FAKTS is an AI-assisted assessment tool. This page describes how the platform is built and run, in the spirit of an AI management system such as ISO 42001. It is written honestly: what we do, what we do not do, and where a person still needs to be the last reader.
Human oversight is available, by design
For anything with editorial responsibility, medical or legal weight, or a regulated audience, a documented human review is available through the Comply layer and the Experts add-on. Comply produces a signed compliance record that names the human reviewer. Experts attach a named subject-matter reviewer to the result. Their sign-off is visible on the certificate.
Surface uncertainty, do not overclaim
The consilium is calibrated to reach for unverifiable before it reaches for verified when the evidence is thin. The five verdicts include a real "I cannot rule" outcome so the system does not paper over uncertainty. Confidence is exposed as a number, and dissent is shown as text inside the result card.
What we do with your content
Content you submit for checking is processed only to produce the result and the artifacts attached to it (the verdicts, the sources, the certificate, the seal). FAKTS does not use your submitted content to train models operated by FAKTS. [Confirm this is accurate before publishing.] The claims extracted from your content are sent to third-party model providers acting as processors so they can respond; that traffic is governed by the terms listed on Sub-processors.
Access controls and data minimization
Your rows are protected by row-level security so one account cannot read another's private data. Access to production data is restricted to a small number of people who need it, with logs and review. Personal data is minimized: FAKTS holds an email, an optional profile, and the content you actively submit. Rate-limit records use a hashed IP, not the raw address. See Privacy Policy for the specifics.
Model changes and calibration
The consilium is a moving thing. Providers update models, deprecate them, and add new ones. FAKTS runs a benchmark of claims with known answers to keep the weighting honest as the panel changes over time. When a model is added or removed, the aggregate calibration is re-fit; individual verdicts are timestamped so an older check remains readable as a snapshot of that moment.
Known limitations
- Freshness. Very recent events sit awkwardly with any source-based method; emerging is the correct verdict for a claim whose evidence is still forming.
- Coverage. Some domains are thinly sourced online; unverifiable is the honest outcome, not a hidden failure.
- Language and culture. The consilium is strongest in English and major world languages; performance on smaller languages is a work in progress.
- Adversarial content. Claims constructed to defeat verification are a real category; the corrections route exists in part to catch them.
Feedback and corrections
Every result surface has a compact "Report an error" action. Reports are reviewed by a person. Confirmed errors are corrected, and the correction is public. See Corrections.
Governance
Editorial control of FAKTS sits with RIZZ. Changes to the method, the verdict definitions, the seal bands, or the source-weighting policy are announced on this page before they take effect and are versioned by date at the top of the page they belong to.