Trust what your documents say
Facticy fact-checks your presentations and reports with a panel drawn from six independent models — none of which trains on your data — cross-checking each other over shared evidence — and marks exactly what to fix, without ever touching your format.
6
independent models on the roster — 3 judge each claim
6
deterministic checks, no LLM
3
trust scores per document
0
training on your data, ever
Three scores, one clear picture
Every document comes back with three independent read-outs — so you know exactly what to trust and what to fix.
Factual soundness
How well the claims hold up against real evidence.
Human authorship
An estimate of how likely the text was AI-written.
Communication quality
Structure, action titles and clarity — a senior partner's read.
And for whoever is on the receiving end
You get fifty decks a quarter and take every one of them on trust. Upload the PDF and we return what stands up and what does not, with the source behind each verdict.
For investment committees, boards and audit teams.
How it works for receiversReal example — a claim that went to all three models
"The Spanish steel market grew 25% in 2024, reaching €48 billion."
Slide 12 · extracted automatically from your deck
Official data shows the market contracted 3% in 2024. The 25% growth figure is contradicted by primary sources.
reuters.com
INE statistics report a decline, not growth, for the sector in 2024. The claim's magnitude is refuted.
ine.es
Both independent and official sources contradict the claimed growth. Verdict: false with high confidence.
reuters.com · ine.es
Final verdict: FALSE (3/3 consensus) — marked on its slide with the correction proposed as a comment.
Same engine, another market — sources are wherever the fact lives
"Dubai residential transactions rose 36% in 2024, to AED 522 billion."
Page 4 · a market report received from an adviser
Verdict: TRUE (2/2, no arbitration needed) — the figure matches the land registry that publishes it.
Checked against dubailand.gov.ae. When a claim's authoritative source is a public registry, that is what the panel is shown — not a press summary of it.
Everything you need to ship with confidence
Claim-by-claim verification
Every figure, date and fact is extracted with its slide's context and checked against official statistics, annual reports and reputable press — one shared evidence base, three independent judges.
Multi-AI debate
Two independent models vote on every claim and a third arbitrates when they disagree; then the panel debates. You see each model's reasoning, its confidence, and how verdicts changed after the debate.
Your format, untouched
The reviewed copy marks each issue on its slide with the proposed fix as a comment — native PowerPoint/Word comments included. Your bullets, fonts and layout stay exactly as they were.
Quality like a senior partner
Action titles, MECE groupings, pyramid structure: actionable feedback slide by slide, with the rewritten title or regrouping proposed — not generic advice.
How it works
From a raw file to a verified, annotated document — the full pipeline, in five steps.
1 · Upload
Drop a PPTX, DOCX, PDF or image (or paste text). Facticy parses every slide, table and native chart — and reads text and figures out of images with vision.
2 · Extract claims
An AI pulls out every verifiable claim — figures, dates, names, statistics — each tagged with its slide and a weight. Opinions and filler are ignored.
3 · Deterministic checks
In parallel, no-LLM engines recompute your maths (CAGR, totals, percentages), check native-chart integrity and terminology, and reconcile every figure against your source Excel model.
4 · Research + AI panel debate
One web-research pass builds a shared evidence base. Two independent models judge every claim on that evidence; where they disagree, a third joins to arbitrate, and the panel debates over up to three rounds until it converges. A model that never answered, or answered without saying why, is shown but does not count towards the verdict — the claim states how many votes actually stood behind it.
Two independent models vote separately and a third arbitrates when they disagree — a claim is only marked FALSE with contradicting evidence, and disagreements go to a directed debate.
5 · Scores + annotated document
You get three scores, a verdict per claim with its sources, and your document back with each issue marked in place — plus optional corrected and plain-language versions.
See it in action
Every flagged claim comes back marked on its own slide, with the correction and sources as a native comment — your formatting untouched.
