Trust & Governance · Published Validation
Three publications in one quarter.
One architecture, validated.
Between March and June 2026, three independent research groups published the empirical case for the physician-governed AI architecture in the HarnessHealth provisional patent draft. Two of them appeared in Nature on the same day.
The papers describe the problem (autonomous AI without an attestation gate) and the safe behavior (advisory output with physician review). They do not describe the specific architectural mechanism that interposes physician attestation between AI action and system execution. That is the harness.
The three publications.
Each describes a different point on the same architectural spectrum — from advisory at one end, to autonomous-action at the other. The architecture in the harness provisional patent draft lives in the middle: physician attestation as a hard, structural gate between AI output and system execution.
THE GAP · Autonomous action without an attestation gate
Nature · 17 June 2026
MIRA — Towards autonomous medical artificial intelligence agents.
Ferber et al. · Kather lab · doi.org/10.1038/s41586-026-10675-5
A GPT-4o agent given 11 clinical tools and over 85,000 possible action choices, operating inside a sandboxed FHIR-native digital health record on 574 real emergency department cases. It placed medication orders, ordered imaging and labs, recommended procedures, and triaged hospital admissions — autonomously, with no physician attestation gate between AI decision and system execution. Diagnostic accuracy 87.8% versus 78.1% for board-certified physicians; procedure ordering 53.5% versus 38.3%; guideline adherence 35 percentage points higher; zero critical medication safety errors across 468 orders.
The paper demonstrates that autonomous medical action is technically real today. The sandbox constraint is a research choice, not a technical barrier. The same agent pointed at a live EHR tomorrow is the regulatory and liability exposure the FDA January 2026 CDS guidance and EU MDR Article 14 assume must be governed — without specifying the architectural mechanism that does the governing. That mechanism is what the HarnessHealth provisional patent draft claims.
On the paper's publication day, its senior author described what his lab built as what the field now calls an “agent harness” — “what we build for clinical agents, long before it was a thing.”— the paper's corresponding author, on publication day, 2026
THE SAFE SIDE · Advisory only, research-only deployment
Nature · 17 June 2026
AMIE Longitudinal — Towards Conversational AI for Disease Management.
Liévin, Palepu, Weng, et al. · Google Research / DeepMind · doi.org/10.1038/s41586-026-10764-5
A two-agent Gemini system — a patient-facing dialogue agent and a long-context management reasoning agent that grounds every plan in 600+ tokenized clinical guidelines and dual-jurisdiction drug formularies. Across 100 multi-visit simulated cases and 21 board-certified primary care physicians, AMIE matched or exceeded physician performance on every management evaluation axis. Management plan appropriateness at Visit 3: 98% vs 81%. Treatment preciseness: 95% vs 67%. Explicit guideline references in 100% of recommendations vs 86% for physicians. Medication reasoning on harder questions: 57.9% vs 47.8%.
Google's stated deployment posture: research-only, not for clinical use. AMIE represents the safe end of the spectrum — physician-level reasoning that stays advisory and would pass through an attestation gate cleanly. The published 100% guideline citation rate is the quality bar the harness architecture lets physicians enforce on every AI output that reaches a patient.
THE EARLY COMMERCIAL PROOF · Physician-in-the-loop in production
Lotus Health AI · 12 May 2026
Physician agreement with AI-suggested care actions in asynchronous primary care.
Stark, Mani, Dhaliwal et al. · Lotus Health AI · lotus.ai
Lotus Health AI is a 50-state-licensed medical practice in which every AI-generated clinical decision — diagnosis, prescription, lab order, referral — is reviewed and finalized by a board-certified physician before it reaches the patient. Its Series A ($35 million, closed February 2026) was led by top-tier venture firms alongside a former U.S. Chief Technology Officer and an OpenAI executive. The May 2026 research paper studies physician concordance with the AI's suggested actions — the validation track for the broader physician-in-the-loop architecture.
