I'm looking through the bezel pictures for our UR5e inspection proposal and the same part changes from good to scratched when somebody reseats it. That's a fairly impressive rate of damage for a still photograph.
Before we start adjusting thresholds, I want quality to examine the real bezels and stop the filenames deciding the answer. What should we keep with each image so the next person can repeat the comparison?
Physical sample ID, its agreed judgement, and the camera, lighting and seating arrangement. Keep disputed labels separate. A consistent filename won't rescue an uncertain physical reference.