It runs the same task multiple times and observes whether reasoning invariants hold under repetition.

No performance claims. No capability or correctness claims.

v1.0 is frozen. If the protocol is flawed, I’m interested in where.