Methodology

What we ingest

StanceGraph works from public statements only: published videos, podcast episodes, newsletters and posts. We do not use private messages, paywalled communities accessed without permission, or leaked material. Each item is transcribed and stored with its publication date and a durable reference to the original, so every downstream record can point back to where it came from.

How claims are extracted

A language-model pipeline reads each transcript and pulls out checkable assertions about health — not opinions about taste or lifestyle, but statements that could in principle be supported or contradicted by evidence. Each extraction is normalized into a claim record with four required fields:

• Statement — the claim rewritten as a single checkable proposition, kept as close to the speaker's meaning as possible.
• Stance — the speaker's position toward that proposition (definitions below).
• Certainty — how strongly the position was phrased, as spoken.
• Receipt — source, verbatim quote, timestamp and the extraction model version. A record that cannot carry its receipt is discarded, not published.

Model versions matter: extraction quality changes as models change, so every record is stamped with the version that produced it, and re-runs with newer models supersede rather than silently overwrite older records.

Stance definitions

Stance describes the speaker's position toward the proposition — not whether the proposition is true.

Certainty definitions

Certainty is a reading of the language used, not of the speaker's private beliefs. It matters because an emphatic recommendation and a hedged musing carry very different weight with an audience.

Evidence grading

Separately from stance, each claim's underlying proposition gets a grade for the published evidence behind it, following the standard hierarchy used in evidence-based medicine — systematic reviews and meta-analyses at the top, expert opinion and mechanism-only reasoning at the bottom:

• A — consistent findings from systematic reviews or multiple well-conducted randomized trials.
• B — at least one good randomized trial, or strong consistent observational evidence.
• C — limited or conflicting studies; small trials, short follow-ups, surrogate outcomes.
• D — anecdote, mechanism-only reasoning, or no published human evidence located.

A grade is a statement about the state of published research at the time of grading, not a verdict on the claim or the person. The grading rubric and the searches behind each grade will be published alongside report cards.

Human review

Model output is not published unreviewed. Before a report card goes public, a human reviewer checks each claim record against its receipt: does the quote say what the record says it says, is the stance label right, is the certainty reading fair, does the conflict matrix reflect disclosures actually present in the source? Records that fail review are corrected or removed, and the review status is stored with the record.

Corrections

A formal corrections policy — how to dispute a record, response timelines, and how corrections are displayed — will be published here before the first public report cards. The commitments it will be built on: every published record can be disputed by the person it concerns, disputes are reviewed by a human, and corrections are shown in place rather than silently applied.