How FactGuard Works
Every verdict is produced by the same evidence-first method: multiple independent verification signals examined in parallel, cross-checking one another, combined by a calibrated ensemble into an honest graded verdict — published with the receipts that produced it. No black boxes. No stamps without evidence.
How a verdict is produced
FactGuard's verification engine runs automated analysis through a consistent, evidence-first sequence. The same method applies to every claim, regardless of its origin or subject.
The claim is parsed and scoped
You submit a claim, headline, or URL. The engine identifies the specific assertion being made — the exact claim statement — and establishes what evidence would confirm, contradict, or complicate it. Ambiguous scope is flagged rather than assumed away.
Evidence is gathered and weighed across multiple signals
Relevant primary sources are retrieved and assessed — original records, reputable reporting, and authoritative references, not paraphrases or unsourced summaries. Simultaneously, the engine runs multiple independent verification signals in parallel. Each signal examines the evidence from a distinct angle, cross-checks the others, and carries its own receipt. No single signal decides the outcome alone.
A calibrated confidence score is assigned
The strength, quality, and agreement of the evidence — across all signals — is converted into a calibrated confidence score from 0 to 100 and a graded verdict label. Reality is rarely a clean true-or-false. The score reflects that.
The receipts are published alongside the verdict
You see the verdict, the confidence score, the cited sources, the reasoning that connects them, and the source-credibility signals that informed the analysis. A verdict without its evidence is just an assertion. Every FactGuard result shows the work — not as a disclosure, but as the mechanism of trust.
Multiple independent signals. Each backed by a receipt.
No single signal determines a verdict. FactGuard evaluates every claim across a proprietary ensemble of independent verification signals — run in parallel, cross-checking one another. Each signal surfaces its own evidence, each receipt is auditable, and the confidence score reflects agreement across the full ensemble, not the output of any single pass.
The shape of this method is described here and illustrated below. The specific signal categories, their weights, model routing, prompts, and proprietary data that power the ensemble are FactGuard's protected IP — see What we publish vs. what we protect below.
Independent, parallel analysis
Each verification signal runs independently on the same claim simultaneously. No signal waits on another; no single signal's output can dominate the result. This design is deliberately redundant — independent agreement strengthens confidence; disagreement is surfaced, not smoothed over.
Evidence-grounded receipts
Every finding a signal produces must be backed by cited evidence — a source link, a quoted text, a referenced record. A finding without its evidence is just an assertion; FactGuard ships the receipt or it does not ship the finding.
Calibrated ensemble verdict
Signal outputs are combined by a calibrated ensemble that reflects the strength, quality, and agreement of the evidence — not a simple average. When signals converge, confidence is higher. When they diverge, the result says so honestly, rather than fabricating certainty from a contested evidence base.
Honest uncertainty disclosure
Where evidence is thin, rapidly developing, or genuinely contested, the system returns a lower confidence score or a "Not Yet Verified" state — rather than forcing a confident verdict. Absence of evidence is disclosed, not hidden.
What the signal ensemble is not: it is not a political-lean detector. FactGuard does not label sources or claims as left, right, centrist, or biased. Each signal surfaces factual evidence, and the reader synthesizes. No aggregate "credibility score" is printed for a source; signals remain disaggregated so the reader, not the algorithm, draws the conclusion.
How many independent signals converge into one honest verdict
A claim or piece of media is examined by several independent signals at once. No single signal decides the outcome — they are weighed and combined by a calibrated ensemble that produces a graded verdict with an honest confidence score. The diagram below shows the overall shape of that method, not its internals.
FactGuard's proprietary 7-step verification algorithm — seven interconnected steps that cross-check one another before a verdict is issued.
Each step is a proprietary signal — shown here by code name only (ARGUS, ATLAS, AEGIS, PRISM, UMBRA, ECHO, LEDGER). The underlying methods and weights are withheld.
We publish the method; the recipe stays protected. This diagram shows the shape of verification — seven interconnected steps, calibrated together — not the individual steps, their weights, prompts, or model routing, which are FactGuard's proprietary IP. See What we publish vs. what we protect.
