The selected bezels are retained. Image labels came first; physical review has not happened yet.

Record how those images were made while the parts are being reviewed: holder, seating, lighting and capture settings. Physical judgements improve the labels, but cannot recover undocumented presentation conditions.

Will the person who reloads the bezel help define the repeat-placement cases? They need to reproduce the intended seating after handover.

Yes. Quality has now judged the retained examples; one real scratch is not where the bright streak was circled.

Keep that scratch's location tied to the physical bezel during the capture comparison; reducing the bright streak is not success if the confirmed defect becomes less visible.

And keep a separate validation selection once tuning begins. Familiar examples are useful for development, but they shouldn't be the only evidence used to judge the finished inspection.

Agreed. Repeat placement still changes scratch visibility, so we're reviewing the holder and lighting before selecting thresholds.

Who retains the physical examples and configuration record when this moves from development to routine inspection?

Quality keeps the examples; maintenance keeps holder and lighting revisions. No commissioned inspection claim while the visibility comparison remains unfinished.