For your health — patient or professional
THE HARNESSFOR YOUR HEALTH.
Any model. A physician behind it. Yours by design.
You already use AI for your health — asking about a symptom, a lab, a medication. So do your doctors, drafting the note and the prior auth. The harness is the layer that makes it accountable: your health conversations stay yours, and a licensed physician stands behind anything that matters.
The mechanism
HarnessHealth in 15 seconds
- 01Your AI drafted it. Who signed?CMS WISeR: no denial may be made solely by technology.The gap
- 02HarnessHealth holds the draft.Until a named physician signs, revises or declines it.Held
- 03Then a receipt anyone can verify.NPI, timestamp and document hash, anchored on hashcare.com.Receipt
- 04Run a sample in the sandbox.Synthetic cases, in your browser. Nothing sent or stored.Sandbox
- 05HarnessHealth.The layer that makes health AI accountable. harnesshealth.aiThe layer
Sandbox and demonstration today; no production review yet. Run a sample on the pilot page →
Two doors, one layer
Whoever you are, there’s a door for you.
Everyone’s health runs through AI now — the patient asking questions at midnight, the clinician drafting notes at scale. The harness is the same layer underneath both. Start where you fit.
Your doctor, in your pocket.
Ask about a symptom, a lab, a medication. The harness keeps the conversation yours — not training data for a company you'll never meet — and puts a licensed physician behind the answers that matter.
iPhone: next. Harness Desk runs on your Mac today.
- Start here: check an AI summaryPaste your original and the AI’s summary. It flags a changed dose or a dropped warning, in your browser.
- What preventive care is due for youThe USPSTF and CDC tables for your age, in your browser. A sourced list, not an order; bring it to your clinician.
- Pocket, on your iPhone and MacYour health alerts read on your own devices; nothing leaves unless you share it
- Why your AI health chats aren’t privateA 3-minute read that will change how you use ChatGPT
- Find a licensed physicianEvery U.S. physician, searchable — no signup
- Plan care for yourself or a parentWorker-owned care for aging at home
- Keep your family’s papers togetherThe ComfortCard family file, on your device
- See the health dashboard demoA demonstration with sample data, not your records
Or tap Ask chanio (bottom-right) to ask a question about this site now. It is an AI assistant, not a clinician.
Your clinic, in a swipe.
Bring any model your team already uses. Every clinically meaningful output is attested by a named physician — NPI-bound, hash-anchored, auditable — before it reaches a patient or a payer.
Sandbox open now. Paid production reviews when the queue goes live.
- Start a pilotSandbox today. Production terms are set in the pilot. A BAA before any patient data.
- See the harness workThe Fidelity Gate: watch it hold an unsafe summary
- Why the law wants a human nowWISeR, seven 2026 state laws, the FDA paper — with sources
- How it works, in 90 secondsDraft → signed → shipped, plainly
- Score your AI governance12 questions, a scorecard you can keep
- Review from your licenseThe pilot runs on synthetic cases
“You need a harness to orchestrate the whole thing.”
— Jensen Huang, CEO NVIDIA · June 2026What we build
Four parts of one stack.
- 01A named physician signature on each consequential AI output, from a physician licensed where the patient lives who is free to declineThe signature
- 02A peer-review marketplace with named cliniciansClinicalSwipe
- 03An audit chain that links every AI output to the physician who attested itHashcare
- 04Domain-tuned agents for surgical practice, registry abstraction, and clinical documentationSurgeonValue
Together: the operational stack that makes physician-attested AI deployable at scale.
The problem
AI in healthcare has a governance problem.
Three failure modes are driving institutional anxiety about AI deployment.
Unsigned clinical output
AI drafts a care plan, a prior authorization letter, or a medication recommendation. Nobody reviews it. It reaches the patient. When something goes wrong, the liability chain is unclear.
Where the law now requires a signature →Credential laundering
An AI model trained on medical literature generates confident clinical language. No NPI number is attached. No licensed physician reviewed it. The clinical authority is borrowed from training data, not from a real doctor.
How an NPI gets bound to an output →Fragmented identity
A patient engages with AI across six different health apps. There is no unified identity, no shared medical context, no way for a physician to maintain a coherent longitudinal view of that patient's health journey.
Who owns the conversation →This isn't only our read. The World Economic Forum's 2026 AI at Work names accountability — not capability — the binding constraint on AI adoption in regulated industries, and economist Dan Davies calls the failure mode the “accountability sink”: a decision produced at machine speed that no one can be held to. The harness closes that gap — every clinically meaningful output is held until a named physician signs it, or it never ships.
Where the harness sits
The records already reach Claude. The output isn’t governed.
Anthropic and connectors like HealthEx already bring a patient’s consented records into Claude — securely, over the federal TEFCA network, via MCP, never used for training. That solves the input. What stays unsolved is the output: when the model interprets a lab, drafts a care plan, or suggests a medication, who signs it?
Consented records in
A patient connects their full history from 50,000+ provider organizations to Claude. Consent-scoped, never trained on. The input layer is solved — securely, over the federal TEFCA network.
See the connectorThe model reasons
The record becomes context. Claude can summarize it, explain a lab, draft a routine. Powerful — and, on its own, ungoverned for clinically meaningful output.
