Fact-check evidence sources and tiers
The twelve specified evidence sources, the T1–T6 credibility model, and the two-independent-sources rule every verdict must meet. Part of the MediaMeter specification.
This describes the fact-check evidence model in our specification — the sources and rules a verdict must satisfy. Design, not deployed capability.
A verdict is only as good as its evidence chain. Our specification fixes both where evidence comes from and the minimum bar a TRUE or FALSE finding must clear.
The credibility model
Evidence sources are ranked in tiers by their institutional reliability and verifiability:
| Tier | Class | Examples from our registry |
|---|---|---|
| T1 | Highest-credibility primary and reference sources | IPCC reports; government statistical services (World Bank, WHO, Eurostat, NOAA); IFCN-signatory fact-checks; Google Fact Check Tools |
| T2–T3 | Strong academic and reference sources | PubMed, Semantic Scholar, Wikidata/DBpedia |
| T4–T5 | News and aggregators | GDELT, news APIs — usable for discovery and context, not for verdict-level proof |
| T6 | Low-credibility material | Unverified claims, circular reporting — never verdict-supporting |
The twelve specified sources
| Source | Type | Class | Status |
|---|---|---|---|
| Google Fact Check Tools API | API | aggregator | enabled |
| IFCN Signatory Fact-Checks | API | factcheck | enabled |
| Wikidata / DBpedia (SPARQL) | SPARQL | reference | enabled |
| World Bank Open Data | API | government | enabled |
| WHO Global Health Observatory | API | government | enabled |
| Eurostat | API | government | enabled |
| NOAA | API | government | enabled |
| IPCC Reports | scrape | academic | enabled |
| PubMed | API | academic | enabled |
| Semantic Scholar | API | academic | enabled |
| GDELT Project | API | news | enabled |
| NewsAPI | API | news | disabled — pending licence review |
The rules a verdict must satisfy
- Two-source rule — a TRUE or FALSE verdict requires at least two independent T1–T3 sources. Two outlets repeating the same wire count as one; the error code
FACTCHECK_SOURCE_COUNT_BELOW_THRESHOLDfires when the rule can’t be met. - Evidence-chain publication — every verdict carries the claims, the sources considered, and the reasoning — so a reader can check the same evidence the system did.
- Human review for high-stakes findings — FALSE and MISLEADING verdicts require a human reviewer before publication.
- Honest uncertainty — where sources disagree or fall short, the verdict is UNVERIFIABLE, not a coin-flip guess.
Retrieval is specified as hybrid (lexical BM25 + dense semantic similarity, multilingual-e5 embeddings) with deduplication thresholds — because the first verification problem is often finding the prior fact-check or primary document, not judging it.
Last reviewed: October 20, 2026. Part of the MediaMeter knowledge base — editorial reference material, not product documentation.