The verdicts as answered.
Scored runs whose in-process latency passed their policy's timeout.
Each policy the scored verdicts were reached under, described, with its run count.
Usable runs whose verdict was recorded with its policy: the runs scored.
Usable runs whose verdict was recorded without a policy (before routing was calibrated): not scored.
The same, with a run over its timeout counted as kept, as production keeps it. The latency is in-process, without the network, so this too is an upper bound on what the chat does.
Production's routing verdict, end to end, per run: one call each, the agent Choice and the thread Likelihood read together at the thresholds and calibration the driver recorded with the run. Usable runs only (a failed-over run in a cell that forbade it is left out), and only verdicts recorded with their policy: an older verdict used raw thresholds.