Omission detection for clinical AI · In development
An AI answer to a clinical question can be accurate, well cited and still miss the one thing that matters: a contraindication, a dose limit, a monitoring step. We are building a verifier that checks each answer against its evidence for what it should have said, and shows you what is missing.
01 · Ask
A clinician puts a dosing question to an AI tool. The question is ordinary. The answer will be quick.
Illustration · not a live result
Question
Starting metformin for a 72-year-old with type 2 diabetes. What dose?
02 · Answer
It streams in: a starting dose, how to increase it, a ceiling. Each sentence carries a citation to a source the tool retrieved.
Illustration · not a live result
Question
Starting metformin for a 72-year-old with type 2 diabetes. What dose?
Answer
03 · Check
A citation check asks whether each sentence is backed by its source. Here, each one is.
This is where most safeguards in clinical AI stop, because nothing the answer says is false.
Illustration · not a live result
Question
Starting metformin for a 72-year-old with type 2 diabetes. What dose?
Answer
All sentences supported
04 · Read in reverse
We ask the opposite question: what in the sources should have been in the answer? In the drug label, one passage is material to this question, and the answer never stated it.
Illustration · not a live result
Drug label · Contraindications
Contraindicated when eGFR is below 30 mL/min/1.73 m².Severe
The answer above never stated this.
05 · Flag
The missing points appear under the answer, each with its severity and its source. One is severe, so the answer is not marked verified. What to do next stays with the clinician.
Illustration · not a live result
Question
Starting metformin for a 72-year-old with type 2 diabetes. What dose?
Answer
Start at 500 mg once or twice daily with meals. Increase by 500 mg weekly as tolerated, up to 2,000 mg a day in divided doses.
Not stated in this answer
06 · Measure
The verifier is in development and has not been evaluated. Before we claim anything about it, it will be tested on a locked evaluation set reviewed by clinicians. These are the targets, and we will publish the results whatever they show.
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.
>80%
of severe clinical-AI errors were omissions
≥85%
of severe omissions caught
≤10%
of complete answers wrongly flagged
<1 s
for the check, after the answer streams (p95)
The problem
Most safeguards in clinical AI test what an answer says: is each sentence supported by a cited source? An answer that leaves out a contraindication passes that test, because nothing it says is false.
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.
How it works
When a paper, drug label or guideline enters the library, we pull out its clinically material points: contraindications, dose limits, monitoring, interactions, red flags and population caveats. Each one is typed, rated for severity and linked to the passage it came from.
For each question we work out which of those points apply to it, then check whether the answer states each one. Two models from different families judge every point, and a point counts as missing only when both agree.
Missing points appear under the answer with their severity and source. If a severe point is missing, the answer is not marked verified.
Who it's for
Design-partner pilots
We are looking for a small number of hospitals, medical centers, clinicians and research groups to shape the verifier and its evaluation with us.
of severe omissions caught
of complete answers wrongly flagged
