ClinicalSwipe runs independent review of clinical AI by licensed, NPI-verified physicians — sampled, documented, and signed. Reviewers are paid per review, never by outcome, and no decision enters the record without documented clinical findings.
Three programs, one protocol. Built for the rule before the rule.
Request a scoped audit Or scan your own copy free →Eli Lilly took a strategic position in Abridge — the most-funded clinical copilot — as it expands into billing and payer workflows. When the intelligence has a sponsor, the signature that proves independence has to come from somewhere else. (STAT)
Per a June 2026 Congressional Research Service brief, every clinical generative-AI product in production today operates in non-device lanes — lanes now under active congressional debate. Documented human review is the readiness posture for whatever rule comes next.
FDA's Predetermined Change Control Plan guidance makes ongoing performance monitoring part of the submission itself, and FY2026 appropriations law directs FDA to report on its post-deployment monitoring authorities. Continuous sampled human review stops being optional.
Anthropic's Claude for Healthcare ships HIPAA-ready infrastructure with agent skills for FHIR development and prior-authorization review. When every operator can generate clinical output at scale, compliance is solved plumbing — and the bottleneck moves to judgment: which clinical assumptions a workflow carries, and who reviewed them before scale.
The fastest-moving idea in AI engineering right now is loop engineering — you stop prompting and instead run agents in a loop: scheduled execution, isolated workspaces, persistent memory, and a separate verifier agent that checks the work. As Claude Code's creator put it, “I don’t prompt anymore. I write loops.”
“The most consequential design choice in a loop is splitting the agent that writes from the agent that checks. A model grading its own output is too generous.”
— The verifier is the part that earns trust. (The New Stack, June 2026; Anthropic harness research)
In software, that verifier can be a second model with different instructions. In medicine it cannot. When the loop generates a letter of medical necessity, a prior auth, a triage decision, or patient-facing copy, the checker that “earns trust” has to carry malpractice liability, a state license, and an NPI — because that is who a payer, a board, or a court asks. A second AI grading the first is exactly the self-generous loop the principle warns against.
And the other half of the warning — “a loop running unattended is also a loop making mistakes unattended” — is the whole reason this exists. ClinicalSwipe is the verifier node you drop into a clinical loop: a licensed, specialty-matched physician as the checker, reached by the same API or MCP connector your agents already speak, returning a signed attestation with documented findings.
The six loop-engineering primitives — the ones now shipping inside Claude Code and Codex — map straight onto a clinical loop. You own five of them. The sixth is the one you can't run yourself.
Sampled-output review by specialty-matched physicians, with a sponsor-bias panel: does your model's suggestion distribution drift toward investor or partner products? You receive a signed report with documented findings on every sampled case — the independence stamp a sponsored copilot cannot self-issue. Includes a workflow-assumptions review: the clinical premises your intake, follow-up, and triage flows carry — technically compliant copy can still be clinically wrong — examined by specialty physicians before you scale them.
One-time or annual →Continuous sampled review of a live model's outputs by licensed reviewers — quality-scored, drift-flagged, attested quarterly. Maps to FDA's PCCP monitoring expectations and gives health-system AI governance committees an instrument instead of a policy binder.
Subscription →FDA clears AI/ML-enabled devices at a pace regulatory trackers put at roughly two dozen a month — and every submission needs credentialed readers and adjudicated ground truth. Specialty-routed reader panels, single-reader to 3-of-3 consensus, with a written adjudication protocol.
Per-study →Regulatory consultancies: bring your client's study to a credentialed reader network — panels can run white-label under your engagement. Start with the same form.
Paste an intake script or follow-up message into the free Workflow Assumption Scanner — it surfaces the hidden clinical assumptions a physician would flag. Runs in your browser; nothing is sent or stored.
Rubber-stamp review approves almost everything — which is exactly what fails an FDA, payer, or board examination. Every ClinicalSwipe engagement runs on the same rails:
Every reviewer is verified against the CMS NPPES registry and reviews only in their specialty.
Documented findings, every verdictA rejection cannot enter the record without a documented clinical reason. Findings are required, not optional.
Timestamped under the reviewer's NPIEach decision is signed and timestamped to the individual clinician — examinable years later.
Consensus tiersSingle-reader for routine sampling; 2-of-3 and 3-of-3 panels with adjudication where the stakes demand it.
Independence as structure, not branding: reviewers are paid per review — never by verdict — and the platform is being organized toward reviewer ownership, so the auditors own the audit house.
Tell us what needs reviewing. A clinical lead scopes the sample, panel, and protocol — no sales sequence.