Make the model write annotated tests too, verify that the annotations plausibly could match the test code, run the tests, feed the failures back in, and iterate until all tests are green?