Retain the source identifiers, masks and processing settings used to create the comparison. Without them it can be difficult to determine whether a changed output reflects the landscape, the observations or the method.
Use a synthetic scene to understand the mechanics of an analysis, then keep that evidence separate from performance on real imagery. Demonstrating the code path does not establish the accuracy of an operational finding.
Review the errors that matter to the decision and record the limitations that remain. A bounded, repeatable finding is more useful than an impressive image with uncertain meaning.

