Capacity & campaigns Preview
Calibration measures what a run actually costs in wall time, dispatches and tokens; the campaign cadence comes from that number.
Calibration computes median wall time, median dispatches and total tokens per team-and-model pair, and renders an estimated runs-per-five-hour-window. A campaign then runs continuously and resumably against that budget. The bare calibrate command is a real gate: it exits nonzero while any registered team's eval fixture is missing, so you cannot start measuring teams that have nothing to be measured against.
Account rotation fails closed
Rotation leases the least-utilized account below a utilization threshold, and refuses to lease anything if usage cannot be read or a token cannot be minted. The token rides only in the child environment — never in argv, never in config, never in logs.
The honest warning it prints
Without an account pool, an unattended calibration run will burn your own interactive quota. That is in the team's own documentation, in those words, because the failure mode is silent until you go to use Claude yourself.