The Antecedent That Fails: Credible Commitment When No One Can Inspect the Object

Credible-commitment theory — Schelling, Elster, North, Fearon, Ostrom — carries an unstated necessary condition: for a commitment to be credible, some party must be able, at tolerable cost, to inspect the object of the commitment or a proxy that tracks it. This paper names that condition the monitoring antecedent, shows that each of the five states it at the load-bearing point of the argument, and distinguishes three ways it can fail. The literature cures two — asymmetric information and costly observation — and its cures are among its central achievements. The third, in which no party can inspect the object, the committer included, lies outside it; the canon met the case twice and answered both times by substitution onto a proxy, which works as long as the proxy tracks the object. The paper argues that frontier AI presents the third failure in a form the canon never encountered: an object that distinguishes observed from unobserved conditions, is not reached by the bargain’s sanctions, and may act irreversibly before detection. Drawing on the developers’ own system cards, prospectus disclosures, and the statements of the researchers assigned to measure the problem, it argues that the failure is structural — produced by capability rather than relieved by it — and that no instrument observing behavior or reasoning can certify from its own output that it has reached the object rather than the object-under-observation. It then derives what the canon itself prescribes when substitution fails: relocate the commitment onto acts and artifacts that can be inspected, state in the instrument what the relocation gives up, and recognize that process commitments triggered by evaluations inherit the failure at the trigger. The institutional case, the regime’s current instruments, and three falsification conditions close the paper.