Healthcare & life sciences • customer story

Sutter Cardiology — Patient education drafting with governed release

A Type-III (CRSP-governed) deployment that accelerates patient education content creation without crossing into diagnosis, treatment recommendation, or clinical decision authority. Outputs remain drafts for review, structured to be audit-ready under an Ask → Evidence → Release (AER) control loop.

Context

Cardiology teams needed consistent patient education materials that were accurate, comprehensible, and compliant with institutional policy—without implying clinical decisions or individualized medical advice.

Patient-facing content Non-diagnostic posture Human clinician authority

Type-III wedge

Start with drafting (safe to approve), then expand into an enterprise governance layer for additional clinical AI tools. Patient education is an ideal wedge: high volume, high scrutiny, and clearly bounded.

Land → Expand Documentation-first Audit survivability

Boundaries (non-negotiable)

  • No diagnosis. No patient-specific assessment. No triage.
  • No treatment recommendation. No medication or dosing suggestions. No “what you should do.”
  • Draft-only posture. Outputs are proposed language requiring clinician and compliance review.
  • Explicit accountable roles. Content must name the human reviewer/approver roles, not imply autonomous authority.
  • Over-reliance mitigation. Language discourages treating drafts as authoritative clinical guidance.
Operational rule: The system is allowed to draft educational content and supporting rationale, but is not allowed to decide, instruct, or personalize clinical actions.

CRSP architecture (what is governed)

CRSP (Coherent Reflexive Super-Prompt) treats governance as a first-class artifact: prompts, role boundaries, policy controls, evidence requirements, and release criteria are versioned, reviewable, and logged—so the organization can answer “why did the system produce this text” with traceable controls.

  • Context bundle (versioned): domain posture, prohibited outputs, role boundaries, risk clauses, required citations, tone constraints.
  • Policy layer: what the agent may draft, what it must refuse, and what requires escalation.
  • Evidence rules: allowed reference sources (e.g., institutional guidance, approved education handouts) and freshness expectations.
  • Defeaters: conditions that block or downgrade release (e.g., patient-specific claims, dosing, urgent care directives).
  • Audit log: immutable record of inputs, policy version, evidence references, and release decision.

AER gate: Ask

The request is normalized into a bounded drafting task: educational content template, audience level, required disclaimers, and prohibited advice.

{ "ask": { "domain": "SUTTER_CARDIOLOGY", "task": "Draft patient education content (non-diagnostic)", "audience": "patient", "reading_level": "8th_grade", "must_include": [ "general information only", "seek clinician guidance", "emergency warning language" ], "must_not_include": [ "diagnosis", "patient-specific advice", "medication changes", "dosing" ], "accountable_roles": [ "Cardiology Clinician Reviewer", "Clinical Compliance Reviewer" ], "policy_version": "crsp.sutter.cardiology.v1" } }

AER gate: Evidence

Evidence is gathered only from approved sources (institutional education library, clinical policy excerpts, approved public references if allowed). Evidence is stored as a bundle and linked to the draft.

{ "evidence": { "bundle_id": "evb_2026_01_24_001", "sources": [ {"type":"institutional_handout","id":"SUTTER_EDU_AFIB_2025_10","freshness_days":106}, {"type":"policy_excerpt","id":"SUTTER_PATIENT_EDU_STYLE_GUIDE_v3","freshness_days":420} ], "notes": [ "Exclude any patient-specific interpretation", "Use neutral language; avoid directives" ] } }

AER gate: Release

The draft is generated with required disclaimers, explicit boundaries, and structured “review hooks.” If defeaters are triggered, the release is blocked or downgraded (e.g., return a refusal + safe alternative such as “suggest talking to a clinician”).

{ "release": { "status": "APPROVE_FOR_REVIEW", "output_type": "patient_education_draft", "includes": [ "general educational overview", "questions to ask your clinician", "when to seek urgent care", "non-advisory disclaimer" ], "defeaters_checked": [ "no_diagnosis_claims", "no_dosing_or_medication_changes", "no_patient_specific_instructions" ], "review_required_by": [ "Cardiology Clinician Reviewer", "Clinical Compliance Reviewer" ], "audit": { "policy_version": "crsp.sutter.cardiology.v1", "evidence_bundle_id": "evb_2026_01_24_001", "timestamp_utc": "2026-01-24T00:00:00Z" } } }

Defeaters (block or downgrade)

These conditions prevent release or force a safe rewrite:

  • Patient-specific tailoring (“Based on your symptoms…”)
  • Diagnosis language (“You have…”)
  • Treatment directives (“Start/stop/change medication…”)
  • Urgent triage commands (“Do X immediately…”) without clinician handoff language
  • Unsupported clinical claims not grounded in approved evidence bundle

Mitigation language (release-safe)

Built-in phrasing patterns that reduce misuse and over-reliance:

  • Draft-only framing: “This is general information to discuss with your clinician.”
  • Authority boundary: “Your care team is the final decision-maker.”
  • Escalation path: “If symptoms worsen, contact your clinic / seek urgent care.”
  • Question prompts: “Ask your clinician about…” instead of “You should…”