For Healthcare Organizations Deploying AI

Runtime evidence for clinical AI deployments.

Ambient scribes, CDSS, clinical chatbots, agent workflows — the AI committee approved them, but no one can show what they actually do at runtime. Glacis runs inside your infrastructure, instruments one deployed workflow with runtime controls, and produces signed evidence your AI committee, HIPAA security officer, and 42001 auditor can verify — with zero clinical payload egress.

Three weeks, one deployed clinical AI workflow, zero clinical payload egress.

The reality

The gap between AI committee approval and runtime reality.

Clinical AI is already in your environment. In the ambient scribe your clinicians use. In the CDSS your radiology team relies on. In the chatbot your front desk handed to patients. In the agent workflows procurement approved two quarters ago.

Most oversight today is paper — AI committee minutes, vendor questionnaires, quarterly attestations. None of it shows what your deployed workflow actually does on a Tuesday afternoon, or which guardrails fired when a clinician asked the model an off-label question.

The Sprint closes that gap on one deployed clinical AI workflow in three weeks — runtime controls instrumented in your stack, signed evidence receipts, and an evidence pack your AI committee and HIPAA security officer can verify.

How the sprint runs

Inside your infrastructure. On one deployed workflow.

Glacis runs beside your existing safety stack — instrumenting runtime controls, signing each control outcome, and packaging evidence your AI committee and auditors can verify without seeing protected health information.

Runtime controls in your stack

PHI redaction, consent verification, scope-of-use limits, content filtering — the controls your AI committee specified now execute and emit signed evidence at runtime.

Zero clinical payload egress

PHI, clinical notes, prompts, and responses stay inside your environment. Only signed hashes and verification metadata leave — designed to minimize BAA scope.

Evidence pack for the AI committee

Timestamped, third-party notarized, cryptographically signed receipts assembled into a packet your AI committee, HIPAA security officer, and external auditors can verify on their own.

What the auditor gets

Receipts at runtime. Pack on demand.

Each control execution emits a signed receipt. Receipts assemble into evidence packs an AI committee or 42001 auditor can verify against the issuer’s public key, offline.

SDK · TypeScriptWrap a clinical control.

import { attest } from '@glacis/runtime';

const receipt = await attest({
  workflow: 'scribe.note',
  policy:   'health.scribe.phi.v3',
  decision: 'REDACT',
  rules:    ['phi.redact', 'consent.ok'],
});

// → signed OVERT receipt; PHI never leaves the box

AI committee packet · OVERT 1.0What the auditor inspects.

Evidence PackAmbient Scribe · Q2 2026
Verified
Issuer
did:web:notary.glacis.io
Workflow
scribe.note · deployment v2.1
Receipts
4,812 control executions
Redactions
1,103 PHI redact events
Schema
overt://schema/v1.0/runtime-attestation
PHI egress
None · payloads hashed locally
ED25519 · ed25519-2026-q2 · chain depth 4,812

What changes for your team

The same clinical workflow. New evidence behind it.

For your compliance team

Stop reconstructing what happened from logs and interviews. Every AI interaction that passes through your controls generates verifiable evidence automatically. Audit prep becomes report generation, not archaeology.

For your board

Answer “how do we know our AI is safe?” with evidence, not assurances. A dashboard of verified control executions, not a policy document. Confidence backed by cryptographic proof.

For your clinical teams

No workflow changes. Glacis observes your existing controls — it doesn’t replace them. Your clinicians keep working exactly as they do today. Evidence generation is invisible to end users.

For your regulators

Give them what they actually want: proof that your governance isn’t just documented, it’s operational. Evidence they can verify without trusting your word. Third-party verifiable, not self-attested.

Fail-closed by default

If consent hasn’t been verified, the request doesn’t proceed. If PHI detection fails, the pipeline stops. Glacis enforces your policies — it doesn’t just report on them.

Built for what’s coming

EU AI Act classifies most clinical AI as high-risk. Colorado, Texas, and CMS are watching. Regulators want evidence governance happened — not documentation it was planned.

Architecture, not policy

Zero clinical payload egress.

PHI, clinical notes, prompts, and responses are hashed locally inside your infrastructure. Only signed hashes and verification metadata cross the boundary.

Patient data, PHI
Never transmitted
AI prompts, responses
Hashed locally only
Clinical notes
Never transmitted
Cryptographic commitments
Yes (no PHI)
Designed so Glacis never has access to plaintext PHI · confirm against your specific HIPAA analysis

Related guides

Healthcare-specific framework guides.

Scope a clinical AI evidence sprint on one deployed workflow.

Fixed scope. Three weeks. One deployed clinical AI workflow. Runtime controls instrumented inside your stack, evidence mapped to HIPAA, the AI committee charter, and emerging state AI laws. No clinical workflow disruption.

Or learn how continuous attestation works →