Six independent signals. Every receipt shown.
FactGuard returns a calibrated confidence score derived from multi-signal ensemble analysis — corroboration, source quality, provenance, framing, omission, track record — grounded in peer-reviewed research. Every claim ships with the citations that produced it. You can audit the evidence. That is the design.
69.6% agreement when major fact-checkers cover the same claim (Lee et al., 2023). FactGuard surfaces cross-checker convergence as a named receipt, not a hidden factor.
A stamp is not a verdict.
Most automated fact-checking tools produce a single-pass verdict — one model, one query, one output. That is fine for filtering noise. It is not defensible for a journalist writing a story, a platform deciding what to elevate, or a researcher testing a hypothesis.
The research on fact-checker reliability is clear: what earns credibility is cross-checker agreement, multiple cited sources, and auditable transparent reasoning — not a fast classification. A verdict without its evidence is an assertion wearing a label.
69.6% agreement when major fact-checkers cover the same claim — rising to near-unanimous once rating-scale artifacts are stripped out (Lee et al., 2023, HKS Misinformation Review). The signal is strong when the method is sound.
13.1% of apparent disagreements between major fact-checkers are pure rating-scale artifacts — not genuine conflicts. Consistent, small, mapped verdict bands are not a simplification; they are precision.
6.5% claim overlap between major checkers (Snopes & PolitiFact, 2016–2022). Most novel claims arrive with no prior corroboration to reference — the engine's multi-signal analysis carries the weight.
Research citation (CC BY 4.0): Lee, S., Xiong, A., Seo, H., & Lee, D. (2023). "'Fact-checking' fact checkers: A data-driven approach." HKS Misinformation Review, 4(5). DOI: 10.37016/mr-2020-126. Interpretations are FactGuard's own.
Multi-signal. Evidence-first. Auditable.
Every FactGuard verdict is produced by evaluating six signals in ensemble. No single signal determines the result. Each signal ships with the receipt that backs it.
Corroboration — the strongest trust signal
Where multiple IFCN-accredited fact-checkers have independently evaluated the same claim and reached the same verdict, that convergence is surfaced explicitly — linked to each original source. Where no prior corroboration exists, the result says so; absence is not papered over with a confident verdict.
Source quality & diversity
The engine evaluates how many distinct, independent origins support the evidence base — distinguishing genuine corroboration from an echo of a single source amplified across outlets. Named and linked.
Provenance & media authenticity
For image- or video-based claims: provenance signals are evaluated before any verdict is issued. Where authenticity cannot be determined with confidence, the result says so explicitly.
Framing & loaded language
A claim can be technically accurate while misleadingly framed. The engine identifies emotive or loaded phrasing by quoting the specific phrase from the source with a neutral rephrase alongside it. No framing finding ships without the quoted source text.
Omission analysis
What is left out of a claim is often its most important dimension. Every omission finding ships with its citation, or it doesn't ship.
Track record & corrections history
Where the source has a documented corrections or fact-check record, that record is surfaced — bounded by sample size ("based on N fact-checks on record") rather than as an absolute judgment.
On political lean: FactGuard does not label sources or claims as left, right, or centrist. The research literature is unambiguous that perceived partisanship is what drives audiences to dismiss fact-checking. No lean labels appear anywhere in the product, including via API.
Confidence that means something.
A confidence score is only useful if it is calibrated — if 70% confidence actually reflects 70% of the evidence pointing in one direction. FactGuard builds confidence scores from signal agreement, not from a model's self-report.
Verdict bands are intentionally small and mappable. The research shows that fine-grained, non-comparable rating taxonomies are the single largest source of phantom disagreement between checkers. Nuance belongs in the supporting receipts, not in an ever-expanding label set.
Not arbitrary choices — documented findings.
FactGuard's method is directly shaped by what peer-reviewed research shows actually makes fact-checking credible and useful.
Cross-checker corroboration = the trust lever
"When multiple fact-checking organizations consistently agree…the public is more likely to trust their assessments." (Lee et al., 2023.) FactGuard surfaces independent checker convergence as a named, linked receipt.
