Three retained. Images identify plates, not the claimed damage locations. Quality hasn't examined those locations yet.

Use the appearance requirement for that examination. The camera folder cannot supply its own ground truth. I would keep a region unclassified when there is no settled physical judgement, rather than forcing every example into pass or fail.

I had a manually accepted cover labelled scratched from its bright streak in another folder. The contradiction was useful evidence about our labelling, not evidence that every bright streak on every cover was harmless.

Owen, useful distinction. Tim, keep the original labels with their source, outside the confirmed tuning examples.

Will whoever normally seats the plates take part in the capture comparison?

Yes, our inspection operator. Quality found one actual line, but its acceptance classification is still pending.

Tim, did quality map that physical line to the corresponding image region? A line on the plate may not be the bright band that originally attracted attention.

Mapped separately. Actual line is near an edge; bright band crosses the middle on that image.

That gives the imaging comparison two different observations to follow. Reducing the band is not automatically helpful if it also hides the edge feature, although the product judgement on that feature is still needed.

Record holder revision and camera settings. Otherwise reseating comparisons can quietly become different-setup comparisons

Any agreed reference classification yet, Tim?

Quality rejects the edge line under our requirement. Other queried regions remain unclassified. No camera threshold decision yet.

Keep an accepted reference too when one is available. You need to investigate false rejection as well as detection of that identified reject feature. Your current three plates do not establish the breadth of production variation.

Accepted-reference selection added to the plan. Thanks Owen; we will track line visibility, not simply reward the cleanest-looking image.

That is a clearer comparison. I would still keep tuning and later validation examples distinguishable, so the next report states whether the method is being adjusted or independently checked.