Lotus operates the architecture in the HarnessHealth provisional patent draft, in primary care. The harness extends the same primitive across surgical encounter coding, prior authorization, Letters of Medical Necessity, registry abstraction, and clinical documentation — anywhere AI output must be billable, signable, and defensible. Different clinical context. Same architectural mechanism. Same regulatory logic.
What the harness adds
The papers describe the gap.
The patent draft claims the mechanism.
MIRA shows an autonomous medical AI agent placing orders without an attestation gate. AMIE shows advisory AI that stays out of execution because Google explicitly limits it to research. Lotus shows a licensed medical practice manually reviewing every AI suggestion before it reaches a patient. None of the three describes the architectural mechanism that makes physician attestation a structural, non-bypassable property of the system itself.
The HarnessHealth provisional patent claims that specific mechanism: an AI orchestration layer in which clinically meaningful output cannot exit the system without a named physician's cryptographic attestation, with volume guardrails to prevent rubber-stamping, NPI-bound timestamps for every signed output, and a hash-anchored audit chain that lets a regulator, payer, or plaintiff reconstruct exactly who attested what, when. The mechanism, not the policy.
The reason the architecture matters now is that autonomous medical AI is moving from research into attempted deployment faster than the regulatory frameworks that govern it. The FDA January 2026 CDS guidance assumes physician oversight without defining the architectural mechanism to enforce it. The EU AI Act 2024/1689 mandates human oversight for high-risk AI under Article 14 without specifying implementation. The harness is the implementation.
01
Provisional patent (drafted): Physician-Governed AI Harness with Hard Intercept
Provisional. Mechanism, not policy.
Read documentation02
Attestation network: forming
Named physicians attest each output. NPI-bound timestamps.
Read documentation03
Audit chain: hash-anchored, externally verifiable
Hashcare. Tamper-evident by construction.
Verify on hashcare.comMarket convergence · 2026
The whole industry is building this stack — and stopping at the same line.
In 2026 the front half of clinical AI was commoditized from every direction. The sensing, the interpretation, the de-identification, the builder — and now the safety harness itself — each shipped for free, open, or on-device. Every player stops at the identical line: the point where a licensed human takes responsibility for the output. That line is not a harder model, and it is not a better harness. It is a medical license, and a commons cannot hold one.
macOS 27 ships a Foundation Model in the terminal — clinical reasoning, on-device, free.
Stops at: No license on the output.
SensorFM: a foundation model on 1 trillion minutes from 5M people; its agent matched clinicians.
Stops at: Clinicians rated it — none signed it.
9.4M installs, Apache-2.0, on-device everywhere; PII models took 1st and 2nd on an independent benchmark.
Stops at: No receipt, no liability.
An “Agent Factory” previewed at HIMSS26 — every health system gets a canvas to wire AI workflows.
Stops at: The protocol knowledge has to come from somewhere.
40+ bills in 2026: Delaware bars AI from licensure, Texas mandates human EHR review, California requires disclosure.
Stops at: They mandate oversight without an implementation.
A physician-built neurosymbolic stack — symbolic retrieval before every LLM call — took a small model from 30.2% to 95.8% safety-intervention on psychosis-bench. Allowed US patent. Benchmarks published openly.
Stops at: It “never gives definitive diagnoses.”
An open model run inside the firewalls of 300+ hospitals within weeks — the fastest clinical-AI rollout ever attempted. Its own physicians warned in JAMA that the model’s tendency toward “plausible but factually incorrect outputs” carried substantial clinical risk.
Stops at: The liability analysis lands the damage claim on the treating doctor, not the model.
A named physician signs the output (NPI-bound, hash-anchored); the raw PHI provably never left the device.
Stops at: This is the line. This is us.
Every layer of the stack is being built and given away — the model, the sensing, the de-identification, and now the safety harness itself, patented and benchmarked in the open. Except one. A commons can build a harness. It cannot hold a medical license or assume the liability. What cannot be copied is not the harness — it is the signature inside it.
The macro proof · Stanford AI Index 2026
The most-cited independent measurement in AI puts numbers on the same line.
