Verification methodology
Every published fact has at least two source groups on different websites. Sources from the same mapped parent company are grouped together before counting.
The problem: two sources aren’t always two sources
The industry minimum for “verified” is two independent sources (see the IFCN Code of Principles and Full Fact). Almost nobody defines what “independent” means. That gap creates four well-documented failure modes:
- Citogenesis. A false claim added to Wikipedia gets copied by multiple news outlets, each citing the others as “independent.” The Ronnie Hazlehurst case had the BBC, The Guardian, The Times, The Independent, The Stage, and Reuters all copy a fake edit within days.
- Wire-service duplication. A single AP or Reuters story is republished by 200+ outlets. Different hosts, same underlying evidence.
- Aggregator confusion. Statista and Encyclopedia.com republish government or vendor data. Citing them is really citing whoever they pulled from.
- Ownership clustering. WSJ, MarketWatch, and Barron’s are all News Corp. USA Today Network runs 260+ local papers. Same tank, different logo.
Our grouping partly addresses two of these (shared parent companies for the websites we have mapped; Wikipedia mirrors and aggregators on our host lists). It does not address wire reprints, copied wording, citogenesis through news outlets or shared studies.
Our approach
DeepInquiry treats source counts as a starting point, never as the answer. Every fact in our corpus is graded through a multi-layer verification policy that asks a stricter question than “how many outlets said this?” It asks how many separate source groups stand behind it.
1. Group before counting
Every candidate source is grouped by website, and by parent company for the websites we have mapped. Outlets under one mapped parent company count as one group. The pipeline then works with groups, not URLs. This does not detect several outlets reporting one study or story; see what this policy does not solve.
2. Set aside mirrors and aggregators
Wikipedia, its known mirrors, and content aggregators are removed from the corroborating set whenever a source of another kind is available. If only such sources exist, they are kept.
3. Score the difference between groups
Each pair of remaining groups is compared on website or mapped parent company and on kind of source (primary, peer-reviewed, trusted news, trade, aggregator, other), and weighted by kind of source: primary and peer-reviewed count slightly more; trade, aggregators and unclassified sites count less. Geography is not scored. The field is named independence_score; the name is historical. It measures differences in website or mapped parent company and kind of source, not independence of evidence.
4. Escalate exceptional claims
Borrowed from Wikipedia’s WP:V: “any exceptional claim requires multiple high-quality sources.” Every claim is classified at extraction time:
- Everyday — well-known, non-contested. Standard bar applies.
- Notable — surprising or counter-intuitive. Higher bar plus at least one primary source.
- Exceptional — challenges consensus, involves a public figure, or is time-sensitive. Highest bar, primary source required, no Wikipedia, mirror or aggregator source removed.
5. Grade the output
The API doesn’t hand you a boolean. Every fact ships with a verification tier, a 0-1 score (field name independence_score), the number of separate source groups behind it, any removed-source codes, and the claim tier. You get to decide the bar for your product.
| Tier | What it means | Available on |
|---|---|---|
| Gold | At least three separate source groups (four for exceptional claims), a high score, at least one primary or peer-reviewed source, no Wikipedia, mirror or aggregator source removed. Some claim types are capped at a lower tier. | All plans |
| Silver | At least two separate source groups at the standard bar (three for exceptional claims). | All plans (default floor) |
| Bronze | At least two source groups with a lower score — useful for research, discovery, and low-stakes use. | Growth and above (opt-in via ?min_tier=bronze) |
| Unverified | Below the Bronze bar. Exposed only for audit and research purposes. | Scale and above (audit mode via ?include_unverified=true) |
The bar is a score threshold that rises with the claim tier (everyday, notable, exceptional). Tiers measure how many different websites and mapped parent companies stand behind a fact; they do not establish that the evidence is independent. See what this policy does not solve.
What this policy does not solve
We publish limitations because a fact-verification product that claims to catch everything is one you shouldn’t trust.
What this does not do. Our grouping separates different websites, and different parent companies for the websites we have mapped; every other site counts as its own owner. It does not detect wire reprints or copied wording, and it does not reliably detect several outlets reporting one study or announcement. The score is coarse: a high value does not mean the evidence is independent. We did not test whether the claims are true. Tiers measure how many different websites and mapped parent companies stand behind a fact; they do not establish that the evidence is independent.
- The primary-only case. A Fed statement, a court ruling, or an SEC filing has one authoritative source by construction. Our tiers need at least two source groups (three for Gold), so such a fact reaches a tier only when a second group is present.
- Sustained multi-outlet fabrication. The Alan MacMasters toaster hoax was accepted by the BBC, The Scotsman, the UK government, and 12+ books over a decade with no traceable Wikipedia citation. Agreement between sources is not proof of truth. The tier is Gold, not Truth.
- Everyone cites the same primary study. Two news outlets reporting the same Nature paper rest on one study. Our grouping does not detect this, so a fact can show several source groups that trace to one original study. We have not solved this.
What we don’t publish here
We publish enough for you to evaluate the policy, cite it in your own product, and hold us accountable to it. We do not publish exact thresholds, cluster keys, the scoring formula, or our proprietary source and mirror lists — the pieces a competitor would need to clone the corpus rather than build alongside it. Enterprise customers with a signed NDA can request a deeper technical brief; email api@deepinquiry.ai.
Sources this policy draws from
- IFCN Code of Principles
- Wikidata: Help:Sources (the anti-cascade rule template)
- Wikipedia: Verifiability (WP:V) (exceptional-claim rule)
- Wikipedia: Perennial sources (mirror and aggregator blocklist seed)
- Citogenesis (the failure mode this policy targets)
- NewsGuard rating criteria (domain-quality prior seed)