How it works
A language model is a fine reader and an unreliable witness. So the pipeline is built to make it read, and to check what it says it read — with the last word going to code, not to prose.
We take the assertion exactly as you wrote it. If you supplied context — who said it, where, what it was answering — that rides along, because “the deadline moved to April 30” means different things in different rooms.
One angle looks for sources that would establish the claim. The other explicitly looks for sources that would refute it. Both run at once, up to 6 candidates each. A checker that only asks the first question ends up agreeing with everything it is shown.
Not snippets — the page text, with the passages the search engine judged relevant pulled to the front. If you set a recency window, anything older is never fetched, so a well-sourced but stale page cannot produce a confidently wrong answer.
The adjudicator is told it has no other knowledge, must quote verbatim, must copy URLs exactly from what it was shown, and must include the strongest evidence on BOTH sides where both exist. Omitting a contradicting source to produce a cleaner verdict is explicitly forbidden.
Each returned evidence item is checked against the pages we actually fetched. A URL that was never in that set is deleted — not corrected, not re-fetched, deleted. Titles and publication dates are then overwritten from our record, so a real quote can never appear next to an invented date.
A decisive verdict with no surviving agreeing evidence degrades. Under high strictness, one uncorroborated domain degrades to mixed. And if nothing survives at all, the verdict becomes unverifiable, the adjudicator's prose is thrown away — it was arguing from citations we just refused — and you are refunded.
What comes back
Open the URL. Search the page for the quote. That is the whole point — you never have to take the verdict on trust, because the thing that produced it is attached to it.
One source speaks to it.
The corroboration rule
Two sources only count as two if they come from different registrable domains. The comparison strips subdomains and handles two-level suffixes, and where it is unsure it treats things as the SAME publisher — the ambiguity is spent on being stricter, never looser.
A publisher quoting itself on a second page is one voice. So is a wire story reprinted under two of a group's mastheads, when those masthead pages share a domain.
The four verdicts
And the honest limits
It reads what has been published, so it inherits the web's blind spots: very recent events, paywalled reporting, non-English coverage, and anything true but unwritten. It will not tell you a claim is false because it found nothing — that is what unverifiable is for, and why we do not charge for it.
8 credits a claim. Nothing found, nothing charged.