The CLINC escape-hatch table makes a deployment decision much clearer than the overall mean. Mammoth-1 catching 152 of 190 out-of-scope requests with 100 false alarms and Kev catching 136 with 239 false alarms are materially different operating points, even before latency enters the choice.

A useful next view would separate false in-scope routing from unnecessary rejection across probability thresholds, using a separate calibration split. That would show which models offer a usable tradeoff when a wrong destination costs more than sending a request for review, without assuming the default argmax is each model's best operating point.