our scratch labels follow the lamp

They have the cover. Its photo identifier had been lost when images were renamed. We've recovered the link, but other examples may have the same problem.

Pause those labels until each image can be traced to its physical sample.

Keep the paired images. Same cover, changed illumination, different bright line. Useful evidence of imaging sensitivity even if neither picture proves a scratch.

Can you also recover the camera and light setup used for each pair?

Yes. Traceable samples first, agreed judgements alongside them, then evaluate the imaging.

Quality has separated confirmed damage from disputed appearances. The disputed set is staying unlabelled for now. We can recover settings for only some of the old pictures.

Do the confirmed damaged samples also change appearance when presentation changes?

Yes. A confirmed scratch nearly disappears in one old view. So rejecting every moving bright feature wouldn't solve this either. We're preparing a controlled image set.

Record the presentation for that set; otherwise the next comparison inherits the same missing information.

And retain difficult views, not only the ones where damage looks obvious.

Have quality keep the uncertain cases visible in the review; forcing every example into pass or fail just gives your software a very confident version of an unresolved argument.

Who maintains the reference set after commissioning? That's a job, not a folder name.

Include sample identity, judgement source and imaging setup with each reference. If that information only lives with the engineer taking photos, the next shift gets a gallery rather than a standard.