How the fact-checks work

Where the evidence comes from

Claims are checked against PubMed, the index of biomedical literature maintained by the US National Library of Medicine (~36 million peer-reviewed records), with Europe PMC as a secondary source. These are neutral indexes of published research, not publishers: nobody can pay to appear in a verdict here. Every cited paper is re-fetched independently before publication, its quoted passage is matched against the actual record, and its retraction status and citation count are checked against OpenAlex, the open scholarly database. A retracted paper is flagged in red and can never support a claim silently.

The two grading frameworks, and who defines them

The definitions of both scales are external and official. The application of those definitions to each paper and claim is done by this platform's analysis pipeline, and every grade shows its one-line justification so you can audit it.

Per paper: Oxford CEBM Levels of Evidence (1-5)

Defined by the Centre for Evidence-Based Medicine at the University of Oxford. It grades a single study by its design: Level 1 is a systematic review of randomized trials; Level 2 a randomized trial; Level 3 a non-randomized controlled study; Level 4 case-series and case-control studies; Level 5 mechanism-based reasoning or expert opinion with no new human data. Journal prestige does not set the level: a famous journal's opinion piece with zero measurements is Level 5.

Per claim: GRADE certainty

Defined by the GRADE Working Group and used by the WHO, Cochrane, and most clinical guideline bodies. It rates the certainty of the whole body of evidence behind a claim - high, moderate, low, or very low - starting high for randomized evidence and low for observational, downgrading for bias, inconsistency, indirectness, and imprecision. A claim supported only by mechanism or animal data is "very low".

The verdicts

Supported: the best available evidence says what the speaker said. Overstated: same direction, but weaker, smaller, or more conditional than claimed. Contradicted: the best available evidence disagrees. Needs context: roughly right but misleading without qualification. Unverified: no publication matching the claim was located - which does not prove the claim false.

Honesty rules the system enforces

The analysis is performed by an AI pipeline under hard, code-enforced constraints: it can only cite papers whose records it actually retrieved during that verification; every citation carries a verbatim passage from the record, re-verified independently; DOIs are never invented; an episode cannot publish while any check remains unresolved; and the speaker's exact spoken sentence is preserved next to every claim, with a link to the exact moment in the episode. Grades and verdicts are judgements produced by this platform - the frameworks are official, the application is ours, and every step leaves a visible trail you can follow to the source.