How a paper becomes evidence on this site

How papers are selected

Selection is not editorial. For each article topic, the pipeline runs a literature search and then narrows it mechanically, in three stages:

• Search. A PubMed query for the topic returns every candidate record — for the Candida article, 4,069 of them.
• Similarity. Each candidate's title and abstract is embedded with a local embedding model and ranked by similarity to the article's question. This is a coarse filter; it is fast and it is not clever.
• Reranking. The top candidates are scored a second time by a dedicated reranking model, one paper at a time, against the article question. The best eight go forward.

Where a paper's full text is openly available, it is used: first from PubMed Central, then from Unpaywall. Where it is not, only the abstract is used, and that limitation is recorded with the paper rather than hidden.

How claims are pulled out of a paper

A language model reads the paper in sections and proposes discrete claims. Each proposed claim must carry a verbatim quotation from that paper, and the quotation is then checked against the paper's own text by the software, not by the model.

• The quotation must actually be in the paper. One contiguous, unbroken passage — no ellipses, no stitching two sentences together. If the quotation cannot be found in the source, the claim is discarded and never becomes evidence.
• Typography is forgiven; wording is not. The check folds the typographic differences that survive a journal's typesetting — thin spaces, en dashes, curly quotes — and collapses runs of whitespace. It does nothing else. A paraphrase still fails.
• Discarded claims are counted, not hidden. Every run records how many claims were dropped for an unverifiable quotation, and that count is published with the run's measurements.
• Unreported is not inferred. Dose, route, comparator, population and outcome are recorded as unknown when the paper does not report them. The model is instructed never to fill them in.

Each surviving claim keeps the paper's PubMed identifier and the character offset where its quotation begins. That is what the small numbers beside a sentence carry: not just which paper, but where in it.

What this site does not claim

Papers are triaged internally by a model score that is never displayed on this site; every article claim carries its primary-source citation instead.

• Nothing here is medical advice. A laboratory result is a laboratory result, and it says nothing about whether anyone should take anything.
• A laboratory or animal finding is never written up as if it applied to people. Where a claim's study system is a cell line, a microbe or a computation, the software forces that distinction, whatever the model proposed.
• No language model here is a peer reviewer. Models read papers, propose claims and quote them; the software checks the quotations. This site does not describe that work as medical review.
• Everything on this page is provisional. A claim can change when a paper's full text becomes available or an error is reported — which is what the corrections policy is for.

How to check any sentence on this site

Each small number beside a sentence is a link to the paper it came from, and it carries that paper's PubMed identifier and the position of the supporting quotation in the source text. If a sentence and its paper do not match, that is a correction we want: the corrections policy and the contact address are on the About page.