I've found that some labels were assigned from the images themselves. The samples are still available, but those labels have no corresponding physical inspection results, making the reference circular.

Mark those labels uncertain and have the samples inspected properly. Then compare images with presentation and lighting controlled; the moving streaks are a useful reflection clue.

That clue needs a limit: presentation can also change how a real scratch appears. Movement in the image isn't enough to reclassify the sample as acceptable.

Correct. I meant a reason to investigate the imaging, not a new defect label. Physical acceptance stays with the agreed inspection method.

Should uncertain images be treated as acceptable, or kept somewhere separate while their status is established?

Retain their sample identities and undecided status. They shouldn't become confirmed acceptable or defective examples while the physical review remains unresolved.

I've checked the image folders and found that exposure and presentation both changed between them. Their improved appearance can't be attributed to either change individually.

@LinAllen0270 I had a separate inspection setup where the 'better lighting' folder used easier samples too. Its name implied a lighting conclusion that the comparison didn't support.

@NadiaArcher0378 Which details did you retain with each image? I can record sample identity and settings, but I'd like a practical record people will actually maintain.

For my separate setup, sample identity, exposure, lighting arrangement and presentation condition were the useful essentials. They let us understand comparisons without relying on recollection.

@NadiaArcher0378 Those records can support the vision specialist's comparison of image conditions. Holding the representative sample set constant will make an acquisition change easier to assess.

Would a learned model just cope with the lighting differences?

It still needs trustworthy labels and representative evaluation. A different model doesn't establish whether your reference scratches are real or your acquisition comparison is fair.

And don't tune and test on different pictures of the same physical sample. That isn't the independence the result would suggest.