Architecture Sutter Health Healthcare Demo • NRC + AER + Preparedness

Align Legal Obligations, Frontier Safety Gates, and Interaction Governance

This architecture turns an OpenAI model into a governed institutional artifact generator—not a generic chatbot. OpenAI Preparedness controls what may be deployed; CRSP NRC/AER controls what may be released per interaction, with defeaters, required checks, and audit traces suitable for a Sutter Health-style review committee.

View the stack #overview NRC contract #user-llm Runtime flow #aer Sutter Health demo #cardiology

Layer 1 — Legal obligations

EU AI Act Defines prohibited practices, transparency duties, and high-risk obligations.

What must be proven What must be blocked What must be logged

Layer 2 — Frontier safety gate

OpenAI Preparedness Threat models + capability evaluation + safeguards before deployment.

Eligibility decision Deployment restrictions Monitoring expectations

Layer 3 — Interaction governance

AER

CRSP NRC + AER Per-output entitlement: ASK→EVIDENCE→PROPOSE→ENTITLE→RELEASE.

Defeaters Required checks Audit trace

NRC as user–LLM contract

Non-persona, institutional routing

What the contract changes

In a normal chatbot, the user asks and the model answers. In a governed system, the user selects (or is assigned) an NRC lane, which functions as a normative contract: it specifies allowed moves, forbidden moves, accountability anchor (e.g., CASRC), and defeaters. The model is only entitled to produce artifacts inside that contract.

Contract primitives: - Inferential scope constraint (allowed/forbidden moves) - Epistemic posture (first-order draft vs second-order risk analysis) - Accountability anchor (audience + review body) - Defeater activation profile (what invalidates release) Result: - Outputs become reviewable institutional artifacts, not “answers.”

Example NRC (Sutter-Health: misuse scan)

NRC.Sutter_Health.TYPEIII.RETEAM.RISK_SCORE.AUTHORITY Allowed moves: - identify_failure_modes - define_defeater_triggers - propose_governance_mitigations Forbidden moves: - clinical_advice - treatment_guidance Accountability: - Audience: AI Safety Review Committee (CASRC) Required checks: - audit_fields - monitoring_plan - review_cadence Defeaters: - AI-as-authority - missing uncertainty handling - missing monitoring

Runtime flow

AER orchestration + entitlement gating

AER pipeline (per interaction)

Phase	Function	Enforced by NRC
ASK	Bind NRC, confirm audience & data posture	Scope + prohibitions must be explicit
EVIDENCE	Surface assumptions & uncertainty	Uncertainty required for high-risk lanes
PROPOSE	Draft artifact within allowed moves	No forbidden move content allowed
ENTITLE	Run defeaters + required checks	Missing checks defeats release
RELEASE	Publish only entitled output	Audit fields attached; log persisted

Preparedness overlay (deployment gate)

Preparedness governs whether the model and deployment posture are acceptable for the domain. In this architecture it is a precondition for runtime calls: model eligibility, restrictions (logging/access), and required safeguard status are checked before AER proceeds.

Preparedness → enforcement points: - Model version pinned + allowed capability surface - Required safeguards present (logging, access control, refusals) - Red-team eval suite results recorded - Rollback plan + monitoring hooks established PF answers: “Can we deploy this capability?” AER answers: “Can we release THIS output?”

Sutter Health demonstration

Cardiology risk score governance, not clinical advice

Scenario

A toy cardiology risk score system emits a banded risk score with uncertainty flags (e.g., LOW_DATA, OOD). The goal is to produce a CASRC-ready misuse scan explaining how clinicians might over-trust the score, and what governance/UI/workflow mitigations prevent over-reliance.

Context artifacts (explicit): A) CardioRiskScore_Toy_v0 spec (toy, non-clinical) B) CASRC Charter (review body + responsibilities) Fixtures: F2: HIGH + LOW_DATA F3: MED + OOD

What “success” looks like

Requirement	Why it matters
No clinical directives	Prevents accidental protocol generation.
Defeater triggers	Explains when reliance is invalid.
Monitoring + audit fields	Operationalizes accountability.
Review cadence	Prevents drift blind spots.

Implementation notes

Minimal components for a demo build

Services

1) NRC selector + compiler 2) AER orchestrator (ASK/EVIDENCE/PROPOSE/ENTITLE/RELEASE) 3) Preparedness policy check (eligibility + restrictions) 4) OpenAI gateway (policy-enforced client wrapper) 5) Audit logger + review queue (CASRC)

Data artifacts

- NRC library (YAML) + compiled agent configs - CASRC charter artifact - Risk score toy system spec artifact - AER audit log (transcript + defeater hits) - Release artifacts table (versioned outputs)

Template invocation (demo)

NRC: NRC.Sutter_Health.TYPEIII.RETEAM.RISK_SCORE.AUTHORITY Artifact type: misuse risk analysis Audience: CASRC Sensitive data posture: none (hypothetical) Context artifacts: [CardioRiskScore_Toy_v0, CASRC_Charter] Fixtures: [F2, F3] Prompt: Analyze misuse risks when clinicians ignore LOW_DATA or OOD flags. Must produce: failure modes, defeater triggers, mitigations, audit/monitoring, review cadence.