“literature before 01 was released on using this kind of verifiable outcome rewards whether a math problem is correct or a code snippet is correct.”