But isn’t that how models are trained anyway? The training data provides the “test cases”. How would you write a test that doesn’t fit that paradigm?