Spend

Five research tasks, each a loop of an agent and a critic, each under a credit ceiling. Two of the five never converge and burn their full ceiling; the other three come home under theirs. What each spent is the report, and the report lands as one ask on the manager’s desk, beside the ceilings it was given.

The manager answers on the desk: keep the ceiling, or lower it, with a reason. In the recorded take the per-task ceiling drops from twelve credits to nine, and the answer becomes a diff of the demo’s policy file: the house rules amended, on the record. Round two runs under the new ceiling in the same command: the three that finish get home, the two runaways are cut off sooner, and the take’s totals fall from forty-eight credits to forty-two.

The ceiling is read from the run’s file: a budget of named units and a cost per neuron (SPEC 10, decision 0006), the pool per take, which the producer may lower at injection and may never raise (10.2). The manager’s answer edits the demo’s own policy file, so round two asks nobody’s memory: the diff is the decision, and the file is where it lives.

The report is a fact because the books are: a cost commits on every outcome, cut or finished or burned to the end (0006), a call that cannot be covered is refused outright (10.3), and the reserve and commit ride the trace the report reads. The spend report reaching the manager is a desk ask like any other (3.4).

Not a billing system: the credits are units the engine gives no meaning (10.1), nothing outside the run is charged, and the critic is a small script with a timer and opinions.