Recover the batch identity with each sample number before trusting the labels; otherwise the same numbered bezel may be two different parts in your folder.
Rejoin the images from the original capture sheets, then have quality confirm the physical decisions; the corrected filename won't tell you whether somebody called a reflection a scratch.
Keep that accepted bright-band example in the imaging comparison alongside the confirmed edge defect, so the next view has to handle both rather than merely look quieter.