I know physical reference decisions come before thresholds. Our FR10 bezel examples still call shifting bright streaks scratches, though the cell isn't commissioned. How should the image set show which labels remain unverified?
Give the annotation a status separate from the image itself, with the part identity and physical reference decision linked when available. A bright region can be a candidate for review without becoming a confirmed defect in the tuning set.
And make that status visible to whoever opens the images, not only in a separate spreadsheet. I've watched people print the picture and lose the caution that lived in the folder notes.
Once quality has identified physical reference locations, compare repeat presentations without changing those decisions to suit the picture. Reducing a bright streak is only helpful if an actual known scratch remains observable under the conditions you intend to use.
Could the training set keep a separate unresolved group meanwhile? I'd rather show a learner why we can't decide than make them choose pass or fail from that glare.
Yes. Unverified tiles now export with that status and stay outside threshold tuning. Physical bezel review and repeat-presentation comparison are still outstanding.