Task guide

Should I use Jev for claim and citation verification?

Verdict for the sample idea

Sample idea

Our LLM writes answers with citations. For each sentence in the answer, check if the cited source passage really supports the sentence, and flag unsupported sentences before the answer reaches the user. About 50,000 sentences a day.

Explanation

Two tasks hide behind the word "verification". Only one of them works.

The first is: does this passage support this sentence? You send both texts. Jev compares them and returns a probability. That is a closed judgment on evidence that you supply, and it fits well. At 50,000 sentences a day, an LLM judge is slow and expensive, and its answer is prose that you then parse. TypeSafe's build guide uses this exact pattern in its composite scoring example, with one Noul for "citations are supported" and one for "contradicts context".

The second is: is this sentence true? Without a source, Jev has only what it absorbed in training. It will still return a confident number. Do not build on that. A valid typed answer is not a true answer.

Shape it as one call per sentence and citation pair, with two or three Nouls. Does the passage state or directly imply the claim? Does the passage contradict the claim? Does the claim contain a number or a date that the passage does not contain? The third question catches the most common LLM error, where the sentence is right in spirit and wrong in the figure.

Split the answer into sentences in code. Then split compound sentences into single claims, if you can. "Revenue grew 12% and the company hired 200 people" is two claims, and a passage can support only one.

The common trap is numbers. Jev reads "12 percent" and "0.12" as text, and TypeSafe lists numeric formats as a weakness. When a claim depends on an exact figure, extract both figures in code and compare them in code. Let Jev judge the meaning, and let code judge the arithmetic.

Send low-confidence pairs to a slower model, not to the user.

Related tasks

• Should I use Jev for RAG relevance filtering?
• Should I use Jev for invoice totals and arithmetic?
• Should I use Jev for summaries and writing?

Verdict produced by Jev on 2026-09-19, rubric version 2. Text written by a person. How the tool works: the report.