Healthcare AI governance for regulated clinical workflows

Constitutional AI Governance for Healthcare

ClinicalGuard gives healthcare teams deterministic guardrails around LLM-powered clinical decision support with MACI separation of powers, tamper-evident audit logging, and built-in checks for PHI, drug safety, and adverse events.

20 constitutional rules 10 PHI identifiers detected API-key protected Docker and Fly.io deployable

Quick start

pip install acgs-lite

Embed constitutional rules in YAML, route decisions through MACI roles, and retain a chain-valid audit trail for every governance event.

Purpose-built for teams that need trustworthy AI oversight before recommendations reach clinicians, patients, or regulators.

The problem

Ungoverned healthcare AI creates compliance and liability exposure.

$1.5M

HIPAA fine exposure

A single PHI-handling failure can become a reportable event with severe financial consequences.

2

Regulatory pressure

Healthcare AI faces two oversight fronts at once: privacy obligations and FDA-style expectations for evidence, controls, and escalation.

0

Tolerance for unauditable decisions

If a recommendation cannot be reconstructed after the fact, the liability sits with the deploying organization.

How it works

Three steps to governed clinical AI.

Define

YAML rules

Encode healthcare policy as a 20-rule constitution covering evidence tiers, PHI handling, dosing, escalation, audit, and clinical safety.

Enforce

MACI separation

The system separates proposer, validator, and approver roles so no agent can self-approve or bypass governance.

Prove

Audit trail

Every decision is written to a tamper-evident chain with constitutional hash, timestamp, reasoning, and integrity verification.

Compliance

Built for security, auditability, and regulatory review.

Framework Coverage Operational evidence
HIPAA 9/15 auto-covered controls for healthcare AI governance workflows PHI detection, audit logging, access control hooks, de-identified workflows
FDA AI/ML Aligned with evidence-tiering, monitoring, escalation, and adverse-event review expectations Off-label routing, dosing checks, MedWatch-style adverse event signaling, human review gates
SOC 2 Ready for control mapping across security, change management, and traceability Tamper-evident logs, API-key auth, persistent audit storage, deployment hardening

Two-layer validation

Clinical reasoning plus deterministic constitutional enforcement.

ClinicalGuard combines an LLM layer for semantic clinical reasoning with a deterministic governance layer for what cannot be left to model judgment alone.

LLM clinical reasoning

Assesses evidence tier, drug interactions, dosing concerns, step therapy, and risk tier in free-text clinical proposals.

Deterministic constitutional rules

Enforces MACI validation, PHI restrictions, escalation logic, audit requirements, and reproducible rule checks with chain integrity.

Features

Purpose-built safety controls for healthcare AI.

PHI detection

Flags 10 HIPAA Safe Harbor identifier types including SSNs, MRNs, DOBs, phones, email, and device identifiers.

Drug interaction checks

Routes major interaction, contraindication, dosing, narrow therapeutic index, and step-therapy concerns before approval.

Adverse event detection

Detects severe allergic reactions, overdose, hospitalization, death, and other MedWatch-relevant safety signals.

Tamper-evident audit

Persists audit entries with constitutional hash and chain validation so investigators can verify nothing was altered.