Most AI coding benchmarks still ask the question: did the agent produce code that passes the current tests? This is a useful ...