How a claim check should work

A useful check is a chain of questions, not a single confidence number.

A claim is a question

A fact-check starts before a search box. First, identify the claim. What exactly is being asserted? Is it about a date, a number, a person, or a cause? A sentence can contain several claims. Separating them makes the result easier to inspect.

Next, preserve the original wording and context. A clipped post may change meaning. So may a headline, translation, or image caption. A useful record keeps the input visible. It also records when the check happened, because sources and events change.

It helps to classify the claim before looking for an answer. A historical claim needs a different source trail from a prediction. A quote needs a transcript or recording. A statistic needs a definition, population, and time period. A causal claim needs more than two events that happened together. These distinctions prevent a search from answering a nearby question instead of the one a person asked.

Sources do the heavy lifting

A model can help find material and compare passages. It does not turn an unverified source into an authoritative one. Checkers should prefer primary documents when available. That may mean a filing, a transcript, a dataset, or a direct statement. Independent reporting can add context. It should not be treated as independent merely because it appears on a different website.

Good source selection is a method, not a magic list. The IFCN Code of Principles describes commitments around nonpartisanship, transparency, and corrections. The First Draft verification guide offers a practical way to assess source and content context.

Search results also have an ordering problem. The first result is not necessarily the best result. It may be recent, popular, or optimized for the query. A careful workflow opens the document and checks its author, date, evidence, and intended audience. It compares a source's claim with the claim being checked. It does not count repeated wording across copied pages as several independent confirmations.

Verdicts need reasons

A label without evidence asks for trust. A record with citations invites review. The result should state what supports the conclusion, where uncertainty remains, and which parts were not checked. Language matters. “Unsupported” is not the same as “false.” “Out of context” is not the same as “fabricated.”

Automation is most useful when it shortens the path to relevant evidence. It is less useful when it hides that path behind a score. In Prove, treat the verdict as a compact research brief. Follow the citations. Check dates. Ask whether the sources answer the claim that was actually made.

There is also a difference between evidence and explanation. Evidence is the material that bears on the statement. Explanation is the reasoning that connects that material to a verdict. A source may be relevant without being sufficient. A clear record names that gap rather than filling it with confident language.

Make the check reproducible

Another reader should be able to understand the input, the evidence, and the reasoning. That does not require every person to reach the same view. It does require enough detail to find a disagreement. This is why provenance, timestamps, and links matter.

Save the version of the claim that was checked. Note meaningful edits to the input. Preserve citations with their access dates when the topic is likely to change. If a source is later corrected, the record should make that change visible instead of quietly rewriting history. For high-stakes work, a second reviewer can repeat the search without being shown the first reviewer's conclusion. That simple separation can reveal where wording or source choice drove the result.

A good check can end in uncertainty. That is not a failed workflow. It is a precise description of what the available evidence can support today. Recheck when a primary source appears, a date passes, or the original claim changes.