Model divergence benchmarks

These numbers come from real debates run on the platform — not a synthetic benchmark. Every completed debate scores each agent's agreement and confidence each round; the stats below aggregate that data live.