How do I tell the imaging team which faceplate marks actually fail?

The finishing inspector, using an appearance sheet that isn't in the camera folder, which may explain some of the argument

Bring that inspector and sheet into the trial. The camera team cannot infer the acceptance rule from a folder called bad, however confidently it has been named.

Ask the inspector for examples near the limit as well as obvious rejects. Operators will eventually need to understand uncertain results, not just recognise the dramatic scratch everyone can see.

They've picked some retained samples and want consistent viewing conditions for the comparison, plus one doubtful plate kept out of the accepted and rejected groups

Was the moving line checked on the physical sample?

Yes, no mark there, but there's a small rub at the edge that does fail; I nearly wrote that the plate was good after the first answer

We did exactly that jump once. Found a reflection, got pleased with ourselves, and stopped looking at the actual part. The real damage was somewhere else entirely.

Keep the rub location attached to that sample's images. Otherwise the training set may learn to reject a bright line that merely happens to appear on a genuinely bad plate.

Also show the operator what to do when the image is not usable. A sample decision and a camera-quality problem need different next steps, even if neither should become an automatic pass.

Have you tried putting the same retained plate back in the proposed presentation several times? Curious how repeatable it is before lighting changes get all the attention.

Only by hand so far, and we didn't record its orientation closely enough; the next static comparison will use a repeatable nest with the imaging person setting it up

Keep the failed captures too. They show the variation you need to explain.

I would separate those exploratory captures from the approved reference set, while keeping the same sample identity. Otherwise a later export can recreate the original naming mess with better photographs.

Done, exploratory and approved views separate, same sample IDs; thanks, Chloe, we had been about to overwrite the awkward pictures instead of keeping them

Did the repeatable nest settle the changing reflection, or is that trial still ahead?