bayesianIt checks the message claim by claim, and it shows its work.
Paste a text and the engine splits it into atomic claims, works out what each word refers to, and judges every claim against a knowledge base of rules backed by evidence. Each verdict is an opinion in three parts, belief, disbelief and uncertainty, so a claim nobody has evidence for reads as uncertain rather than 50% true, and a claim that is literally true but misleadingly framed is flagged next to its standard description. Every verdict shows the rules that decided it and each source behind them, with its kind, date and dependence group, plus the fingerprint of the parameters that produced the numbers.
What a verdict looks like
One sentence from the test set: every word in it is defensible, and the sentence is still a lie. Open the full verdict page.
Loading the sample…
Results from the media run
We ran the validator on 735 news articles in fast mode, with the framing review on. The articles came from the five most-read outlets in each of 50 countries and were fetched on 15 and 16 October 2026. We then sent one article per outlet through the research step.
Without research, the fast pass decides almost no claim against the world. In the media ranking it therefore measures how a text frames its claims, not whether they are right. Every contradiction found in the verified sample lists its source, so a person can check it. The full report covers the method, the rubric and what is still missing.
Independence rules
Verdicts follow an audit model: they are never bought, and the commissioner does not shape the result.
- Anyone may commission a verdict.
- The result is published whatever it says. Publishing makes it permanent; the commissioner cannot edit or withdraw it.
- The commissioner is named on the verdict page.
- The full chain is shown: every rule, every source, the parameter fingerprint.
- Right of reply: the subject of a verdict can answer, and the reply goes into the knowledge base as evidence.
- A client may add evidence but can never delete or outweigh the base evidence.
- No ads, and no sponsor chooses what gets checked.
The verdict rates the message, not the messenger: it judges claims, not outlets or authors.
Two passes
Fast pass
About 20 seconds per article. The engine parses the text, runs the framing review and looks up the knowledge base. It catches misleading framing and anything the knowledge base already knows. A claim with no evidence yet stays uncertain.
Verified pass
Takes minutes and runs as an async job. The engine researches the most checkable gaps (identifiers, numbers and dates first) against primary records, stores what it finds as evidence, and judges the text again. Gaps still open can then be investigated further, with as much research as you choose.
What it does not do yet
- A fast pass decides little without research: in the calibration baseline, 90% of claims came back as “insufficient evidence”.
- Calibration labels for documents were assigned by a model from primary sources and are provisional. The benchmark rows are small samples. See Calibration.
- This demo runs on one machine through the
claudeCLI, with a daily research budget per API key.
For developers
The website runs on the same JSON API you can call yourself: POST /api/v1/evaluate for a fast pass, POST /api/v1/jobs for a verified pass, and POST /api/v1/verdicts to publish a verdict. The same engine is also available as an MCP server. API documentation.