Evals

The eval gate: a candidate version against the live one on the same cases, before it may take a canary.

How the gate runs here

Runs start in GitHub Actions, from a version's page. Each agent has its suite: τ³ tasks for Support, labeled routing cases for the Dispatcher, a carrier-delay day for Ops and Insights.

dispatcher: dispatcher, 40 cases x 1 on mediumsupport: tau3, 3 cases x 1 on mediumops: scenario, 1 cases x 1 on mediuminsights: scenario, 1 cases x 1 on medium

Candidate	Against	Suite	Status	Scores	Cost	Requested
insights v2	insights v1	scenario, medium	passed	1 vs 1	$0.22	Sep 30, 19:48 by operator