August 12, 2026
BEM-1
Our first family of biomedical models: two that read biomedical writing, and two that write it.
Where BEM began.
BEM-1, the Biomedical Epistemic Model, was our first family of biomedical models: two that read biomedical writing and two that write it. BEM-1 Examine classifies a text’s evidence tier with a calibrated confidence and scores its hedging and claim strength; Examine Mini screens every input first and passes the hardest 1% up, matching Examine’s accuracy end to end at 2.86 times the speed on CPU. Compose writes under a requested study design, and Critique raises the methodological objections a careful reviewer would ask first. On held-out abstracts, Examine reaches 94.76% evidence-tier accuracy. Two findings shaped what came next: hedged writing turned out not to signal weak evidence, and Compose’s study-design conditioning was a null. BEM-2 rebuilt the writer, and blind judges now recognise the requested design in 55% of its texts, against 23% for BEM-1 Compose.
- evidence-tier accuracy for Examine, on held-out abstracts
- 94.76%
- macro-F1 for Examine Mini on 31,527 abstracts it never saw in any form
- 92.00%
- of inputs Examine Mini settles alone in the cascade, at 95.09% accuracy
- 99.0%
- faster on CPU than Examine alone, with no loss of accuracy
- 2.86×
The family.
BEM-1 Examine
ReadsClassifies a text’s evidence tier with a calibrated confidence, and scores its hedging and claim strength.
BEM-1 Examine Mini
ReadsA compact version of Examine that screens every input first and passes the hardest 1% up.
BEM-1 Compose
WritesA small biomedical chat model that writes under a requested study design.
BEM-1 Critique
WritesRaises the methodological objections a careful reviewer would ask of a claim or abstract first.
Results.
Measured by Essarion, including where BEM-1 fell short.
Evidence tier, by study design
F1 for Examine Mini on held-out abstracts, in their natural mix of study designs.
The cascade
Seconds to tier 200 abstracts on CPU. Accuracy end to end is 94.69%, against 94.66% for Examine alone. Lower is better.
Mini keeps up with Examine
Evidence-tier accuracy on the same held-out abstracts, in their natural mix and balanced across designs.
Overclaim detection, on real text
ROC-AUC for Examine. The first set is 68% rule-based synthetic text, so it flatters; on real writing the score drops.
Writing for a requested design: a null
How often Examine recovers the study design Compose was asked to write in, over 600 generations, against the chance floor.
One finding worth knowing: hedged writing does not mean weak evidence.
We tested whether more hedging comes from lower tiers of evidence, three ways, on about 430,000 held-out abstracts. None of the correlations was significant. Case reports are among the least hedged, because they narrate one patient and make no general claim; meta-analyses are among the most, because they discuss heterogeneity and bias.
The models are private.
BEM-1 is succeeded by BEM-2, available in Essarion Science.
Ask harder questions.
If you are working on similar problems, we would like to hear from you.