Compliance
Governance
Every artifact the platform already produces — sealed runs, judge certificates, dataset provenance, the override ledger — is the evidence. Governance is a read-side projection over the same entities: compliance as a byproduct of engineering, not a parallel paper trail.
Evidence
Generated by daily work
sealed runs · certificates · provenance · overrides
+ Control mappings
AAG-curated packs
control → evidence types + freshness, each with a why
= Compliance posture
Counts + explicit gaps
never a weighted score — gaps name the missing artifact
L1Onboarded
- ·`quality init` done
- ·Starter case green
- ·Dev-loop suite runs locally
L2Gated
- ·Release gate blocks merges
- ·Functional + safety concerns hard-required
L3Regulated-grade
- ·Numeric oracle + action coverage where applicable
- ·Calibrated judges bound (κ ≥ 0.80, certificate current)
- ·Dataset composition pinned
L4Continuously assured
- ·Live enabled: guardrails, watch windows, drift feed
- ·Continuous red-team campaign
- ·Governance evidence current — no open gaps
no agents here
Mandate schedule
dates are mandated top-down; levels are computed from real entities — a gate that gates nothing cannot reach L2| Agent | Current | Target | Deadline | What's blocking |
|---|---|---|---|---|
| WFM Assistant production agents | L3Regulated-grade | L4 | 2027-Q2 | L4: governance evidence current fairness gap open (RSK-004) — blocks L4 |
| HR Policy Advisor pre-production cohort | L3Regulated-grade | L3 | before production | on track — target criteria met |
| Payroll Copilot high-risk agents — no exceptions | L2Gated | L3 | before production | L3: calibrated judge bound comp-answer-judge@v1 shadow — κ 0.71 < 0.80 |
| People Navigator production agents | L1Onboarded | L2 | 2026-Q4 | L2: release gate blocks merges smoke suite advisory only — not wired as a required check |
| UKG Bryte 2.0 production agents | L2Gated | L2 | 2026-Q4 | L3: composition pinned routing utterance corpus refreshes weekly — not yet pinned |
| Benefits Explainer production agents | L2Gated | L2 | 2026-Q4 | L3: calibrated judge bound no domain judge — benefits-answer judge not yet drafted; retrieval floors only |
| Shift Swap production agents | L1Onboarded | L2 | 2026-Q4 | L2: release gate blocks merges dev-loop advisory only — not wired as a required check |
| Candidate Screener high-risk agents — no exceptions | L0Unmanaged | L3 | before production | `quality init` done not onboarded — no quality assets exist |
AAG provides, day one
- ✓CLI + portal, day one
- ✓Concern → backend routing table (versioned, explainable)
- ✓Grader packs + judge templates
- ✓Shared calibrated judges with certificates
- ✓Substrate registry + shared overlay library
- ✓Attack packs for red-team campaigns
- ✓Control mappings per framework (this catalog)
The team owns
- →Cases and suites, in the team's repo
- →Product-specific overlays
- →Thresholds above the AAG floor
- →SME time for domain judges
- →Fixing what the evals find
Non-goals, both ways: AAG never writes team cases; teams never run judge infrastructure.
Rollout is driven top-down with dates (§81) — mandated maturity targets per cohort. The compliance-theater risk is mitigated by construction: a level is a bundle of checkable criteria over real entities, and minimum-viable gaming shows up in the evidence ledger.