Turn an ML paper into an evidence-backed experiment.
Solve ARC puzzles, build the rule, watch AI attempt it