The published evidencePhysician-governed output
Before clinically meaningful output reaches a patient, a licensed physician attests it — NPI-bound, timestamped, audit-anchored. The layer the input pipes do not provide. This is us.
How the gate worksHealthEx put the patient back at the center of their data. The harness keeps a licensed physician at the center of the decision.
The architecture
The harness is the governance layer you don’t see.
Four components. One governed infrastructure. Each card says what runs today and what is still in pilot.
Independently shown at the frontier — 2026
An open agent harness running GPT-5.5 ranked first on the Frontier SWE benchmark, ahead of frontier models in their own harnesses. The same model in its native harness ranked fifth; only the control loop changed. The lesson: independent verification, revisable planning, and disciplined stopping beat raw model strength on work that runs for days. But the control loop itself is now a commodity — every major model vendor ships an autonomous agent that will run a workflow end to end for a monthly fee. What does not commoditize is who the verifier is. Here it is a licensed physician, and the gate is attestation — the one part no vendor can rent you. Zenith, Frontier SWE →
NPI Registry
Every U.S. physician, live from CMS
Any NPI renders a page at harnesshealth.ai/dr/[NPI] on demand from the CMS NPPES registry — nothing is pre-indexed or stored. A physician claims a page by verifying the NPI and binding a signing key at /provider. No signup for a page to exist.
Search itAttestation Engine
AI drafts. A physician attests. Cryptographic record.
Clinical output is held as a draft until a named physician signs it — NPI-bound, hashed, timestamped, and verifiable by anyone on the record’s public /verify/[id] page. Live as a working demonstration; the production review queue is in pilot.
Sign and verify a document yourselfThe site guide
chanio answers here from one shared fact file
One script loads the site guide. It answers from /kb/registry.json, the facts every site guide in the network is written from, and hands people to the right sibling site. It runs on harnesshealth.ai today; other sites are a pilot conversation.
<script src="https://harnesshealth.ai/harness.js" defer></script>API Gateway
Every AI call is governed.
solvinghealth.com/api/chat — the governed Claude bridge. TLS-encrypted transport. Draft-until-attested gating. FHIR-aligned connectors (the Redox sandbox loop on SurgeonValue is the only live one). The intelligence layer is proprietary.
Programmatic access is a pilot conversationFour entry points
The same infrastructure. Four entry points.
Who are you? Each audience gets a different door into the same governance layer.
01For Health Systems
Deploy physician-governed AI with a documented governance framework and a BAA offered to qualifying covered entities. Start in the sandbox; no EHR integration is live.
Evaluate for your health system02For Physicians
Review AI-generated clinical documents between patients , paid the same whether you sign or decline, once live cases open; the pilot runs on synthetic cases. Your NPI is the credential.
See how attestation works03For Developers
Open pipes, signed outputs. FHIR connectors, EHR scaffolding and PROM tools are open code (Apache 2.0). Review routing, receipts and the MCP server run hosted.
Read the developer docs04Trust and Governance
Draft-until-attested gating. The flat per-review fee, the same whether the physician signs or declines. The governance documentation your legal team needs.
Review the governance modelSystem readiness
Live infrastructure, not a roadmap.
What is deployed and running today. What is in process. The honest accounting is a governance signal, not a hedge.
- The Fidelity Gate: check an AI summary against your original, in your browser
- Every U.S. physician searchable via the CMS NPI registry
- An MCP server for physician-governed AI tools
- Physician search at harnesshealth.ai/find
- chanio, the site guide, on harnesshealth.ai
- Patient PROM check-in (SMS delivery not yet live)
- EHR integration: designed, not built — the Redox sandbox loop on SurgeonValue is the only live connector
- BAA: offered to qualifying covered entities
- Production review queue: in pilot on ClinicalSwipe — the sandbox is open now
“The moat is whoever first connects grounded clinical evidence to native multimodal output, real workflow extensibility, and physician-earned trust.”
Techy Surgeon Substack, “Clinical AI Faceoff”, April 2026
That is what we are building.
Who’s accountable
A trust company signs its work.
Every attestation in the network traces to a named, NPI-verified physician — starting with ours.
Blaine Warkentine, MD
Founder
MD and clinical translator — two decades at the intersection of medicine and software, including growing BrainLAB's orthopedic vertical to $250M. Builds the harness in public.
blainomd.comMedical Director
Licensed where the patient lives
The physician whose NPI-verified signature goes on a Letter of Medical Necessity or a clinical attestation, or who declines to sign it. Licensed in all 50 states and DC; each case is checked against the patient’s state before anything is recorded or charged.
How the signature worksQuestions a form can't answer? Use the BAA + evaluation request form.
The governance layer your AI deployment is missing.
Start with a BAA request or connect Claude to the harness. No sales call required.
We publish our system prompts, attestation rules, and audit log schema. Trust comes from showing the work.
HarnessHealth provides AI governance infrastructure for healthcare, not medical advice. Clinical outputs are designed to be held until a licensed physician signs or declines them. Consult a qualified healthcare provider for diagnosis and treatment.
HarnessHealth is an independent company. The harness is built on Claude, Anthropic's AI model, and is not affiliated with, endorsed by, or operated by Anthropic. The engine is theirs — the governance, the physicians, and the accountability are ours.