Load a sample AI-drafted clinical output, run it through the review loop, and see what comes out: an automated check, a physician-grade sign / revise / decline, and an audit-anchored attestation record.
We ran synthetic HSA/FSA determinations locally through two capable open models — including MedGemma 27B, Google’s medical foundation model — under the instruction “never invent a code.” Both were excellent: correct verdicts every time, and they cited the right ICD-10 code whenever a real diagnosis was present. Then, once, a model did this:
VERDICT: DECLINE Diagnosis linkage: Intermittent low back pain (M54.5). Medical rationale applying the but-for standard: … the patient is not currently under a treatment plan and lacks a formal diagnosis on file …
M54.5 is the correct ICD-10 code — for a condition nobody diagnosed. The verdict was right; the linkage was invented. Fluent, plausible, and exactly the kind of sentence that sails through every automated check.
It didn’t happen again. That is the whole point: you cannot spot-check your way past a failure that surfaces roughly one time in fifty, at any temperature, with no warning. The only thing that catches it is a named human accountable for every determination — not a sample.
Synthetic cases only, local models, ~100 generations. Not a benchmark, and emphatically not a knock on any model — both were strong, and open medical weights are exactly why this layer matters now. When capability is this good and this cheap, the scarce thing isn’t accuracy on average. It’s an accountable record on the one draft that slips.
Across states and CMS, an AI output that affects care increasingly needs a licensed, accountable human behind it — and a record that proves it. This is the layer that produces that record.
Utah's Office of AI Policy runs a real regulatory sandbox — healthcare AI is its first focus. A physician-accountability layer is the kind of human-oversight mitigation it exists to bless.
Colorado's health-AI rules require a licensed clinician to review AI coverage denials (ADMT Act, in force Jan 1, 2027) — the accountable human isn't optional.
In force since 2025: a medical-necessity decision driven by AI must be made by a licensed, specialty-matched physician, not the algorithm.
Federal payment and coverage rules keep tightening around documented human review of AI-assisted determinations. The attestation is the artifact that survives an audit.
This demo runs on synthetic samples. In production, a named, specialty-matched licensed physician reviews your actual outputs under a BAA and signs, revises, or declines — with the audit-ready record attached. Tell us what you're building.
We only store what you enter here to follow up. ClinicalSwipe provides independent, licensed physician review of AI-generated clinical outputs; it does not practice medicine.