The convergence above is the qualitative story. Stanford's AI Index 2026 — the field's most-referenced independent report — makes it quantitative: capability and adoption are racing to free, while the real-world, judgment-laden, attested layer stays scarce. Every figure below is Stanford's or the FDA's, not ours.
Only 2.4% were backed by randomized-trial data. The regulator is clearing AI faster than the evidence is being generated.
The Stanford–Harvard ARISE Network reviewed 500+ studies; the rest ran on exam-style questions. Its conclusion: AI works best supporting — not replacing — clinician judgment.
The top-scoring AI agent completed under 70% of real EHR tasks; the report calls the evidence base for reliable autonomous agents “early-stage.” Autonomy isn’t the product yet — attested judgment is.
The models are cleared and adopted. What stays scarce is the attested, judgment-laden, human layer at the bottom of the stack — the medical license on the output. That is the line every player stops at, and it is the harness.
Source: Stanford HAI, The AI Index 2026 Annual Report — Medicine Ch.6 (FDA device data; Stanford–Harvard ARISE Network State of Clinical AI Report and MedAgentBench cited therein). Figures cited with attribution; no charts reproduced.
The other proofs show AI isn't ready for autonomy. This one shows why a human in the loop isn't enough either.
In a study of 20,302 primary-care checkups, visits documented by an ambient AI scribe recorded more psychiatric symptoms than human notes — and were followed by less depression treatment. A clinician was present at every one of those visits. The complete-looking note did not ensure the right action followed. That is the gap between reviewing a fluent artifact and being accountable for a decision — the exact line the harness draws.
Approve is a click. Attest is a record.
Source: Castro, McCoy & Perlis, Psychiatric Documentation and Management in Primary Care With Artificial Intelligence Scribe Use, JAMA Psychiatry, 2026. Composite depression outcome = any of diagnosis, antidepressant prescription, or behavioral-health referral. Figures cited with attribution.
Medicare just proposed to pay for clinical AI.
In the CY2027 outpatient rule (CMS-1850-P, July 2026), CMS created status indicator O1 — “Software as a Medical Service” — the first Medicare payment category for algorithm-driven clinical care, paid on its own.
Payment legitimizes the category — and drags audit and liability behind it. A clinical-AI output Medicare pays for is one a regulator can review and a court can weigh. That doesn't make the output safe. It makes the layer that keeps it accountable — a named physician, an NPI-bound attestation, a tamper-evident record — the thing every paid deployment now needs.
Off-the-shelf models are unsafe for unsupervised clinical use.
Stanford AI Index 2026 · PrIME-LLM (JAMA) →
First state suspension of autonomous AI prescribing.
Utah Medical Board → Doctronic, Apr 2026 →
First Medicare payment category for clinical AI.
CMS CY2027 OPPS · status indicator O1 / SaMS →
Three independent signals — the science, the regulator, the fee schedule — point to the same scarce layer: the licensed human attesting the output. That is the harness.
Source: CMS, CY2027 OPPS/ASC Proposed Rule (CMS-1850-P) — status indicator O1, comments due Aug 31 2026. O1 pays the facility for FDA-cleared software in the hospital-outpatient setting; the harness is the governance layer any paid clinical AI needs to stay auditable.
Who needs this surface.
Health systems
Deploying clinical AI is now a board-level decision. The harness is the compliance layer the FDA CDS guidance and the EU AI Act assume exists but do not specify. Pair it with any LLM stack.
Read morePhysicians
Your name on every signed output. Your NPI bound to every timestamp. The mechanism that lets you participate in AI-augmented care without absorbing the AI-vendor liability stack.
Read moreRegulators and payers
Verifiable, NPI-bound attestation records. Externally auditable hash chain. The exact accountability infrastructure both the FDA January 2026 CDS guidance and EU MDR Article 14 require — without naming.
Read moreDevelopers and AI infrastructure teams
The harness sits between any model layer (Claude, GPT, Gemini, open) and any clinical surface. Open architecture, attested output, audit chain by construction. Build on top.
Read more