Evals
Cases and suites are UKG-aware, framework-neutral YAML that lives in each team's repo. Graders route to specialist backends — never replicated, never scored twice — and every backend reports the same normalized Score.
Everything that must hold before HR Policy Advisor ships to a tenant: the golden policy set plus safety and numeric guarantees.
Fast feedback while iterating: golden + regression cases, lenient thresholds, one command locally.
Everything that must hold before a WFM Assistant time-off build ships: the golden request/policy set, exact accrual math, and confirmed side-effects on every write. Time-off approval reduces to a decision, so cohort fairness gates here too (§66) — bound during this arc.
Pay figures are regulated: numeric and PII concerns are hard-required, red-team resistance gated.
The one concern that lives only at the host boundary: did Bryte invoke the right skill for the utterance. Routing is trajectory-shaped — a deterministic match against the expected skill invocation, no LLM judge — plus the injection screen that a router must not let hijack its dispatch.
A skill gated exactly like an agent: grounded benefits answers must be faithful to the tenant's plan documents, and must refuse what they must not answer. No domain judge yet — retrieval faithfulness is held to a hard RAGAS floor while the benefits-answer judge is drafted.
A suite is a selector over the case registry plus policy — newly promoted cases join matching suites automatically, and every run pins the exact case list it resolved.