Adding a reviewer agent bought one useful objection in five, and two new bugs

Both halves of the loop are rewarded for the same thing. Thoroughness, not by any instruction the operator wrote but by the way the pair is set up ^[14]. A reviewer that returns nothing reads as lazy. A reviewer that returns four true findings reads as diligent, and nothing charges it extra when all four are inert. The builder that answers each finding carefully also reads as diligent. Nowhere in that arrangement is there a penalty for a correct objection that changes no outcome.

Which is why the ratios matter more than the anecdote. One objection in five altered the result, so 80 percent of the review output was accurate, well argued and paid for in wall clock without moving anything ^[1]. Set that against the builder's own defects for the day and the review layer produced two errors for every one it prevented ^[2]. The operator's own summary is the right one: quality control had become the largest moving part in the machine ^[15].

Conflict is self-instrumenting. Agreement is not. He reckons a fight between the two agents would have surfaced inside a week ^[4]. The mutual admiration ran for weeks, because at each individual step it is indistinguishable from care.

The paper trail is where this becomes measurable, and it is the only measurement in the account that does not depend on a model grading itself. 505 prose lines out of 668 puts 76 percent of a machine-readable header into essays ^[3]. Strip the prose and 163 lines remain in a block designed to carry six or eight short fields, so even the residue is roughly twenty times its intended size ^[4]. Fifty of that project's fifty-four documents went unopened on the day of the audit ^[5]. Fifty-nine percent of the projects carried a briefing nobody had touched since rollout while their code kept shipping ^[6]. In the folder of rules governing how the two agents should cooperate, one section's entire content was an explanation of why that section had previously appeared twice ^[13].

The account has limits worth keeping in view. It is one operator, one toolchain, around a dozen products ^[1]. The five-objections figure came from asking the system under audit to score itself, which makes it an order of magnitude rather than a measurement ^[6]. What survives both limits is the shape of the failure. The supervisor did catch the thing a builder structurally cannot, namely whether the code just described is the code actually on the machine ^[3]. That value was real. It arrived attached to a process with no stopping rule and no cost for being right about nothing.