Skip to content

Frontier models know everything except how the work gets done.

  1. Models can name a diagnosis. They struggle with the hard case.
  2. Clinicians weigh evidence and rule things out. Reports keep only the conclusion.

We record the answer and how clinicians reached it.

Four products. Reviewed by people who do the job.

Start with an eval to find the gap. Then buy the data or environments that close it.

SFT data. Written by clinicians, checked by a second one.

Reference answers, rationales, rubrics and labels written by clinicians in the field. A second clinician, who did not write the item, reads it before it ships.

Evals. Benchmarks built from clinical work.

Private and public evals, scored on separate axes and never blended into one number. You see what your model gets wrong and how badly.

RxScribe-BenchCorrectness, %

  1. Spark 1.358.0
  2. 5.6 Sol57.8
  3. Gemini 3.1 Pro55.1
  4. Spark 1.249.5
  5. Opus 547.1

RL environments. Graders written by radiologists and pathologists.

Tasks, tools and graders taken from clinician workflows. Clinicians who do that work check the reward signals.

Grade a prostate needle biopsy

0/11

    Licensed datasets. Existing corpora, with samples first.

    De-identified prescriptions, reports and imaging. Ask for a sample before you commit, or have us collect a new set to your spec.

    Prescriptions12M+
    General pathology reports3.3M+
    Discharge summaries3M+
    Radiology images and reports1M+
    Pathology WSI images and reports500k+

    Can a model read a prescription?

    Medicine-only correctness

    higher is better

    1. Spark 1.358.0%
    2. 5.6 Sol57.8%
    3. Gemini 3.1 Pro55.1%
    4. Spark 1.249.5%
    5. Opus 547.1%

    Fabrication rate

    lower is better

    1. Spark 1.38.3%
    2. Spark 1.28.8%
    3. 5.6 Sol8.9%
    4. Gemini 3.1 Pro14.8%
    5. Opus 516.4%

    Hard-field robustness

    higher is better

    1. Gemini 3.1 Pro51.0%
    2. Spark 1.350.4%
    3. 5.6 Sol49.2%
    4. Spark 1.248.1%
    5. Opus 546.4%

    Hard-field attempt rate

    higher is better

    1. Gemini 3.1 Pro80.6%
    2. Opus 580.1%
    3. 5.6 Sol77.8%
    4. Spark 1.376.4%
    5. Spark 1.275.8%
    Get one dose wrong and the whole medicine counts as wrong.

    Remote work for practicing clinicians.

    Read records, write rubrics and rule on the cases models get wrong. The work is paid and remote. More than 2 million people are in our expert network.

    See open roles
    7,900+
    healthcare professionals

    Nothing personal survives the pipeline.

    Three steps from patient document to a record we can keep.

    How we handle data
    1. 01

      Patient document arrives

      In an isolated environment, names, DOB and IDs still on it

    2. 02

      Personal details removed

      All 18 identifier types stripped

      HIPAA Safe Harbor identifiers

    3. 03

      Only a clean copy is stored

      We never store unredacted patient data

      Encrypted delivery

    Tell us what your model gets wrong.