Conjecture

Attributability is one of two independent constraints on paying for harms that did not occur, and for AI correction capacity the other one binds harder. The two are: whether the avoided harm can be traced to the actor paid to avoid it (epistemic), and whether the actor who forgoes revenue to avoid it keeps the gain (distributive). The health record shows them failing separately, which is what makes them independent rather than one constraint described twice. For correction capacity the distributive constraint is worse than in health: Medicare at least bore the cost and banked the saving inside one payer, whereas the harm from an uncorrected agent error falls largely on parties other than the operator who would pay to prevent it. Prediction: correction capacity gets built where the harm returns to the builder through liability, reputation or deployment rights, and stays thin where it does not, roughly independent of how traceable the error is. The same constraint outside health, in Diamond: clean firms outcompete dirty ones only where regulation is effective and the public cares, that is, only where the external harm is routed back to the firm.

Coordination architecture