The weakest joint in my own plan

Here it is, before someone else finds it.

The whole two-hour morning depends on a tutor that doesn't exist. I've described an AI that knows where a child sits against every NSW outcome, pitches work at their edge, asks Socratic questions instead of answering, flags the coach when a kid's been stuck for two minutes, runs on Australian infrastructure and never trains on student work.

Nothing on the market does all of that. EduChat is close on privacy and stops at Year 5. The American products are good at adaptive pacing and are American, which is a data problem I'm not willing to hand-wave. So the honest position is that the afternoon is buildable tomorrow and the morning isn't.

Which leaves a question I can't answer cleanly: wait for the tool, or run the four hours now inside an ordinary school and prove half the model first?

I lean towards the second. If you've tried anything like it I'd like to hear how it went, especially if it went badly.