Overview
Every item below is backed by a linked entity — a gate, a judge, a risk, a mandate. Priorities are labels drawn from posture, owners, and dates, never a score.
Needs your attention
every item is backed by a linked entityrun_5493 failed numeric — the agent answered 24.0 h at risk, the engine says 22.5 h — and two answers cite the superseded 2025 PTO policy. The PR is blocked.
base_rate disagrees with the engine after the substrate refresh — $31.25/h declared vs $31.75/h derived. SME review required before Calendar@2026.3 adopts.
5 verdicts waiting; the judge is in shadow at κ 0.71 — below the 0.80 gate floor.
The swap-eligibility suite runs but isn't wired as a required check; wire the gate to reach L2.
tr_5a209 and other captured traces are staged for a Presidio scrub and SME sign-off before they become cases.
Maturity vs. mandate
8 fleet entitiesSuites at a glance
live gate posture per agent| HR Policy Advisor — release gate | agent:hr-policy-advisor AND tier:golden | 34 cases | pass |
| WFM Assistant — dev loop | agent:wfm-assistant AND tier:(golden OR regression) | 61 cases | pass |
| WFM Time-Off — release gate | agent:wfm-assistant AND tier:golden AND topic:time_off | 24 cases | pass |
| Payroll Copilot — release gate | agent:payroll-copilot AND tier:golden | 28 cases | fail |
| Bryte routing — release gate | concern:functional agent:bryte-main AND tier:golden | 24 cases | |
| Benefits Explainer — release gate | agent:benefits-explainer AND tier:golden | 18 cases |
Start with the hero flow
what this prototype demonstrates end to endOne UKG-aware case: policy answer, exact accrual math, PII guarantees — authored in ~20 lines of YAML.
OpenGraders route to specialist backends — ADK eval, RAGAS, the UKG deterministic engine, Llama Guard — never scored twice.
OpenA selector picks the cases; thresholds live in suite policy; the same suite runs from a laptop and a pull-request gate.
Open