There is a particular kind of A that has become ordinary. The essay is fluent, the citations check out, the thesis is plausible. Two weeks later the student cannot summarise the argument — not because they are hiding, but because there is nothing to summarise. It was never in their head.
The grade is real. The learning is a fiction. And nothing in the system catches the difference until an exam, an interview or a cold call eventually does — by which point the gap is months wide and the student had no way of seeing it opening.
That is the failure this product is built against. Not cheating, exactly; most of the students it happens to are not trying to cheat. They asked a machine a reasonable question and it answered, and the answering is the problem. The moment where you try, fail, and repair is not a step on the way to learning. It is the learning. An answer machine skips it, politely, every single time.
So the design question was never “how do we make an AI tutor more accurate.” It was “how do we make one that refuses to be an answer machine, in a way a motivated seventeen-year-old cannot talk it out of.” Everything below is downstream of that question.