Auditable transparency is the credibility mechanism
Snopes is cited as "widely recognized as one of the most reliable" specifically for "detailed and comprehensive investigations, which often include multiple sources and references." Transparency is the mechanism. This page and the per-verdict receipt exist because of that.
Verdict comparability eliminates phantom disagreement
98 of 228 disagreements between Snopes and PolitiFact were caused entirely by incompatible rating scales — not by conflicting judgments. FactGuard's small, consistent verdict set is a precision choice.
Non-partisan stance: research-grounded, not marketing
Disagreement between checkers clusters on polarized topics. Perceived partisanship is what drives audiences to dismiss fact-checking. The no-lean-label rule is grounded in this evidence — a specific, well-supported design choice.
Track record bounded by sample and checker bias
Fact-checkers have systematic selection emphasis. A track record derived from fact-checks inherits that lens. FactGuard bounds the signal by N and shows methodology caveats.
Most claims have no prior corroboration — by design
Only 6.5% of claims overlap between Snopes and PolitiFact across six years. The multi-signal engine — corroboration where it exists, honest "not yet" where it doesn't — is the right architecture, not a lookup table.
Research citation (CC BY 4.0): Lee, S., Xiong, A., Seo, H., & Lee, D. (2023). "'Fact-checking' fact checkers: A data-driven approach." Harvard Kennedy School (HKS) Misinformation Review, 4(5). DOI: 10.37016/mr-2020-126. Open access at HKS Misinformation Review → Licensed CC BY 4.0. Interpretations and product choices on this page are FactGuard's own.
Verification at the level the work requires.
FactGuard is designed for work where a wrong verdict has real consequences.
Journalists & newsrooms
Pre-publication claim checks with auditable receipts — cited sources, corroboration status, framing and omission analysis. Results that show the evidence so the journalist can evaluate it, not just trust it.
- Live fact-check support during breaking news
- Corroboration surfaced per-claim, with source links
- Framing and omission signals, quoted and cited
Research & policy teams
Systematic claim verification for research workflows, policy briefings, and due-diligence checking — with consistent verdict taxonomy and exportable receipts.
- Consistent verdict bands across all claims
- Calibrated uncertainty — no false confidence on thin evidence
- Per-verdict primary-source links for citation
Platforms & API integrators
Embed multi-signal verification into content pipelines, moderation workflows, or reader-facing tools via the FactGuard API — structured outputs, consistent verdict vocabulary, and auditable receipts at scale.
- Structured JSON output with all signals and receipts
- No partisan labels — safe for non-partisan platform contexts
- Cached architecture — high-volume efficiency on repeat claims
A compounding data advantage.
FactGuard is not built around a single model call. The verification quality compounds with use — in ways that are structural, not cosmetic.
Proprietary corroboration corpus
First-seen timestamps, cross-checker concordance data, and corrections records accumulate with every claim processed. This corpus is a proprietary signal that no single-model competitor can replicate from a standing start.
Calibration that improves with feedback
Confidence scores are grounded in signal agreement — and as signal weights are refined against observed outcomes, calibration improves. A tool that shows its uncertainty estimates are accurate is harder to dismiss.
Cached repeat-claim efficiency
The verification engine caches results at the claim and corpus level. Repeat claims — which dominate misinformation cycles — are served at near-zero marginal cost.
Growing source credibility panel
Track record, corrections history, and ownership signals are accumulated per domain over time — building a first-party record not available from any third-party dataset.
A note on the engine: the signals and method described here are published openly — auditability is the trust mechanism. The exact signal weights, model routing, prompts, confidence thresholds, and proprietary corroboration data are FactGuard's protected IP. See the full methodology page for the publish vs. protect framing.
API access, newsroom pricing, and team plans — in progress.
FactGuard Pro is live for individual users. API access and team pricing are under development. Reach out to discuss your use case, join the waitlist, or ask about integration requirements.
API documentation, volume pricing, or integration questions: [email protected]