How Facticy verifies

1.Principles

• Evidence rules. No verdict comes from an AI model's memory. Every high-weight claim is checked against sources retrieved from the live web at analysis time, and every verdict links to those sources so you can judge them yourself.
• Disagreement is shown, not hidden. Two independent AI models read every claim, and a third arbitrates whenever those two disagree — the extra opinion is spent where there is actual doubt. If the panel still doesn't converge, the claim is marked Disputed, with each model's reasoning. A manufactured consensus would be more comfortable and less useful.
• Limits are stated. A fact-checker that presents itself as infallible is exactly the kind of claim Facticy exists to flag. The limitations section below is not fine print — read it.

2.Claim extraction

When you upload a document (.pptx, .docx, .pdf or an image), Facticy identifies the verifiable claims: figures, dates, statistics, and facts attributed to named entities ("Spain's GDP grew 2.7% in 2024", "company X was founded in 2011"). Images are read with vision to recover their text and figures.

Deliberately excluded: opinions, predictions, purely internal metrics with no public referent, and rhetorical text. We verify the verifiable; we don't grade your style.

3.Source retrieval & trust hierarchy

Each high-weight claim triggers a live web search that builds a shared evidence base. Each source is classified by its independence, and that classification drives the verdict:

• Official / primary — statistics offices, regulators, peer-reviewed research, international organizations, central banks, official registries. The strongest evidence.
• Independent — established press with a corrections policy, wire agencies, recognized reference works.
• Interested (the subject's own) — a company's own reports/filings about itself: treated as the interested party's version, valid as the primary record for its own published figures, but not as neutral truth for contested claims.
• Down-weighted or excluded — content farms, authorless aggregators, forums, anonymous blogs, and pages whose only merit is ranking in search.
• A legal, regulatory, market-size or ownership claim needs at least one official or independent source; an interested source alone yields "unverifiable" (or "partial"), never a confirmation.

4.Review by a panel of independent models

The retrieved evidence and the claim go to a panel of three independent models — chosen from a privacy-vetted roster (Claude, GPT, Gemini, and more). Two of them assess every claim independently; the third is brought in whenever those two disagree, and the panel then debates. That the third opinion is bought only on contested claims is deliberate, not a shortcut: where two independent models already agree, a third vote can confirm the reading but not change it, so the extra scrutiny goes where it can actually change a verdict. Every model in the roster is vetted: none is allowed to train on your data. More than one model reduces the risk of a single model's blind spot becoming a verdict.

To be clear about what this is and isn't: agreement between models raises confidence in the reading of the evidence; the truth still comes from the sources, not from machines agreeing with each other. The models do not search the web themselves — they judge the shared evidence, so a claim is searched once, not once per model.

5.Verdicts

Facts change over time. You can mark a document as monitored: Facticy periodically re-checks its claims and alerts you when a verdict changes — a once-true figure that no longer holds.

• Verified — trusted-source evidence supports the claim and the models agree on that reading.
• False — the evidence contradicts the claim. Absence of evidence is never "false" — that's "unverifiable".
• Partial — partly supported: some elements confirmed, others not, or true only under a specific definition/period.
• Disputed — sources contradict each other, or the models disagree. Both sides are shown with their sources; you decide.
• Unverifiable — no sufficient evidence in reliable sources. Common for internal data and very niche claims. Not the same as false — but if you present it as fact, the burden of proof is yours.

6.The document score

A document gets three independent 0–100 scores — factual soundness, human-authorship estimate, and communication quality — each shown with its breakdown rather than as a single opaque number. The factual score is a weight-adjusted average of the per-claim verdicts, and the claim explorer shows the count of each verdict so you can see how the number was reached.

7.Limitations — actually read these

• Sources can be wrong. Trusted sources publish errors too. Facticy weights reliability; it doesn't guarantee it.
• Very recent facts are thin. Events from the last hours or days may lack quality-source coverage yet, and will show as unverifiable.
• Context can beat the system. Irony, accurate figures used misleadingly, technically-true-but-deceptive claims: we catch some, not all.
• Languages. The system operates in English and Spanish; source coverage is strongest in English.
• Facticy is a first pass. For legal, medical, financial or editorial decisions with real stakes, expert human review is not optional. Our job is to make that review start with a map instead of blind.

8.Your documents

You control retention: keep files until you delete them, delete the original right after analysis, or auto-delete after 30 or 90 days. Files are never used to train any AI model — the AI providers do not train on API data, and we don't either. Encrypted in transit (TLS) and at rest, hosted in the EU (Cloudflare R2, EU jurisdiction; compute in Germany). The processing sub-processors are listed publicly.

9.Found a wrong verdict?

Every verdict in your analyzed document has a feedback control — agree/disagree, suggest the correct verdict, add a comment. Each report is reviewed and error patterns feed system improvements. A fact-checker that doesn't accept corrections doesn't deserve the name.

10.Changelog

July 2026 — initial public version of this methodology.