Interesting project, but why didn't they just run the test for every input that is less than 64MB? I bet it is cheaper (and way more understandable) than generating 180 million lines of proof with frontier AI...