Removing humans from the reinforcement-learning loop improves capability rather than merely reducing cost, which puts centaur architectures in direct competitive tension with the fastest available capability path.

Why. Dissolves the apparent conflict between human removal improving capability and the centaur commitment: the RL result is a phi-p finding and the centaur argument is a phi-b argument, so they are not competing claims about one quantity.

A superseded claim stays admitted: replacing is not deleting.

Why. Narrowing the relation I first recorded as rebuts. Azhang's evidence is scoped to domains with dense objective verifiers, where o3 discarding human-crafted domain knowledge performed best. It therefore limits when the centaur commitment holds -- human-in-loop carries a capability cost, not merely latency, in cheaply verifiable domains -- rather than defeating the commitment in domains where the reward signal is contested, which is where the collective's judgement work sits. Both edges are left standing so the overcorrection is visible in the record.