This is my submission for the AI Alignment Awards 2023 competition that won a first round prize.
I wrote a LW post describing it.
Original submission is in this repo, but you should probably just read the post (I now think lots of things in the .pdf
might not make that much sense).