How do I stop this faceplate rocking during our UR5e camera comparison?

Same plate rocks. Two clean-looking ones don't. None has a recorded quality decision.

Keep the original pictures as well as the folder labels. On another job the presentation copies had changed what an edge looked like. That won't settle your plate's condition, but it can explain part of the image disagreement.

And keep the awkward plate in the holder comparison, Ella; a test made only of the two that sit nicely would be rather kind to your fixture.

It's kept separately and identified. Quality is checking it; I'm checking where it touches the holder.

Good. Look for an unintended contact, not just a missing clamp. A part can rock because it touches something that was meant to clear.

Lin, would you remove the guessed labels entirely or retain them as image history? I wouldn't want someone mistaking old guesses for current decisions.

Retain the originals and their history outside the approved reference set. Make the current quality decision explicit on the controlled sample record. The training or tuning set should not inherit the old guesses.

Oscar, thanks for distinguishing history from the working set. Deleting the guesses would also hide how they spread, but nobody needs them offered as truth again.

Quality accepted the rocking plate's surface. Its folded edge touches a holder corner that should clear.

Then have the fixture designer compare that edge with the permitted part geometry. Don't modify the holder around one sample until you know what range it must accommodate.

That accepted plate is a useful awkward example now. Do you also have a confirmed reject to check the eventual lighting still reveals the damage that matters?

Quality selected one with a rejected nick. Holder change not designed yet. No threshold work started.

When the presentation is repeatable, compare both identified samples under the candidate lighting. Include repeated loading, rather than selecting one well-positioned image of each. Acceptance decisions and image conditions should remain traceable separately.