Three plates are still here and their image filenames identify them. Other photographs were supplied without sample identities, so we cannot trace those back reliably.

Keep the untraceable photographs outside the reference set for now. Has quality judged the three retained plates using an agreed physical inspection method?

Keep repeated views of each identified plate together. If the folder treats every photograph as a different sample, it can hide the fact that one physical plate changes category when it moves.

Quality has not judged them yet. I have grouped the views by plate rather than by the proposed scratch label; that makes the contradictory labels quite visible.

Use a small controlled comparison before collecting hundreds more images. Ask the inspection specialist to assess presentation and lighting with known acceptable and rejected examples, preserving the real defect visibility rather than merely reducing bright pixels.

I agree on the smaller comparison. Hazel, do you have any independently judged reject available, or would that also need to be selected with quality?

Quality has now judged the three retained plates: two acceptable and one rejected for a physical scratch. They have marked the scratch location. Lighting and presentation work is still pending with the vision specialist.

Keep that location and the physical decisions attached to the plates through the comparison. The next image condition should reveal the known scratch consistently, not persuade everyone that its bright neighbours have vanished.