How Facticy reviews documents

Extract the claim faithfully

We parse the document and extract factual claims with their original passage and location. A useful comparison preserves the subject, date, unit and scope. Opinions, forecasts and attributions need different questions; a future outcome is not an observed fact.

Extraction can fail. A claim that does not faithfully represent the original is disclosed as an extraction error and excluded from the document’s factual counts.

Retrieve evidence before judging

References and structured sources can provide evidence directly. When needed, research searches for candidate sources. We open the source and keep the exact passage, its link and the date it was read. A URL or quote produced only by a model is not accepted as retrieved evidence.

Sources are labelled as official, independent, interested or the document author’s own. These are provenance labels, not guarantees. An author’s publication can establish what that author said; it does not independently corroborate the underlying assertion.

One assessment and a second check for false claims

One model compares the claim with the available source passages. Another provider independently checks proposed false claims, using the same evidence without seeing the first opinion. Both must confirm the contradiction before it is published as false.

A failed reviewer or an inconclusive comparison leaves the claim unresolved. Two reviewers reading the same material do not constitute two independent sources, and their agreement can still be wrong.

A limited second search prioritizes unresolved claims. Another assessment is requested only when that search adds new evidence; repeating an opinion over the same evidence does not count as progress.

Three visible outcomes

Partial or disputed saved verdicts are included in unresolved. Extraction errors are disclosed separately. Historical analyses without a record of source retrieval are treated as unresolved in the current review; previously generated annotated files may still retain their original verdicts.

• Supported — retrieved evidence supports the claim within the stated period and scope.
• False — retrieved evidence contradicts the claim, and a second review confirms the comparison.
• Unresolved — evidence is insufficient or the comparison remains inconclusive. The reason may be missing sources, access restrictions, an internal method, a forecast or an operational failure. Missing evidence is not evidence of falsity.

Coverage and an opinion on quality

The report opens with false claims and their original location, explanation and source passage. It shows supported, false, unresolved and analyzed counts. A result with no detected false claims describes only the verified portion; zero coverage is stated explicitly.

The editorial review considers the document and final factual findings. It provides a brief opinion, strengths and up to three priority improvements. If it cannot be completed, the review is labelled partial or unavailable rather than assigned an invented score.

Cost and privacy

The upload screen shows an estimate, not a guaranteed dollar maximum. Call and search limits and account quotas constrain usage; retries still incur costs and their earlier charges remain recorded. A run stopped at a safety limit can fail without a complete report.

Confidential mode disables web search and image analysis. Document text still goes to AI providers, so this is not a fully local analysis. External verification will be limited. You can choose to retain the file, delete the original after analysis, or delete it after 30 or 90 days. See the privacy and subprocessor pages for processing terms.

Limits of the method

• Sources and models can be wrong. A retrieved passage proves what was read, not that the interpretation is correct.
• Retrieval is bounded: some dynamic sites, paywalls, images and large PDFs cannot be read completely. Source selection and claim extraction can miss relevant context.
• English and Spanish interfaces do not guarantee equal source coverage across languages or subject areas.
• We do not claim a general accuracy rate, coverage rate or percentage cost saving. Those measures require evaluation on representative documents.
• Use the report to guide your own review, especially where a decision depends on the interpretation of the evidence.

Historical public reports and feedback

Saved public reports and Lab studies retain the outputs and measurements of the version that produced them. Their dates and methods matter; historical scores and panel results do not describe the current workflow.

Each claim’s details let signed-in users agree, disagree or suggest a correction. Feedback is useful for investigation and evaluation; it does not silently replace the original result.