How a verdict gets made

We first split the document into discrete, checkable claims. To keep this repeatable, we run several extraction passes in parallel and keep only the claims a majority of passes agree on, matched by their exact location in your text. No single pass decides what gets checked, and no model ranks claims by "importance." The claim list is capped in document order, and its length prices the check. The quote you see before submitting is the number of extracted claims times a per-claim rate.

Each claim goes to a panel of models from different providers. The model catalog lists them. Every panelist answers on its own and never sees the other models' votes. In Model Fact-Check, a panelist argues from its own knowledge. In Live Research, it searches the live web and must cite the sources it relied on. Each vote is a verdict on a 7-point scale, plus the model's reasoning and confidence. The scale runs true, mostly true, mixed, misleading, mostly false, false, and unverifiable.

When the panel is unanimous and confident, we do not confirm it on the spot. We prompt a cheap challenger model, never a member of any panel, to refute the consensus. If it cannot, we confirm the verdict. If it can, or if the panel was split or nuanced to begin with, a separate judge model weighs every vote, rationale, and piece of cited evidence, with the challenger's dissent on the record, and writes the final consensus verdict. Dissenting votes stay visible in the report.

The confidence on a verdict is calibrated. We compute it from observable signals rather than taking a model's word for it. It starts from the judge's stated confidence. Four auditable factors then lower it: how much the panel agreed, whether the agreeing models' written reasoning concludes the verdict they voted for, how authoritative the cited sources are, and whether those citations support the claim. The same inputs always produce the same score.

• High (≥ 0.8). Strong agreement and, on researched checks, citations that hold up. Act on it.
• Moderate (0.5-0.8). A real signal with caveats: some panel disagreement, weaker sources, or partial support. Read the dissent before relying on it.
• Low (< 0.5). Treat it as flagged for human review, not as a verdict to quote.

The verdict and the confidence are separate. A claim can be confidently mixed, or only tentatively false. "Unverifiable" means the panel could not settle the claim either way. It is not a softer word for false.

We measure the pipeline two ways: on public claim-verification datasets with ground-truth labels, and on curated documents seeded with a known mix of true, false, and misleading claims. The headline metric is verdict accuracy against those labels, with the 7-point scale collapsed to supported, refuted, or unverifiable so datasets are comparable. We also measure calibration: whether an 80%-confidence verdict is right about 80% of the time. And we track run-to-run stability, because a fact-checker that changes its verdicts between identical runs is not one you can build on.

Benchmark results. Our first major quality milestone is planned for October 2026. The full benchmarking suite runs against that build. The numbers, datasets, and per-tier breakdowns will publish on the benchmarks page, and we will update them when the model lineup changes.

• Models can be wrong together. Panel diversity reduces shared blind spots but cannot remove them. Models trained on overlapping data can repeat the same popular misconception, and the consensus then looks stronger than it is.
• Knowledge cutoffs. Model Fact-Check argues from training knowledge and can be stale on recent events. Use Live Research for anything time-sensitive.
• The web itself can be wrong. Live Research weighs source authority and checks that citations support the claim, but a falsehood repeated by plausible-looking sources can still get through.
• Extraction is a judgment call. Claim extraction can split a claim awkwardly or miss an implied one. Very long documents are capped at a fixed number of claims, in document order.
• Opinions, predictions, and vague claims are not checkable facts. They should land as "unverifiable" rather than be forced into true or false.
• Reports do not replace an editor. For high-stakes publication decisions, treat a report as a structured second opinion with an evidence trail, and keep editorial responsibility with a person.

Checks run with a fixed temperature and seed. Claim extraction is majority-voted across parallel samples. The calibrated confidence is a pure function of its inputs. Providers introduce some nondeterminism we cannot remove, but identical documents produce stable claim lists and prices. Every verdict keeps its full audit trail: individual votes, the challenger's attempt, the judge's reasoning, and citations.