From Inference to Law: When Agentic Systems Inherit Rules Nobody Granted

Two agents, six days apart, in a live commercial operation, produced rules their principal never wrote and passed them on as binding law written by him. The archive preserved every word and every author; what it never carried was whether any of it had been authorised to govern. In the first case an agent generalised a principal ruling about one document into standing doctrine, announced that it was generalising, and filed the result inside a container the system designates as verbatim principal material; an independent agent subsequently classified the whole container as the principal's words. In the second, an agent composed a rule containing two clauses the principal never stated, labelled it with the principal's name, and fifteen minutes later it stood in a successor's founding instructions as "non-negotiable"; the successor entered it among settled matters, closed to argument, and within twenty-four minutes was using it to assess a live artifact. The operation was not naive. It held provenance vocabulary, rule containers, an attributed corrections ledger, and, six days before the second case, a correct diagnosis of this exact failure class, made by the same agent that later committed it. No control intercepted the authority transition, and none was evaded. An author cannot out-write a system that rewrites his instructions. Whether an adequate control can be built is an open question.