Market overview
Slide 12Contradicted by INE/Reuters: the market contracted 3% in 2024. Suggested: “declined 3% to €38bn”.
ine.es · reuters.com
Six verdicts, no false certainty
Every claim gets one of six verdicts — and "no outside source can settle this" is a first-class answer that comes with the reason why.
Verified
Trusted-source evidence supports the claim and the models agree.
False
The evidence contradicts the claim. Absence of evidence is never "false".
Partial
Partly supported — true only under a specific period or definition.
Disputed
Sources or models contradict each other. Both sides are shown with their sources.
Unverifiable
No outside source can settle it — and the card says which kind: your own methodology, licensed data, a bucketed range, or simply not found. Not the same as false.
Misextracted
The claim we built does not faithfully represent the passage it came from. Our defect, not your document's: it scores nothing and is never marked on your file.
How we verify
The full method — sources, models, verdicts and every limitation, in plain language.
"Unverifiable" is three different things. We tell you which.
A report that comes back "41% unverifiable" reads as a tool that gave up. The same analysis, split by why, is a statement about your document — and most of it is in your favour.
Unverifiable by construction
A figure computed with your own published methodology, sourced to licensed data, or stated as a bucketed range cannot be confirmed by any public number — ever. Properly sourced and unfalsifiable is a finding in your favour, and we say so on the card instead of implying a gap.
Method undisclosed
When a figure does not say how it was computed, a reader hits the same wall we did. That is the group worth acting on before you publish, and it is counted separately from everything else.
The honest residue
What is left is what we genuinely could not find. Stated as exactly that — never dressed up as a verdict, never counted against you as if it were false.
Same analysis either way. One version reads as a shrug; the other tells you which claims a reader could challenge and which they never could.
Your documents stay yours
Built for confidential material from day one.
EU-hosted & encrypted
Stored in the EU and encrypted at rest.
Never trains AI
Your content is never used to train any model.
Confidential mode
Verify fully offline — no web search, no images leaving.
You control retention
Keep it, delete after analysis, or auto-delete in 30/90 days.
Questions worth asking
The things people check before uploading a confidential document.
How does Facticy decide if a claim is true? +
Each high-weight claim triggers a live web search that builds a shared evidence base, ranked by source independence (official, independent, or the subject's own). Two independent models then judge that evidence, a third arbitrates where they disagree, and the panel debates. The verdict comes from the sources, not from a model's memory, and every verdict links to those sources. The models come from a privacy-vetted roster — none of them may train on your data — and the methodology page lists it.
What can I upload? +
PowerPoint (.pptx), Word (.docx), PDF, and images (PNG/JPG/WebP) — images are read with vision to recover their text and figures. You can also paste raw text or verify a single claim without uploading anything.
Are my documents used to train AI? +
No. Your files are never used to train any model — the AI providers do not train on API data, and we don't either. You can also run in confidential mode, which verifies without any web search or images leaving our infrastructure.
Where are my files stored, and for how long? +
In the EU (Cloudflare R2, EU jurisdiction; compute in Germany), encrypted in transit and at rest. You control retention: keep files until you delete them, delete the original right after analysis, or auto-delete after 30 or 90 days.
What do the verdicts mean? +
Six: Verified, False, Partial, Disputed, Unverifiable, and Misextracted. "Unverifiable" never stands alone — the card says whether the claim rests on your own methodology or licensed data (properly sourced, and no public figure could ever confirm it), whether the document does not disclose how it was computed, or whether we simply did not find a source. It is not the same as false, and we never present a lack of evidence as a confirmation. "Misextracted" is our own defect — a claim we built that does not match the passage it came from — and it scores nothing.
Does this replace human review? +
No — it is a first pass. For legal, medical, financial or editorial decisions with real stakes, expert human review is not optional. Facticy's job is to make that review start with a map instead of blind. The methodology page lists every limitation in full.
Check your next document before your audience does
Free to start — you only pay for what you analyze.
Create free account