This is fascinating! Using prover-verifier games to improve the legibility of language model outputs sounds like a game-changer. It's intriguing how focusing on making outputs verifiable by weaker models also helps humans evaluate them better. This balance between correctness and clarity could have huge implications for AI reliability. Anyone else think this could be a big step towards more transparent AI systems? Would love to hear your thoughts!