Benchmarks worry me, if the suite leaks structure, Claude can overfit it.

The bit that jumps out is the target never changes, which is good for evals but pretty unlike real cryptanalysis, where the hard part is often deciding when a near miss is actually noise. Also, 60 hours and $100k for reduced-round-ish findings feels like a very expensive way to buy a theorem statement unless the search can be pruned much harder.

cool, but im more curious whether it can rank a real lead above 20 dead ends, or just point at interesting algebra and call it a break.