WFM Assistant

AGENTWorkforce ManagementL3Regulated-gradehigh-risk

Handles accruals, balances, and time-off requests, including transactional writes.

Workforce Core · Google ADK · production

Concern posture

native units from the latest sealed gate run — never a blended score

run_5522

NM44 / 44 checks exactAC9 / 9 asserts · confirmations presentFNIN_ORDER ✓ · 0.88 avg semanticFR2.3× refusal-rate gap

Maturity

level 3 · Regulated-grade

Portfolio mandate →

Numeric oracle + action coverage — run_5522 — numeric + side-effect asserts 100%

Calibrated judge bound — task-completion-judge@v3 · κ 0.87, certified

Composition pinned — 61 cases pinned on Calendar@2026.2

L4: Live enabled — watch windows + drift feed active

L4: continuous red-team — campaign running, no zero-tolerance findings

L4: governance evidence current — fairness gap open (RSK-004) — blocks L4

L3 → L4: governance evidence current ✗ — fairness gap open (RSK-004) — blocks L4. Mandate: target L4 by 2027-Q2 · production agents.

What gates it

suites, policy, and the latest sealed verdict

WFM Assistant — dev loop2 of 4 concerns hard-required · resolves to 61 cases

Gate

run_5522· 60 / 61· 2026-07-07 09:41Dev loop →

What watches it

production watch windows, drift monitors, guardrail parity

Open

Live

2 monitors · 0 drifted· drift monitorsLive →

Drift monitors (Evidently)

Retrieval groundedness drift· retrieval groundedness — accrual answersstable

Response-length drift· response length (token count)stable

Guardrail classifiers — deployed ⇄ offline parity (§56)

Llama Guard 4in-sync

Prompt Guard 2in-sync

Presidioin-sync

What attacks it

red-team campaigns, posture, open findings

Open

WFM Assistant — continuous watchcontinuouszero-tolerance breach1 open finding

Who judges it

LLM judges bound via this entity's suites — trust earned by calibration

Open

task-completion-judge@v3gate-eligibleκ 0.87vs ≥ 0.80false-pass 1.7%vs < 2%

What it costs

serving economics — always beside quality, never alone

Open

Cost / interaction

$0.019

hard

p50 / p95

1.2 s / 1.8 s

p99 2.6 s

Weekly volume

1.8M / wk

production

Projected annual

$1.78M / yr

at production scale

Its evidence

the governance profile and the ledger entries it produced

Compliance profile

Full compliance profile on file · risk owner P. Weasley · next review 2027-06-24 · open profile →

Recent ledger entries

2026-07-07 09:41run_5522 sealed — WFM Assistant dev loop · pass 60/61

2026-06-24 10:05WFM Assistant high-risk classification affirmed at annual review

2026-05-30 11:12task-completion-judge@v3 certified — κ 0.87, false-pass 1.7%