Plates retained with IDs; labels came from image sorting, not physical inspection. Quality has not classified them yet.

Keep them unclassified then, rather than converting a doubtful scratch into an accepted surface simply by deleting the word

And preserve the full images and capture conditions. A small crop may hide the presentation change that moved the streak.

What do the training notes currently tell the operator about those examples? I'd stop them becoming an unofficial standard while quality is still deciding what the actual surface contains.

Only a draft image sheet so far. I'll mark the judgements unverified before anyone uses it in training.

That handles the teaching risk. For the capture comparison, can the holder and camera owners record seating and settings so they can change one condition deliberately rather than compare unexplained views?

Clara, yes, but the controlled images still need independent product judgements; repeatable glare is not a confirmed scratch either

Both tasks remain open: quality's physical judgements and a controlled presentation comparison. Same plate number will link the records.

Include the actual marked region in that link too. One rejected plate can have harmless reflective areas elsewhere.

Ella's point would help the lesson writer. A whole-part reject label isn't permission to circle whichever bright line looks most convincing in the slide.

Thanks Ella and Yasmin; region references are in the request. No physical classifications returned yet, so the draft stays unverified.

Once those decisions exist, compare visibility of confirmed features and responses on accepted regions separately. Keep a later validation selection distinct from the examples used to adjust the method.