A controlled evaluation that pushes a model to its limits by testing edge cases and challenging scenarios.