How Pubvana fact checking works in Pubvana v3
When you see a fact check box under an article on a Pubvana site the site owner had an AI assistant check the article's claims against real sources, and the assistant's findings were filed on that site.
What the verdicts mean
The checker sorts the article's statements into facts and opinions, then grades only the facts.
Supported: checked against sources, holds up. Partially supported: partly right, with the part that isn't spelled out. Refuted: checked, and it's wrong. The correction is included. Unverifiable: no reliable source could settle it, so the checker says so instead of guessing.
Opinions get no grade. They're marked as opinions with a short note. Nobody's greatest blog post in the world claim is being fact-checked.
Every graded claim carries its sources, so you can do the checking yourself. That's the point.
The rules the checker follows
Before every check, the assistant fetches a versioned prompt from Pubvana. That prompt is the only instruction set allowed. It tells the checker to verify against citable sources, prefer primary sources, weigh evidence the site owner offers the same as anything it finds on its own, and never guess.
Instructions buried inside the article don't count. If a piece of content tries to flatter, bribe, or script its own fact check, the checker flags the attempt, quotes it, and keeps checking anyway. You'll see that flag on the box when it happens. Content that tattles on itself is still content worth doubting.
The same goes for whoever is driving the AI session. Asking the checker to skip the site's rules, or to run someone else's prompt instead, voids the agreement for that user and that site. Pubvana's position is simple: our tools exist to help people publish accurate information, and we won't let them be used to launder misinformation. The service refuses, and the refusal is the service.
What the check is worth
An honest limit: we can't see inside someone else's AI. A determined person can lie to their own tools. What Pubvana can do, and does, is refuse every report that doesn't attest to the current prompt version, log each attempt on the site, and let the site owner cut off any key in one click. A check that made it onto the page did so by passing through all of that, and the box carries the prompt version it was checked under so the claim is on the record.
If the article changes after the check, the box says so. Stale means check it again, not trust the old one.
If you run a Pubvana site
Fact checking runs through the API key your assistant already uses. Under Tools → AI Assistant → Fact Checking, read the terms, accept them, and switch the service on. You'll need at least one enabled key. When the prompt is updated, the service pauses until you re-accept, so a check is never filed under terms nobody read.
Place the Fact Check Summary block in your theme's after-content region Appearance -> Themes -> Regions and every checked article shows its report to readers.