clinicians.build · interactive · august 7, 2026

840 Dots

Stanford open-sourced an evaluator that checks whether AI-generated medical text is faithful to its input. Everything the evaluator knows about being wrong comes from these 840 examples. Here they all are.

Primary source: Aali et al., MedVAL: Toward Expert-Level Medical Text Validation with Language Models
npj Digital Medicine, published Aug 4, 2026 · preprint arXiv:2507.03152v5 · Stanford University
Every count below is from the paper’s Tables 2 and S3

The complaint about clinical AI for two years has been that nobody built the measuring stick. This week somebody did, and put it on GitHub: an evaluator trained without a single physician label, statistically non-inferior to one human expert, small enough to run on a laptop.

The interesting part is not the model. It is what the model was measured against.

The evaluator got distilled. The taxonomy didn’t.

The whole benchmark, one screen

Each dot is one physician-annotated output · six tasks · 840 total
Examples
840
Clinical tasks
6
Physicians
12
Graded unsafe
44.4%
Every dot is one AI-generated clinical output that a physician read and graded on a four-level risk scale. Level 3 and 4 mean a human has to review or rewrite it. That is 44.4% of the set — the benchmark was built to contain errors, deliberately.

How long is a benchmark question?

Click Input length above and the dots resize to the average length of what the evaluator is handed. Then look at the same numbers laid out to scale.

Average input length per task · tokens · drawn to scale
The medication-answer task — where distillation produces its largest gains — gives the evaluator a ten-token question. The radiology impression task gives it fifty. Only the ambient-scribe task, at 1,497, is the length of a short clinical document.

A real emergency department encounter is not a paragraph. It is a chart: prior notes, an active problem list, a medication reconciliation, labs with trends, imaging with priors, and the three things the patient said that nobody wrote down. Every benchmark on the market is a pre-digested excerpt of that, and this one is more honest about it than most — the authors say plainly that their task suite “does not encompass the full spectrum of clinical document types and scenarios.”

Ninety of the dots

Click The 90 that carry the claim. Fifteen examples per task were graded by more than one physician. That subset — 90 of 840 — is the entire basis for both the inter-physician agreement statistics and the headline that the evaluator is non-inferior to a human expert.

What the 840 actually encode

Twelve physicians — four internists, three bilingual residents, four radiologists and a radiology resident — graded every output against a taxonomy they defined: four risk levels, and eleven named error types under three families.

The taxonomy — the part that did not get distilled
FamilyErrors
Hallucinationsfabricated claim · misleading justification · detail misidentification · false comparison · incorrect recommendation
Omissionsmissing claim · missing comparison · missing context
Certainty misalignmentoverstating intensity · understating intensity
Most common error in the whole set: fabricated claim, 45.7%. Then missing claim, 14.0%, and incorrect recommendation, 12.6%. Error count rises with risk grade, from an average of 0.14 errors at level 1 to 3.24 at level 4.

The model is a commodity now. The judgment about what constitutes a dangerous omission in a discharge summary is not, and it doesn’t come from an engineer.

The build

The evaluator is free and the benchmark is free, which means the defensible asset is an annotated corpus for your own population. Take 100 outputs from the tool you already run. Two clinicians. These four risk levels. Grade them independently and then look at where your two annotators disagreed — that number will tell you more about your deployment than any published F1.