Home Platform
The verification layer, in detail.
Harmony Verify sits between your model and the clinician reading its output. You decide what gets routed and how fast it must come back; we handle reviewer matching, structured clinical review, evidence linking and the record that proves it happened.
Pipeline
What happens to a routed output.
Submission
Your application posts the generated output together with the clinical context it was derived from, the task type and the urgency tier. Identifiers can be tokenised on your side so that direct identifiers never reach us.
Triage & routing
Cases are classified by clinical domain and complexity, then matched to a reviewer whose licence, board certification and scope of practice fit the case. Conflict-of-interest and workload rules apply before assignment.
Structured review
The reviewer works a fixed rubric rather than writing free prose. Every dimension is scored, every correction is attached to the exact span it applies to, and every clinical judgement is linked to its source.
Adjudication
Cases that meet escalation criteria — reviewer uncertainty, high severity, or a customer rule you define — go to a second independent reviewer. Disagreements are recorded, not silently resolved.
Return
The verdict, corrections, citations and reviewer attribution come back over the API and appear in your dashboard. Your application decides what to do next: block, amend, warn or release.
Aggregate signal
Every record feeds your quality view: error rates by specialty and task type, severity distribution, drift over model versions, and reviewer agreement to keep the review layer itself measurable.
Review rubric
Judgement, made comparable.
Free-text clinical opinion does not aggregate and cannot be audited. Every Harmony Verify review is scored against the same dimensions, so results are consistent between reviewers and meaningful in bulk.
| Dimension | The reviewer is asking | Failure signal |
|---|---|---|
| Factual accuracy | Is every clinical assertion correct and supported by the provided context? | Unsupported claim, fabricated citation |
| Guideline concordance | Does this align with current society guidance and the approved label? | Guideline drift, off-label without flag |
| Safety | Could acting on this cause harm — interaction, contraindication, dose? | Interaction missed, unsafe dose |
| Completeness | Is anything material to the decision missing? | Omission of contraindication or follow-up |
| Calibration | Is the confidence expressed proportionate to the evidence? | Overconfidence, false certainty |
| Scope | Has the model answered beyond what it was asked or is competent to answer? | Scope creep, unrequested diagnosis |
Rubric weighting is configured per deployment — a coding assistant and a triage assistant do not fail in the same ways, and should not be scored as though they do.
Integration
One call in, one record out.
Harmony Verify runs asynchronously alongside your product. Submit an output, receive a verification identifier immediately, and take the completed record by webhook or by polling.
Choose what gets verified
Full review of every output is rarely the right economic answer. Most teams start with a sampled share to establish a baseline error rate, then move to targeted review — routing only what their own confidence scoring, task type or risk tier marks as worth a clinician's time.
- Sampled — a fixed share, for measurement and quality evidence
- Triggered — routed when your own signals flag risk
- Comprehensive — every output, for the highest-stakes surfaces
| verification_id | Stable identifier for the record |
| verdict | verified · verified_with_amendment · corrected · rejected |
| rubric_scores | Score per review dimension |
| corrections[] | Span, category, severity, replacement text |
| evidence[] | Source cited for each clinical judgement |
| reviewer | Credential, specialty, licence state, signature |
| timestamps | Received, assigned, completed |
Built to fit an existing stack
Verification is a service call, not a platform migration. Nothing about your model, hosting or interface has to change — you keep your product, your data flows and your customer relationship.
- REST submission with webhook callback, or polling if you prefer
- Runs asynchronously — your user-facing latency is unaffected
- Reviewer workspace for your own clinical staff, if you have them
- Record export for quality files and customer governance reviews
Adult presenting with unilateral calf swelling and pleuritic chest pain. Draft advised routine outpatient follow-up within one week.
Reviewer: presentation is consistent with possible pulmonary embolism. Routine follow-up is unsafe here — guidance amended to immediate emergency assessment.
Reviewer network
Who is actually reading your output.
The value of this layer is only as good as the clinicians in it. Reviewer quality is something we treat as a measured, ongoing obligation rather than a claim made once at onboarding.
Verified credentials
Licence and board certification are verified at source before a reviewer sees a single case, and re-checked on a schedule. Practice scope is recorded so routing can respect it.
Specialty matching
A cardiology case goes to a cardiologist. Routing is on demonstrated specialty and subspecialty, not on general availability — the common failure mode of generic expert marketplaces.
Calibration & audit
Reviewers are periodically given cases with established answers, and inter-reviewer agreement is tracked continuously. Reviewers who drift are recalibrated or removed.
Questions
The ones buyers actually ask.
Does this slow our product down?
No — verification runs asynchronously and out of band. Your user gets your response at your normal latency. The verification record arrives afterwards, and your application decides what to do with it: annotate, amend, escalate, or hold a specific class of output until it clears.
Do you need protected health information?
We need enough clinical context for a reviewer to judge the output — but that is rarely the same as needing identifiers. Most integrations tokenise or strip direct identifiers before submission. Where PHI is genuinely required, it is handled under a business associate agreement with access limited to the assigned reviewer. See Trust for the detail.
Is Harmony Verify a medical device?
Harmony Verify is a verification and quality assurance service for the companies building clinical AI. We do not generate clinical recommendations and we do not practise medicine. Whether your product is a regulated device depends on what it does and where you sell it — that determination stays with you, and our records are designed to support the evidence you need for it.
What happens when a reviewer is wrong?
We assume it will happen, and instrument for it. Escalation criteria send high-severity and low-confidence cases to a second independent reviewer, disagreement rates are tracked per reviewer, and every record is attributable — so a bad judgement can be found, corrected and learned from rather than quietly absorbed.
How is this priced?
Per verification, with rates that vary by specialty, review depth and turnaround tier, plus a platform fee at volume. Design partners are onboarded on terms agreed case by case. Tell us your volume and task mix and we will come back with concrete numbers.
Who owns the verification records?
You do. Records are yours to export, retain and present to your customers and regulators. We retain what we need to run the service and to measure reviewer quality, under the retention terms in your agreement.
Early access
See it against your own output.
The fastest way to judge this is to route a sample of your real output through it. Send us your task mix and we will scope a pilot.