What you can check, and what is still open.

The product

• ✓ In placeA live service: read a draft referral, see the published evidence or an ask. Free for GPs: no account, and a work email once to read your own letter. Try it
• ✓ In placeA vendor API with a published OpenAPI 3.1 specification. The spec
• ✓ In placeSandbox keys that cost nothing and answer from recorded readings. Developers
• ✓ In placeA refusal contract: the 5 clinical requests are refused, text with any of 9 kinds of identifier is refused before it is read, and unknown requests are refused. Pinned in CI and checked against production after every deploy. The contract

The evidence

• ✓ In placeFigures from each publisher’s own file, never merged across publishers: Republic of Ireland by hospital, NTPF, monthly; Northern Ireland by HSC Trust, Department of Health, quarterly; Scotland by NHS Board, Public Health Scotland, monthly; Wales by health board, Welsh Government, monthly; England by provider, NHS England, monthly.
• ✓ In placeEvery figure carries its publication date, on screen and in every API response. API responses also carry a hash of the file the figures came from. See every figure
• ✓ In placeAn archive of 13 monthly NTPF releases, and a replication across all of them. How we checked
• ✓ In placePublishers checked for new releases three times a week; a newer release is reviewed by a person before it is served.
• ✓ In placeWaits are shown as the publisher's own counts by time already waited. No average wait is ever derived, because no publisher gives one.

The readers

• ✓ In placeTwo independent readers from different model families; when they disagree, Vouch asks rather than choosing. The evaluation
• ✓ In placeA held-out evaluation of candidate reader pairs, with the cost of asking and both scoring rules shown.
• ○ OpenNext evaluation run: the reader pair serving live traffic today. The published results cover three readers and every pairing of them.
• ○ OpenClinician labels: one reviewer, a final-year medical student, has labelled 12 held-out letters; two GPs are asked next. Today each invented letter is labelled with the specialty it was generated to.

What it will not do

• ✓ In placeChoose a hospital, set urgency, rank patients, book an appointment or send a referral. The API refuses each of these outright.
• ✓ In placeDiagnose, or predict how long a particular patient will wait. The figures describe people already on each list.
• ✓ In placeDecide whether to refer. It assumes that decision is made, and is not for urgent or emergency referrals.

Data

• ✓ In placeThe letter text is not stored. There is no field for it. Privacy notice
• ✓ In placeHosting in Google Cloud, europe-west1 (Belgium). Where each reader processes the text is stated in the privacy notice: today one reader is in the United States and one on a global endpoint.
• ○ OpenAn EU-only reading path: built and tested, and deployed once EU model capacity is confirmed. Until then, Vouch does not describe itself as EU-only.
• ○ OpenA data processing agreement, signed with each pilot before any real referral is read. Until one is signed, invented letters only.

Regulatory and clinical

• ○ OpenMedical-device qualification: the intended purpose is drafted and ready for an independent regulatory opinion under EU MDR and UK MDR.
• ○ OpenClinical validation: a pilot is where the effect on waiting times and patient outcomes gets measured. Until then, Vouch claims none.

A pilot, measured

Written on 28 September 2026, before any pilot has started, so the numbers reported afterwards are the ones named here and not the ones that looked best.

Amended on 30 September 2026, still before any pilot: each reading counts one outcome, its first choice. Before this, a second choice counted again on some paths, and a specialty typed into the box after an ask was not counted at all.

• ✓ In placeCounted by Vouch, as counts and never as letters: readings; readings where the two readers agreed and where they asked; when they asked, which of the two named specialties the clinician opened first, or neither; when they agreed, how often the clinician opened a different specialty instead; records saved. One outcome per reading: a later choice, or going back to a letter already read, is not counted again, and choices on the invented examples are counted apart from real readings.
• ✓ In placeCounted from the service's own log lines, which carry the decision and the specialty and have no field for the letter or the person. Exported weekly during a pilot and reported as counts, with how many clinicians took part and how many readings each made, so one enthusiastic user reads as one.
• ✓ In placeReported by the practice, not by Vouch: where each referral went, and whether anything about the wait changed. Vouch is beside the workflow and not in the send path, so it cannot see the destination and does not try.
• ✓ In placeReviewed by clinicians, not counted by Vouch: a blinded sample of letters, including ones the readers agreed on, since agreement is not correctness; which asks were useful, which were unnecessary and which were missed; the time each check added; and what the clinician did differently.
• ✓ In placeNot claimed from a pilot of this length: any effect on waiting times or patient outcomes. Whether patients who can travel benefit more than those who cannot is a question for the practice's own data, and the report will say so rather than answer it.

Ask about a pilot

One referral workflow, four to six weeks, with a written report at the end. Scope, who sees what, who reviews the asks and the exit criteria are agreed in writing before it starts. Tell us who you are and where you would like to start; a person replies.