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.

@IsabelBell0648 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.

@BrunoArcher0365 A scratch's visibility can change with presentation too. Don't let 'moving streak' quietly become the rule for relabelling something acceptable.

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

@BrunoArcher0365 So uncertain isn't the same as good? Where do those images go?

@NoahChen1151 Keep them explicitly undecided, with the sample link. Don't use them as confirmed examples of either class until the physical review settles them.

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.

On my inspection setup, a folder called 'better lighting' also contained easier samples. The folder name was doing a lot of unearned work

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

@OwenBennett0747 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.

@BrunoArcher0365 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.

I'll group images by physical sample before dividing the tuning and evaluation sets. The current filenames wouldn't reliably expose the same sample appearing in both.

Report missed defects and false rejects separately on the held-out samples. One agreement percentage can hide which mistake your method is making.

How should the evaluation report handle samples whose physical status is undecided? I want them visible without assigning an unsupported error label.

Show them as undecided with their count, outside confirmed-class error rates. That keeps the coverage limit visible without inventing a ground-truth label.

I can explain the reference uncertainty better, although the example images still don't support a performance judgement. Physical classification and image-condition comparisons remain outstanding.