I suspect this is going in the wrong direction. Telling a sandboxed AI to have a long conversation with a student to ensure they actually know what they're talking about, while giving minimal hints away, seems like the scalable solution. Let students tackle the material however they will, knowing that they will walk into class the next day and be automatically grilled on it, unaided. There's no reason a similar student couldn't have their essay fed into the AI and then asked questions about what they meant on it.

Once this becomes routine the class can become e.g. 10 minutes conversation on yesterday's topic, 40 minutes lecturing and live exercises again. Which is really just reinventing the "daily quiz" approach, but again the thing we are trying to optimize for is compliance.