01
Start with a specific gap
Choose a model boundary and freeze a holdout before generating training data. A useful gap is concrete: for example, small text lost under a bounded range of motion blur, or identifiers missed after page cropping.
A broad synthetic corpus may improve coverage, but generation volume alone is not evidence that the model learned a useful representation.
02
Make the allocation inspectable
A recommended training mix should state condition weights, severity ranges, source templates and mixture order. Preserve enough metadata to trace any training sample back to the generation recipe.
The example 35% motion blur, 25% crop, 20% perspective and 20% compression mix on this site is illustrative product UI.
03
Retest without moving the benchmark
Keep related source groups out of both training and holdout. Compare the updated model against the same evaluation set and policy, retaining the baseline and changed model configurations.
Report regressions as well as gains.