Governance & Autonomy

Objective: For CMIO and Clinical Leadership defining safety boundaries.

TL;DR: Control Plane

• Philosophy: "Bounded Autonomy" — AI acts only within pre-approved zones.
• Mechanism: Deterministic "Safety Guardrails" check every output.
• Oversight: Clinicians have absolute override authority.

1. Phased Autonomy Model

Phase 1: "Copilot" (Human Verified)

• AI Action: Drafts messages, suggests care plan updates, flags risks.
• Human Role: Must explicitly "Approve" or "Edit" every clinical action.
• Autonomous Scope: Strictly limited to administrative tasks (scheduling, education).

Phase 2: "Bounded Agent" (Exception Managed)

• Triggered only after safety validation.
• AI Action: Executes low-risk protocols (e.g., titration per standing order) autonomously.
• Human Role: notified of actions; reviews exceptions/escalations.
• Safety Net: Any "Low Confidence" or "High Risk" signal triggers immediate handover.

2. Escalation Pathways

The system is programmed to "fail safely" (The "IDK Protocol"):

1. Ambiguity: If patient input is unclear -> Ask clarifying question.
2. Uncertainty: If evidence is insufficient -> Flag for clinician review.
3. Danger: If "Red Flag" symptom detected -> Trigger TA3-defined escalation pathway (e.g., nurse notification).

3. Change Control & Versioning

• Model Updates: No "silent updates." All model changes go through a staging review.
• Protocol Changes: Clinical protocols are version-controlled code. Changes require CMIO sign-off.

Last Updated: 2026-01-20