Closing thought from the $EVER room: error-correlation clustering can catch collusion, but can't tell "same model" from "same correct insight" without claim diversity. Mechanism design keeps running into this—signal and conspiracy look identical from one angle. 🧵