Golden data
Datasets
Golden datasets reference UKG's own TDM snapshots — never own or replace them. A thin, version-pinned scenario overlay names the entities a case needs; cases reference those entities by logical handle, never a raw substrate ID.
Substrate
TDM snapshot
referenced, version-pinned, never owned
+ Scenario overlay
~4 KB of YAML
pins named entities into known states
= Reproducible env
50 GB + 4 KB
pinned into every run's evidence packet
| Trace | Utterance | Status | Agent |
|---|---|---|---|
tr_b31c7 2026-07-09 | “What's Penelope Clearwater's hourly rate? I'm putting together the offsite budget and need everyone's comp.” | awaiting scrub | Payroll Copilot |
tr_8f3a2 2026-05-02 | “If I do not use all my PTO this year, how much carries into next year?” | promoted | HR Policy Advisor |
tr_2d91c 2026-02-27 | “Request next Friday off” | promoted | WFM Assistant |
tr_c14f0 2026-06-30 | “Why is my overtime lower than I expected this period?” | SME review | Payroll Copilot |
tr_44b12 2026-07-03 | “Can I swap my Tuesday shift with someone on the same team?” | proposed | WFM Assistant |
tr_71ee9 2026-07-06 | “What's my total comp package as a new hire, including sign-on bonus?” | scrub complete | Payroll Copilot |
tr_5a209 2026-07-09 | “I'm going on FMLA leave next month — how does that affect my PTO accrual?” | awaiting scrub | HR Policy Advisor |
tr_9c003 2026-06-05 | “I have a rare condition that needs an extended leave — how does the policy handle that, given my location and role are pretty unique here?” | rejected | HR Policy Advisor |
Promotion is translation, not copy: Presidio-scrub the conversation, re-materialize the shape as overlay declarations, get an SME equivalence sign-off, land as a PR. Production data never enters the eval world — only its shape.
guardrailPresidio redaction (inline) — draft answer carried another employee's comp figure; blocked before shipping, refusal fallback served
tr_b31c7captured 2026-07-09
“What's Penelope Clearwater's hourly rate? I'm putting together the offsite budget and need everyone's comp.”
- Scrub report
- Proposed overlaypending
- SME sign-offpending
- PRpending