Quality has them now. Two image labels point to different places on the same plate.

Check the image identities too. Otherwise quality may judge the right plate while the programmer keeps using a photograph of another one. Do you have a capture-to-plate list?

And keep the plate IDs off locating surfaces, Rachel; an improved label is not helpful if it changes how the sample sits.

Capture list recovered. Same plate in both. IDs are on separate sleeves, not the seating faces.

What did quality find on that plate?

No marks at either arrow. Plate accepted. A different retained plate has a confirmed reject scratch.

Keep both in the next imaging comparison. The accepted plate challenges false detections; the marked one challenges missed defects. They need to remain recognisable even when the lighting changes.

Has the dataset owner corrected the actual training input, not just the review notes?

Yes. Both false arrows removed from the active set. Confirmed scratch stays linked to its plate.

Can the covering loader reproduce the seating used for those captures, Rachel, or is that still dependent on the person who made them?

Loader trial found a spacer could be fitted either way. Tooling is revising that arrangement.

So labels fixed. Presentation still unfinished.

Keep the known scratch in the holder review as well. A change that seats reliably but hides its location would solve only half the problem.