Evidence
Measured and corrected beats marketed and unmeasured
How to read our claims
Four labels
- Measured
- Measured in a published study or by us, with the source named.
- Vendor-stated
- Stated by the company concerned and not independently checked.
- Target
- A goal we are building toward. Not yet achieved or measured.
- Estimate
- Our own estimate, with the basis stated.
Why omission
What the research found
Across 20 large language models and 4 widely used clinical AI tools, errors of omission accounted for more than 80% of severe errors.
Source: First, do NOHARM (NOHARM benchmark), arXiv:2512.01241, v4 revised 13 Jul 2026
Citation checks and sentence-level fact checks cannot catch an omission, because an omission has no sentence to check. That gap is the reason this product exists.
Our targets
The bars the verifier has to clear
Catch at least 85% of severe omissions on a locked, clinician-reviewed evaluation set.
Flag no more than 10% of complete answers as missing something.
Return the completeness check within one second at the 95th percentile, after the answer has streamed.
Method
How we will measure it
The evaluation set combines NOHARM's published sample cases with cases we write with clinicians. Each case has a reviewed list of the material points a complete answer must state, with their severity.
For each case we test several answers: a complete one, one missing a severe point, one missing a moderate point, and one padded with points that don't apply. Clinicians review the answers, and the set is locked and fingerprinted before the verifier is scored on it.
We will publish the first evaluation whatever it shows, including where the verifier falls short.
What we will report
| Metric | Target | Measured |
|---|---|---|
| Severe-omission recallOf the answers missing a severe point, how many the verifier flags. | ≥85%Target | — |
| PrecisionOf the points the verifier flags as missing, how many really are missing and material. | Reported, no target set | — |
| False-flag rateHow often a complete answer is wrongly flagged as missing something. | ≤10%Target | — |
| Judge agreementHow often the two independent model families agree on a point. | Reported, no target set | — |
| LatencyHow long the completeness check takes after the answer has streamed. | <1 sTarget | — |
| Measured results are published here when they exist. | ||
Status
Where we are today
The verifier is in development and has not been evaluated. There are no results to report yet. When there are, they will be on this page. If a claim on this site turns out to be wrong, we correct it on this page and say what changed.
Design-partner pilots
Help build the evaluation set
Researchers and clinicians can contribute cases, review answers and check our method before we publish.
- ≥85%Target
of severe omissions caught
- ≤10%Target
of complete answers wrongly flagged