Overview
Eval suites
Section titled “Eval suites”Evaluation suite catalog (list, get, versions, publish, unregister) — part of the admin control plane. Full versioned CRUD, mirrors policies 1:1. Registry-only: eval-run execution + per-kind grader dispatch live on the eval-runs tag. evalKind selects the shape of spec — closed enum, extended additively (accuracy, pairwise, regression, human-review, benchmark, custom). Runs against these suites are started through the eval-runs tag. Caller-plugged via EvalSuiteRegistryBinding.