Collect target observations with a defined scope and a reference process appropriate to Site B's quality decision. Preserve information about product family, production batch, camera configuration, time, and unresolved cases. Consider which units are independent; many nearly identical images of one item do not represent many independent defect events.
Separate development and evaluation material in a way that reflects intended use. A later production period or independent batches may be more informative than a random image split when adjacent images share the same item or setup. Keep fitted preprocessing and adaptation choices inside the development process.
Retain a target evaluation set for the final comparison and document exclusions before examining results. If the evidence covers only one material and one lighting arrangement, state that boundary. Additional sites or conditions require their own supporting observations rather than inheriting a generalization claim from this study.