What we publish vs. what we protect
A credibility brand lives or dies on the trustworthiness of its method, not the opacity of its engine. We open the method and the evidence. We protect the implementation details that make the engine hard to replicate.
We publish openly
- The full method — how the ensemble works at a high level, what the verification approach measures, and the evidence standard required to ship a finding
- Per-verdict receipts — every result includes the cited sources, the confidence score, and the reasoning that connects them; where independent prior verdicts exist, they are linked directly
- Honest calibration — confidence scores, "Not Yet Verified" states, and explicit uncertainty disclosures when evidence is thin or contested
- Research grounding — the peer-reviewed findings that shaped the design, with full citation
- Known limitations — where automated analysis is likely to be weaker, what the system cannot determine
We protect as proprietary IP
- Signal identities & weights — which signals the ensemble uses and how each is weighted
- Prompts & model routing — the specific instructions and routing logic
- Confidence thresholds — the decision boundaries that map signal agreement to verdict bands
- Proprietary verification data — claim history records, cross-checker concordance data, and corrections corpus
Why this boundary? Open method + auditable receipts is the mechanism of trust. Proprietary implementation details protect the compounding investment in data, calibration, and accuracy. The distinction between "how we think about verification" (open) and "how we built the engine" (protected) is deliberate.
Principles
The same standards govern every verdict, regardless of the claim or its origin.
Evidence over opinion
Claims are judged against verifiable evidence — not authority, consensus, or preference.
Cited primary sources
Every verdict links to real primary sources you can open and check for yourself.
Calibrated confidence
A graded 0–100 score and verdict label — not a blunt binary — reflecting evidence quality.
Non-partisan
The same method applies to every claim regardless of its political origin. No editorial line.
Transparent reasoning
Every conclusion ships with the reasoning that produced it. No black boxes.
Privacy
An on-device option keeps queries on your device. Zero ads, zero tracking.
Why FactGuard is built the way it is
Our verification approach isn't arbitrary. Each design choice traces back to what peer-reviewed research shows actually makes fact-checking trustworthy and useful.
Independent convergence is the strongest trust signal — so we show it
The study found that when multiple independent fact-checkers examined the same claim and reached the same verdict, the public was significantly more likely to trust the result. Where prior independent verdicts exist, FactGuard surfaces them as explicit receipts. Where they do not, that absence is shown honestly, not papered over.
Auditable transparency, not a black box
Detailed and comprehensive investigations that include multiple sources and references are the specific quality that earns credibility. A verdict without visible reasoning is just an assertion. That is why every FactGuard result includes cited primary sources, confidence assessment, and the reasoning that connects them.
Small, mappable verdict bands — nuance lives in the receipts
13.1% of all apparent disagreements between major fact-checkers were a pure rating-scale artifact. FactGuard uses a small, consistent verdict set so verdicts remain comparable. The nuance belongs in the supporting receipts, not in an ever-expanding label taxonomy.
No political-lean labels — by design, not omission
The perception of partisan bias is what drives audiences to dismiss fact-checking. FactGuard applies the same verification method to every claim regardless of political origin. We do not label sources or claims as politically left or right.
An honest "Not Yet Verified" state — absence of evidence is not evidence
Only 6.5% of claims overlap across major fact-checkers. For claims where the evidence is sparse, rapidly developing, or genuinely contested, FactGuard returns a "Not Yet Verified" result rather than forcing a confident verdict from thin evidence.
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 under Creative Commons Attribution 4.0 (CC BY 4.0). Interpretations and product choices on this page are FactGuard's own and not endorsed by the authors.
Limitations
Automated analysis is powerful, but it is not infallible. We would rather be honest about the edges than overstate what a verdict means.
Verdicts can be incomplete or incorrect. They are produced by automated analysis, which can make mistakes, reflect data biases, or appear confident while being wrong.
New, breaking, or ambiguous claims are harder. When evidence is sparse, rapidly changing, or genuinely contested, confidence is lower and the verdict is less certain.
Always verify what matters. Treat a FactGuard verdict as a starting point — check important claims against the cited primary sources yourself.
For general information only. FactGuard is not a final authority and is not legal, medical, financial, or other professional advice.