Hm, the puzzle_id thing seems… questionable?

Maybe I’m misunderstanding.

Props to HRM authors, and ARC team, for providing easy-to-replicate experiments. This is key to learning from the papers and building on it.