How a verdict gets made

We first split the document into discrete, checkable claims, in two passes: a main pass, then a recall sweep that looks only for facts tucked into asides, subordinate clauses and comparisons. Both run at a fixed temperature and seed, and every claim must anchor to the exact sentence it came from or it is dropped. No model ranks claims by "importance." The 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.

Stamp AI is built on one rule: no model answers from memory. For each claim, in this order:

• Gather. Two live-web searches start at once, built in code rather than by a model: the claim itself, and a hunt for any fact-check, correction or denial of it.
• Screen. Every result is checked for relevance and for prompt injection (text that tries to instruct the checker), and relabelled with whether it actually supports or challenges the claim.
• Rule. A fast model rules on that screened evidence alone, on the 7-point scale: true, mostly true, mixed, misleading, mostly false, false, and unverifiable. A slow answer is retried in parallel on another host rather than dropped, and no answer at all is reported as unchecked, never as "unverifiable".
• Audit. Every source the verdict cites is checked against what the source actually says, the evidence is traced to its independent origins (seven outlets re-running one wire report are one origin), and the confidence band is computed from those signals.

Every claim in a document runs in parallel, so a whole document typically completes in well under a minute.

When a claim has to survive scrutiny, the Research Panel goes deeper. Each claim goes to a panel of models from different providers (the model catalog lists them). Every panelist researches the live web on its own, never sees the other models' votes, and must cite what it relied on. When the evidence has nothing first-hand or on the record, a primary-source probe searches for what the party concerned actually said, including in its own language.

When the panel is unanimous and confident, we do not confirm it on the spot. A challenger model, never a member of any panel, researches the claim through its own search and tries 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 weighs every vote, rationale and piece of evidence, with the dissent on the record, and writes the final verdict. Dissenting votes stay visible in the report.

We show confidence as a band, not a percentage. A percentage reads like a measured probability; our score is computed from observable signals, and what we can measure is how often each band turns out right. The score starts from the ruling model's stated confidence and four auditable factors lower it: how much the voters agreed, whether their written reasoning concludes the verdict they voted for, how authoritative the cited sources are and whether those citations support the claim, and how many independent origins back the verdict. The same inputs always produce the same band.

• High. Strong agreement and citations that hold up. On our benchmarks, Stamp AI's high-confidence verdicts were right 97.2% of the time.
• Moderate. A real signal with caveats: some disagreement, weaker sources, or partial support (right 88.3% of the time on the benchmarks). Read the evidence before relying on it.
• Low. 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 evidence could not settle the claim either way. It is not a softer word for false.

We run Stamp AI, end to end, on AVeriTeC and LIAR2: real-world claims that professional fact-checkers investigated and labelled. On the 3,718 claims they settled as true or false, it called 89.8% right and 7.1% backwards. The benchmarks page has the dataset, the method, the per-label breakdown and the calibration of every band, and we re-run it whenever the pipeline changes.

• Stamp AI rarely says "unverifiable". On claims fact-checkers marked as lacking enough evidence, or as cherry-picked, it usually still picks a side. If a claim is thin or contested, read the evidence, or run the Research Panel.
• Published fact-checks carry weight. When a claim has been fact-checked before, Stamp AI finds that fact-check (it searches for one on purpose). That is usually what you want, but it means an error in a published fact-check can carry through.
• 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.
• The web itself can be wrong. We weigh source authority, trace independent origins and check 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. A verdict is an automated assessment of the claim as written, based on the evidence available when the check ran. 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, and the confidence band is a pure function of its inputs. Providers and the live web introduce some nondeterminism we cannot remove, but identical documents produce stable claim lists and prices. Every verdict keeps its full audit trail: the evidence, each vote, the challenger's attempt and the judge's reasoning where they ran, and citations.