RATSe - Health (Resilient AI Trust Score - Health)

RATSe-Health Runtime Governability Architecture for Clinical AI

Deterministic Control of Probabilistic Clinical AI Systems

Author: Dr. Sharad Maheshwari MD

Institute for Responsible Healthcare AI

Operationalizing governability for high-impact clinical AI systems. Binding global standards to clinical liability and patient outcomes through deterministic runtime control.

Disclaimer: This artifact is for research and governance synthesis purposes. It is not backed by any government organization or regulatory body. Not a certification authority. Not a regulatory body. Not legal advice. The content is an independent sectoral binding profile based on the RATSe framework of BeResponsibleAI & the Institute for Responsible Healthcare AI.

Governance Doctrine

Governance is deterministic control over execution.

Accuracy is statistical.

Safety is structural.

Governability Stack

1. PRIME

🛡️

Admission Control

2. RATSe Runtime Governance

⚙️

Deterministic Control

3. RATSe-Health Profile

🏥

Clinical Binding

4. The Patient

❤️

Immutable Foundation

Admission Control

PRIME

Architecture Compression Flow

Inference

→

ERL

→

DVEE

→

Sidecar Governance

→

Execution

→

Outcome Feedback

→

Governance Adaptation

Core Governability Engines

The structural mechanisms that enforce deterministic safety bounds in production environments.

ERL

Epistemic Reality Layer

Grounds probabilistic inference against objective truth constraints before action is permitted.

• ✔ Medical Ontology
• ✔ Clinical Evidence Base
• ✔ Institutional Policy
• ✔ Operational Reality (Resource Limits)
• ✔ Care Context (ICU, Triage, Rural)

DVEE

Decision Validation & Execution Eligibility

Determines the critical gating question: Can this inference participate in execution?

Inference → Validation → Admissibility → Execution

Safety

Agentic Safety

Governs multi-step, autonomous healthcare agents to prevent cascading clinical failures.

• ▪ Authority Control Plane (ACP)
• ▪ Traceability
• ▪ Orchestration Visibility
• ▪ Execution Boundaries
• ▪ Prevents "invisible workflow influence"
• ▪ Prevents authority escape

Control

Runtime Governance Sidecar

The operational control plane attached to the model. It sits outside the probabilistic core to enforce deterministic rules.

Authority Boundaries Runtime Policy Containment Safe Degradation

OGL

Outcome Governance Layer

Completes the continuous governance loop by structurally connecting clinical failures back to execution restrictions.

✔ Outcome Surveillance

✔ Drift Consequence Monitoring

✔ Governance Adaptation

✔ Execution Restriction Escalation

The 6 Foundational Pillars

Core dimensions of governability adapted for clinical AI.

Regulatory Cross-Walk

Mapping RATSe pillars to global regulatory frameworks.

Pillar	NIST RMF 1.0	EU AI Act	India DPDP Act	GDPR

IRHAI Policy Spine

The constitutional alignment required at clinical scale. These standardized doctrines mandate the boundaries for healthcare AI.

Core Mandate

Compliance is not an endpoint; it is a byproduct of deterministic technical control.

Understanding Governability Metrics

Select a governance mode below to observe how technical maturity directly influences the safety profile, latency of detection, and overall control of a clinical AI system.

🎭

Governance Theatre

Safety as Documentation. Compliance via checklists and meetings. Low technical control.

⚙️

Governance Competence

Safety as Infrastructure. Technical enforcement via code and sidecar monitoring.

01. Governance Maturity

In Theatre mode, maturity is "performative." You have an Ethics Board, but the AI is a black box that engineers cannot stop in real-time.

02. Detection Latency

This is the "Awareness Gap." In Theatre, you realize a model has failed weeks later after a patient is harmed. Latency = Months.

Runtime Control Radar

Detection Speed (Awareness)

© 2026 Institute for Responsible Healthcare AI

Inference may be probabilistic. Execution authority remains deterministic.