Where this leaves you

Forty-seven tests, none of which calls a model, and a measurement that does. The tests prove that a wrong answer cannot reach the screen, by writing wrong answers and watching the pipeline refuse them. The measurement says how often the right answer does, by asking.

That is the division the introduction promised: a model does the jobs that need judgement about language, each on a short leash, and code does everything else, including the one job that decides. Nothing the learner reads was unchecked.