What multi-model consensus catches that single models miss

A single model doesn't fail like a broken calculator, loudly and obviously. It fails like a persuasive colleague with a false memory: fluent, specific, and wrong. It will invent a plausible statistic, misattribute a quote, or date an event two years off — and package the error in the same confident prose as its correct answers. If one model is your fact-checker, its blind spots are your blind spots, and you have no instrument that can see them.

The engineering answer is old and boring, which is why we trust it: independent redundancy. Get several estimates from systems that fail differently, and compare.

Every claim we check goes to a panel of frontier models from different providers — different training data, different cutoffs, different failure habits. Each votes independently, without seeing the others' answers. When one model confidently asserts something the rest refute, that's not noise to be averaged away; it's the single most useful signal in the whole pipeline. A confabulation is rarely shared: the model that "remembers" the invented statistic is outvoted by the ones that never saw it, and the disagreement itself drags the consensus confidence down — visibly, in the report, with the dissenting vote on the record.

Knowledge cutoffs work the same way. A model that stopped learning in 2024 will confidently describe a 2025 state of the world. On a panel, its vote collides with models that know better — and on live-research checks, with models that just looked it up and have to cite what they found.

Here's the uncomfortable part most "ensemble" pitches skip: models trained on overlapping internet data can all inherit the same popular misconception. On those claims the panel is unanimous, confident — and wrong together. Agreement is evidence, not proof.

So unanimity gets the most suspicious treatment in our pipeline, not the least. Before a unanimous, confident verdict is confirmed, an adversarial challenger — a cheap, independent model that never sits on any panel — is prompted to do one thing: refute it. Not re-derive it, not summarize it; attack it. If the challenger finds a crack, the case escalates to a full judgment with that dissent on the record. If it can't, you get a verdict that survived an active attempt to break it, which is a categorically stronger claim than "four models agreed."

When the panel splits — or the challenger lands a hit — we don't take a majority vote. Counting heads treats a well-evidenced minority the same as an outvoted hallucination. An independent judge model reads every vote with its full rationale and cited evidence and synthesizes a verdict on a 7-point scale, where "misleading" — literally true but missing the context that changes the takeaway — is a first-class outcome rather than a rounding error toward true or false. Three shallow "seems true" votes can lose, and should lose, to one vote that traces the claim back to a primary source.

The confidence score you see at the end is computed, not self-reported: the judge's confidence, dampened by how much the panel actually agreed, how authoritative the cited sources are, and whether those citations really support the claim. The full formula is on our methodology page.

Here's a claim from our public sample report, from an article asserting that every French Fifth Republic president except de Gaulle and Macron has a major corruption scandal to their name. The panel split three ways: one model ("sonnet") called it mostly true — the pattern is broadly defensible; two ("gemini" and "gptmini") called it false, working president by president; and one ("deepseek") said unverifiable, because the source article only evidenced two of the six presidents it indicted.

Each model, alone, would have shipped a different answer — and each answer is defensibly reasoned. The judge's synthesis is where it resolves: the claim is solidly true only for Chirac and Sarkozy, weaker for Giscard and Mitterrand, and fails outright for Hollande and Pompidou — and a "every president" claim is falsified by a single covered exception. Verdict: mostly false, at low confidence, with 40% panel agreement on the record — the overreach fails, the judge notes, not the underlying pattern. That's a nuance no single vote contained: the mostly-true model saw the pattern, the false models saw the exceptions, and the verdict keeps both.

None of this asks for trust: every report keeps its audit trail — each model's vote and reasoning, the challenger's attempt, the judge's synthesis, and the citations behind it. Run a document with a claim you already know is subtly wrong and watch where the panel splits. The disagreement is the product.