Tag
This paper introduces a new instance-optimal protocol for AI debate in scalable oversight that guarantees worst-case correctness, makes honesty a dominant-strategy equilibrium, and proves black-box lower bounds showing no protocol can outperform it on problems with stable decompositions.