Search cases, suites, agents…⌘K
DEMO · synthetic data
LLM as judge

Judges & calibration

Every judge is a versioned asset — a plain-language rubric, a model, and a calibration record against UKG domain SMEs. A judge scores in shadow until its agreement bar clears and an SME signs off; degraded agreement demotes it again.

The standing boundary — judges never touch regulated numerics. Every rate, accrual, and pay figure is settled by the deterministic engine; a judge only reasons about language.
Lifecycle— trust is earned forward and lost backward
draft
authored — not yet scoring
shadow1 now
scores everything, gates nothing
gate-eligible3 now
cleared the bar + SME sign-off
drift-watch
live samples double-labeled continuously
recalled
degraded agreement auto-demotes and alerts suites

Historical: policy-answer v3 was recalled 2026-03 — a model swap shipped without recalibration, its weekly drift sample fell below the bar, and it was auto-demoted to shadow. See its version history.

SME time this quarter
6.4 h
across all four judges
Verdicts kept trusted
≈ 3,900
per quarter
Full-human-review equivalent
≈ 325 h
of SME review avoided
Leverage
≈ 50×
review hours saved per SME hour
Judges— one earning trust, three holding it