If your name goes on the piece, a confident summary is worthless. What you need is the source, the sentence, and what argues against it.
For professional verification, the useful part of an AI tool is not the verdict but the evidence chain: every statement either backed by a quote located verbatim in the cited source or marked as unsourced. wyper publishes how often that holds, 73.1 percent verbatim across 633 re-fetched citations, along with the failure modes, so the tool can be audited rather than trusted.
Traceability. Every statement must say what backs it. Ours are labelled either with a sentence located verbatim in the cited source or as coming from the model with no source behind it. That distinction is the difference between a research aid and a plausible-sounding risk.
A published error rate. We re-fetched 633 real citations and checked them: 73.1 percent were verbatim on the page. The rest is published too, because it is what you need for planning: 19 percent of pages block automated readers entirely, through 403s, PDFs and paywalls, and 6.4 percent could not be found at all.
Retrieval that reaches the root. Across 24 claims with the deciding document named in advance, the root was reached in 23. Primary documents were 63 percent of everything returned, legacy media 14 percent, other fact-checkers 1 percent. We also published the first version of that evaluation, which was wrong and made us look worse, next to its correction.
Standard checks are for triage. For a claim that carries a story, Full Spectrum returns the strongest counter-positions, the spread of sources and the open questions, with a Triple-AI reading: three models on one shared evidence set, so a disagreement is a different reading of the same documents rather than a different search result.
The Booster puts twelve independent passes on a single claim, each asking something different, from what the primary documents record through the chronology, the base rate, who benefits, what would falsify it, and sources outside the English-language mainstream, plus a separate multi-hop research run with its own citations. Measured on a real run: 24 counter-positions instead of 3, 48 sources instead of 8, and 36 statements of which 31 carry a verified quote.
There is a public dispute register, append-only and hash-chained, so an entry cannot be quietly edited or removed once it exists. If a verdict is contested, the contest stays on the record. For work that may be challenged later, a corrections trail that cannot be rewritten is worth more than a support address.
Sources are never filtered by origin, and that is a measurement rather than a promise: no mainstream bonus, no preference for official bodies, no shortcut through other fact-checkers. For cross-border reporting this is the practical point, because a story that exists only in English coverage has a hole in it, and our retrieval test deliberately included non-English roots.
Free: 5 checks a day, no account needed. Every statement tells you whether a quote backs it.
wyper