A replay of one recorded session from 26 September 2026: three models (GPT, Kimi, GLM) answer the same question blind, a reader lists where they differ, each model ranks the answers without knowing whose is whose, then they argue one point in turn. Every word is what the models returned. The timing is 8 times faster than it ran and the waits between stages are cut. Nothing on this page calls a model.