Abstract
Generative development is an open loop: a language model produces the code, and readiness is judged by an opinion. This work studies whether a transactional state machine external to the agent — owning the requirements canon, an executable checks table, a double run of every check, and a signed verdict — turns readiness into a reproducible fact. Field experiments ran on a battery of console-utility tasks spanning five task senses and five technology stacks: a three-hand experiment — top and budget models in the system, solo without it — with independent judges, and a September replication on the evolved machine. The system hands delivered verified-twice verdicts on every task, quality invariant to the model class; solo, cheaper on its own tokens, matched the byte-exact reference in a fifth of the judges’ rows — the divergence is freedom of interpreting the brief, not functional defects. Systemic fixes nearly halved the battery’s token cost, the methodology unchanged. On the full path, with the verifying operator priced in, the machine route is also the cheaper one, and the human leaves the loop on most tasks. The closed loop’s benefit is reconcilability, not token economy; an anti-entropy frame is proposed to interpret it.