improving the picture for humans while degrading it for the model, which is the least intuitive failure on this list and the hardest to get anyone to believe.
Where this shows up
What it gets first