Methodology

When a reader submits a statement, an AI model assesses it and returns a preliminary verdict with an explanation and suggested sources. That draft is published in the feed labelled “Preliminary — AI only”. It is a research starting point, not a finished journalistic verdict.

The model has no live web access and a fixed knowledge cutoff. It can be wrong about recent events, precise figures, and the existence of specific documents. Where a claim depends on data it cannot confirm, it is instructed to return Unverifiable and lower its confidence rather than guess. Confidence is the model's own estimate, not an accuracy guarantee.

• True

  True — The statement is accurate and nothing significant is left out.

• Mostly true

  Mostly true — Accurate in substance, but a detail is imprecise or a caveat is missing.

• Mixed

  Mixed — Part of the statement holds up and part does not, or the framing distorts real numbers.

• Mostly false

  Mostly false — Contains an element of truth but leaves a materially wrong impression.

• False

  False — The statement is not supported by the available evidence.

• Unverifiable

  Unverifiable — The claim cannot be settled with evidence: a prediction, an opinion, or data that does not exist or is newer than the model's knowledge.

Verdicts judge the factual substance of a statement only. Party, ideology, tone, and the speaker's motive never move a rating.

Sources are ranked in this order, and higher tiers outweigh lower ones:

1. 1. Primary records — statistical agency releases (BLS, BEA, Census, CBO), bill text, roll-call votes, court filings, agency data portals.
2. 2. Official statements in full context — transcripts, recordings, and complete documents rather than clips or excerpts.
3. 3. Peer-reviewed or institutional research, with funding and method disclosed.
4. 4. Established fact-checking and news reporting, used to corroborate, not to replace primary evidence.

Advocacy material, anonymous posts, and unlabelled charts are never sufficient on their own. Model-suggested sources carry an “not yet verified” notice in the citation panel until a human opens them and confirms the figures.

• Preliminary — AI only — AI draft, published as-is, no human has checked the sources.
• Under human review — A reviewer has claimed it and is opening every source.
• Human reviewed — A reviewer confirmed the evidence, the wording, and the verdict.
• Corrected — A published check was changed after an error or new evidence; the change is logged below.

Researcher submits the claim and gathers evidence. Reviewer independently re-checks each source against the verdict. Editor approves a rating change or a correction. A single person never both writes and signs off on the same rating.

If a check is wrong, we change it and say so. Factual errors that affect the verdict are re-rated and marked Corrected, with a note at the top of the check stating what changed and when. Typos and clarifications that do not affect the verdict are edited silently. We never delete a published check to hide a mistake.

Spotted a problem? Use the feed's status control to flag a check for review, and include the source that contradicts it.