our bad cover folder contains the same cover looking good

Cover is still here, happily. Nobody marked the area, and the person who labelled the folder was judging from the images rather than inspecting it

Right, so those labels are opinions about pictures. I'd have quality inspect the cover and record any actual defect location before using either image as a tuning reference.

Who defines an unacceptable scratch for this product? Identifying a surface feature and deciding that it fails the appearance requirement are separate decisions.

Quality should answer that too. Sorry, I skipped from finding a mark to calling it a defect. There may be allowed surface features we need the camera to ignore.

Agreed, but I'd keep a few controlled pose variations in the trial as well. We need to know what happens when the presentation isn't perfectly repeated, not only get one flattering image.

Quality found no scratch at the bright line's location. There is a small edge scuff elsewhere. We've recorded both, and neither original folder label tells that story

Retain the corrected explanation with the images so a future trainer does not accidentally restore the old labels when rebuilding the dataset.

Done. Old labels remain in the audit history, excluded from our reference set. Next question is how much pose variation our fixture actually permits

Use that permitted variation in the training examples, with quality's decision attached, so an operator can distinguish an uncertain image from an accepted surface condition.

Nice to have the physical cover settle one part of the argument. Are you measuring the presentation variation, or is that still someone wiggling it by eye? We started with the latter and got very creative results.

Measure the seating repeatability first. My 'hold it still' reply made that sound easier than it is.

A simple repeat-seating study should give the fixture and imaging people something shared to work from. Let them define the method; otherwise a camera change could hide a seating problem you're meant to fix.

Has quality confirmed whether the edge scuff is acceptable? The bright reflection is now explained, but that second feature still needs a disposition before the cover becomes a labelled reference.