How we check claims

Every page in the Evidence Library starts with the same pipeline that runs inside the EvidenceLens app. A claim (or each factual claim extracted from a video) is rewritten into a neutral, language-independent research statement. That statement is matched against our index of peer-reviewed publications using semantic search, the closest papers are read, and a language model weighs how much of the evidence supports or contradicts the statement.

The output is a verdict, a confidence score from 0 to 100, the main reasons, what is still missing, and the list of papers that drove the decision. The text on a page is the pipeline's own output. An editor may soften or correct wording, but never changes the evidence a page cites.

The research index is built from OpenAlex, an open catalogue of scholarly works. We keep peer-reviewed research articles and reviews that have an abstract, and exclude editorials, letters, errata, paratexts and other non-research work types. Older literature is included when it is still the best available evidence, but the index is weighted towards recent publications.

Pages link to papers by DOI so you can read the originals. We publish titles, authors, years and study types, never abstracts or full texts.

Every page must pass the same publishing rules before it appears: the verdict uses our fixed vocabulary, rests on enough peer-reviewed papers, is never worded more strongly than the evidence, and carries no advice.

Checked-video pages and reports are read and approved by an editor. Claim pages are published in one of two ways, and the note at the bottom of each page says which. Some are read against the cited papers and approved by an editor. Others are published automatically: on top of the publishing rules they must clear a set of evidence checks (enough papers, high confidence, no sign that the papers lean the other way) and an independent second assessment of the verdict, which has to reach the same conclusion and find the wording not overstated. Topics where a wrong answer could cause harm, and any page that fails a check, are held back for an editor instead.

Automated checks reduce errors; they do not eliminate them. If a page looks wrong, tell us and an editor will review it.

The review date is printed on every page. When the underlying pipeline improves, pages are regenerated and go through the same review again before the update is published.

A verdict describes the strength of published research for a specific statement. “Not confirmed” means the evidence does not back the statement as worded; it does not mean the person who said it lied, and it does not rule out future findings. “Not enough data” means exactly that: the question has not been studied well enough to say.

Checked-video pages judge claims, not creators. We quote at most one sentence per claim so you can find the moment in the original video, and we link to the original so you can watch it in context.

Nothing here is medical, legal or financial advice. Research findings describe populations and conditions that may not match yours. Talk to a professional about your own situation.

If a page misstates a paper, misquotes a video, or checks a claim that should not be here, write to support@evidence-lens.com with the page URL. We correct or remove pages after checking the sources again, and we note material corrections on the page.