Making sense of model results requires engineering judgement, and always will. But what if we could develop testing tools that quickly flag common errors, or big changes, or unusual inputs or results?