BEM-2

Study design

The first thing a careful reader asks of a paper is what kind of study it is. We tested Examine and Examine Mini on papers first indexed in PubMed in 2026, after their training data was collected, and scored their labels against PubMed’s own.

Against seven frontier models given the same instructions, BEM-2 Examine Mini sits at the top of a close group. Reviews of animal or cell research are the one category in dispute: the instructions call them preclinical, PubMed often calls them reviews. Counted either way, every frontier model is within about two points of Examine Mini.

Against open models run on the same 2,000 papers and the same GPU, the gap is wider. Examine leads the strongest of them, Mistral-Small-3.2-24B, by 3.5 points, and Examine Mini leads it by 3.0. BEM-1 Examine leads them too, which says most of that lead comes from specialising on the task.

Speed

Examine Mini grades 200 abstracts in 7.7 seconds on an ordinary laptop CPU, 3.4 times faster than BEM-1’s fastest configuration, with no GPU and no data leaving the machine. Examine, the larger model, takes 21.7 seconds on the same laptop. On one GPU, Examine Mini takes 0.06 seconds, against 0.56 for BEM-1 Examine, and is about 51 times faster than the fastest open model we ran.

Cost

Reading 1,000 abstracts for study design costs about a hundredth of a cent with Examine Mini, and about a tenth of a cent with Examine. The frontier models in the same test cost between 8 cents and $7.87 for the same thousand abstracts.

Test size

Examine is the more accurate reader, and it shows once a test is large enough to tell. It leads Examine Mini by about a quarter of a point on 20,000 papers, on 95,334 and on a million, where that is over 2,600 more papers labelled correctly. The 300-paper test is too small to show it. A single paper there is worth a third of a point and a score can land a point and a half either side of where it settles, so Examine can come out behind, as it does by one paper.

Epistemic Warrant Optimization

A model that writes about research should not claim more than the research shows. BEM-2 Compose was trained with Epistemic Warrant Optimization (EWO), our method for teaching a writer to keep each claim within what its study design supports.

Against an otherwise identically trained version without it, blind judges found overclaiming in 27% of Compose’s texts rather than 44%, and in 37% rather than 48% when fresh judges read the same texts again. Most of the gain is in case reports and reviews; observational texts still make causal claims about half the time. Part of it comes from more cautious writing. Judges also recognised the intended study design less often with EWO, 55% against 71%.

Set against BEM-1, Compose now writes for the requested design far more reliably: judges could tell which design a text was written for 55% of the time, against 23%. Open and frontier models still do this better.

Against frontier writers given ten clinical topics to write about, Compose sits in the middle of the range. It overclaimed more often than four of the seven, significantly so only against GPT-6-Astra, and at a rate similar to Claude Sonnet 5.5.

Claim strength

Examine also rates how strongly each sentence claims cause and effect, from no relationship to direct causation, following the Sumner et al. (2014) coding. This is where the frontier models are stronger. On sentences from scientific abstracts Examine is ahead of GPT-6-Luna and of the open models that do not reason first, and behind the rest. It is a long way ahead of BEM-1 Examine, which scored 50.3 on the same 300 sentences.

Press releases are harder. On 2,074 press-release sentences Examine trails every frontier model and every open model we tested except BioMistral-7B. It was built for research text, and this is where that shows.

Warrant Check

Warrant Check compares a sentence’s claim with what its study design can support. On 1,050 texts written by seven different models, its score separates overclaiming texts from careful ones with an AUC of 0.73, against blind judges. It catches most overclaims and also flags many careful texts, so it works as a first pass that points a person at the sentences worth checking. Detecting exaggeration in press releases remains Examine’s weakest area.