Fact-checking fails when the evidence disappears. MediaMeter keeps the claim, the retrieval path, the sources considered, and the reviewer decision in one chain.
Five steps
Extract
Atomic claims are separated from rhetoric, opinion, and prediction. Check-worthiness ranks what matters.
Retrieve
Candidate evidence is gathered from the web, wire services, official records, and the publisher’s own archive.
Rank
Sources are scored for proximity, independence, recency, and reliability. Weak sources are kept in the chain, not hidden.
Verify
The claim is compared with the best available evidence and assigned one of seven IFCN-aligned verdicts.
Review
FALSE and MISLEADING verdicts require a named human reviewer. The reviewer can accept, edit, or override.
The verdict scale
Conservative by design
A wrong FALSE verdict does more damage than an honest “we could not establish this.” The system is tuned to prefer UNVERIFIABLE when evidence is thin, conflicting, or inaccessible, and it records why.
The evidence chain
Each verdict links to a structured record: the exact claim span, the sources retrieved, the passages relied on, the sources rejected, the model version, the reviewer, and the timestamp. Export formats include JSON and PDF so the chain can travel with a published correction or an audit request.
Claim: “The plan eliminates 40 percent of bus routes.” Evidence considered: plan document (primary), agency FAQ (primary), council transcript (primary), three news reports (secondary). Finding: the plan lists service changes on 12 routes but no eliminations. Verdict: MISLEADING. Reviewer: required.
Technical specifications
Evidence sources
12+ source classes are specified at launch, including official records, wire copy, court filings, legislative transcripts, and vetted fact-check archives. Source coverage varies by country and language.
Languages and precision targets
50+ languages is a target at general availability, not a current capability. The false-precision target is ≥ 0.90 on the evaluation set, reported with confidence intervals.
Human review policy
FALSE and MISLEADING verdicts always require review. TRUE and MOSTLY TRUE verdicts can be auto-cleared only when confidence exceeds the configured threshold and no conflicting primary source was found.