Research prototype, not for clinical use. Not validated by clinicians. Every record here is synthetic.

Every contraindication on the drug's FDA label is checked against the record. Pick an example.

Input

Recorded output

You define typed facts (boolean, number with a threshold, date window) and your own outcome labels as rules (AND, OR, NOT, ATLEAST k OF (...)). The model reads the facts; code applies the rules.

Input

Recorded output

True, false or not stated?

Multiple choice (2 to 60 options)

RxDecide architecture: patient record, normalizer, fact reader model, decision logic, router, output
.861balanced accuracy, 120 real clinical notes
0contraindications missed among automatic answers
.932new decision types, 5 input formats
.978synthetic notes
13×faster than Qwen3-8B asked directly
$0.04per 1,000 decisions on one L4

Decisions: applies does not apply not enough information

Human review is flagged when confidence is below 0.90, when "does not apply" has confidence below 0.99 (clearing a contraindication needs more certainty), when an undefined judgment term such as "severe" is involved, or when a code rule and the model disagree.

Confidence = (K·pmax − 1)/(K − 1) over the K possible answers.

All test numbers are from pre-registered evaluations, each run once. Labels for real notes came from two independent language-model labelers; no clinician reviewed them. Details, limitations and safety findings are on the model card.

This page shows outputs recorded from the real model. To try your own records, run it on any GPU (a free Colab T4 works) or on a CPU (slower):


      
The repository contains the weights and the meddecide package. A FastAPI server is included
        (server.py). See the model card for the full API and every input format.