Question answering (QA) measures consistency by turning each target sentence into factual questions, then comparing answers extracted from multiple reference samples. section 5.2.3
Apply the method in two steps:
- Generate questions about the target sentence, with each question aimed at a specific factual claim. section 5.2.3
- Extract answers from every reference sample and compare them; matching answers indicate consistency, while variation indicates inconsistency. section 5.2.3
Historical date claim
For “the Pantheon was completed in 125 AD,” generate “When was the Pantheon completed?” Answers of “125 AD” across samples support consistency; differing answers indicate inconsistency. section 5.2.3
QA enables fine-grained verification, but depends on question-generation quality: weak questions can miss important claims or target irrelevant details. section 5.2.3





