Verifying LLM Answers Against Their Source

It checks whether an LLM's answer follows from the paragraphs it was given. One Jev request asks three Noul questions about the paragraphs and the answer: answerable (do the paragraphs state every fact the question needs), declines (does the answer say the paragraphs lack the answer), and supported (following the paragraphs step by step, do you arrive at the LLM's answer).

Install openai, ipython, matplotlib and cooksafe>=0.2.0,<0.3.0, then set TYPESAFE_API_KEY and OPENAI_API_KEY. The OpenAI key is used for the model whose answers are checked and, in the measurement, for settling answers that disagree with the dataset reference. A keyless client falls back to cache-only replay.

A few lines of code turn the three Noul probabilities into one risk number where higher means the answer is more likely wrong. If the LLM declined, risk is the chance the paragraphs do answer the question; if it answered, risk is the chance the paragraphs do not lead to its answer. A risk of 0.2 or more sends the answer to review.

Questions come from MuSiQue (Trivedi et al., CC BY 4.0), a fixed random sample of 1,000 development-set questions, each joining two to four Wikipedia questions with 20 paragraphs. Answers were written by gpt-6-luna, gpt-6.1-sol settled disagreements with the reference, and the check ran on jev-1.14.0 on 2026-11-07.

Of the 1,000 answers, 88.5% were right. The check passed 73% of answers and 95.7% of those passed answers were right. The remaining 27% went to review, and that review set contained 84 of the 115 wrong answers, so a share of wrong answers still passed the gate.

It takes a question, its paragraphs and an answer, and returns the three probabilities (answerable, declines, supported), the combined risk, and a pass or review verdict. The cookbook also uses cooksafe helpers such as JsonCache and make_playground_link for caching and inspecting runs.

It is built for answers grounded in retrieved paragraphs, so it does not apply to open-ended generation with no source text. It also needs both a TypeSafe key and an OpenAI key for the full run, and the published configuration used 4 Jev workers because more workers hit the public rate limit.