Methodology

ACTUALLY checks factual claims. It does not infer intent. FALSE is a verdict about a claim, not a finding that someone lied.

One statement may contain many independently checkable claims. Each claim is extracted and scored on its own. Unrelated claims are not collapsed into one overall verdict.

A published fact check is a bundle of claim checks. Rankings score individual claim checks. The ranking period uses the statement date, not the date the check was published.

Verdicts are TRUE, MOSTLY TRUE, MISLEADING, MOSTLY FALSE, FALSE, UNVERIFIABLE, OPINION, and NO CHECKABLE CLAIM. Only rankable factual verdicts enter leaderboards. UNVERIFIABLE, OPINION, and NO CHECKABLE CLAIM are excluded from scores.

Score weights are TRUE 0, MOSTLY TRUE 0.25, MISLEADING 0.50, MOSTLY FALSE 0.75, FALSE 1.00. Repeated equivalent claims are counted once per claim family in a period. The default minimum sample is 10 qualifying claims.

Country scores measure government-member claims in the selected period, not a country's population. Party scores use the speaker's party on the statement date.

Valid previous checks may be reused. Stale or dynamic claims are re-checked. A later true claim does not rewrite an earlier false check.

AI may assist discovery, extraction, matching, and drafting. AI is never itself evidence. Administrators publish checks.

A SourceAccount is where a person speaks. A Source is an evidence-registry entry. Coverage is incomplete. Selection depends on monitored speaking accounts and public submissions. Rankings describe the current dataset and filters, not every politician everywhere.