My example images show bright streaks moving across a polished cover when presentation changes slightly. People have already labelled them scratches. This fixed-camera FR5 station isn't commissioned. I need a reference set somebody can maintain, not thresholds built around whoever named the images first.
Keep the physical covers identified with their images. Ask the appearance owner to judge them under the required viewing conditions before using those scratch labels. Movement of a streak alone doesn't prove a harmless reflection.
I've retained the covers and removed the assumed scratch labels from the working reference list, not deleted the images. Quality has the judgement request. One image can't yet be tied to a cover, so it stays out of the labelled set.
Once those judgements exist, compare the same covers over the intended presentation variation. Keep lighting and fixture conditions with the results so maintenance can reproduce a troublesome view, not merely find its file name