How a verdict gets made

A document is first decomposed into discrete, checkable claims. To keep this repeatable, we run several independent 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, and no model "ranking by importance," decides what gets checked. The resulting claim list is capped deterministically in document order, and its length is what 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 frontier models from different providers (see the model catalog). Every panelist answers independently — it never sees the other models' votes. Depending on the tier, a panelist argues from its own knowledge (Model Fact-Check) or researches the claim with live web search and must cite the sources it relied on (Live Research). Each vote is a verdict on a journalistic 7-point scale — true, mostly true, mixed, misleading, mostly false, false, or unverifiable — plus the model's reasoning and confidence.

When the panel is unanimous and confident, we don't rubber-stamp it: a cheap, independent challenger model — never a member of any panel — is prompted to refute the consensus. If it can't, the verdict is confirmed. 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 synthesizes the final consensus verdict. Disagreement is never hidden: dissenting votes stay visible in the report.

The confidence on a verdict is calibrated: a number we compute from observable signals, not one we take on a model's word. It starts from the judge's stated confidence and is dampened by three auditable factors: how much the panel actually agreed, how authoritative the cited sources are, and whether those citations actually 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 real caveats: some panel disagreement, weaker sources, or partial support. Read the dissent before relying on it.
• Low (< 0.5) — treat as "flagged for human review," not as a verdict to quote.

Note that the verdict and the confidence are different axes: a claim can be confidently mixed, or only tentatively false. "Unverifiable" means the panel couldn't settle the claim either way — it is not a euphemism 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 / unverifiable for comparability), alongside calibration — whether an 80%-confidence verdict is actually right about 80% of the time. We also track run-to-run stability, since a fact-checker that changes its mind between identical runs isn't one you can build on.

Benchmark results: we're iterating rapidly toward our first major quality milestone, planned for October 2026. The full benchmarking suite runs against that build, and the numbers, datasets, and per-tier breakdowns will publish on the benchmarks page — and stay updated as the model lineup evolves.

• Models can be wrong together. Panel diversity reduces — but cannot eliminate — shared blind spots: models trained on overlapping data can repeat the same popular misconception, and 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 actually support the claim, but a well-laundered falsehood with plausible sources can still get through.
• Extraction is a judgment call. Claim decomposition can occasionally 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 aren't checkable facts. These should land as "unverifiable" rather than forced into true/false.
• It's an aid, not an oracle. For high-stakes publication decisions, treat reports as a structured second opinion with an evidence trail — not a substitute for editorial responsibility.

Checks run with deterministic settings (fixed temperature and seed), claim extraction is majority-voted across parallel samples, and the calibrated confidence is a pure function of its inputs. Providers introduce some nondeterminism we can't remove, but the pipeline is engineered so identical documents produce stable claim lists and prices — and every verdict keeps its full audit trail: individual votes, the challenger's attempt, the judge's reasoning, and citations.