A sample of entries checked in both languages, and thirty-one summaries corrected
Three models checked 300 real entries in Korean and, for the first time, in English, and a fourth weighed what they found. Thirty-one summaries said a little more than their verses, and now say only what the verses say.
Every entry summary should say no more than the verses it cites. After the test sheets of the first comparison, three models (Claude Sonnet 5.5, Claude Opus 5.5 and GPT-6 astra) checked real entries: a sample of 300, drawn by a fixed rule from every part of the collection, and 49 of the largest. For the first time the English summaries were checked as well as the Korean ones. Claude Fable 5.1 then judged each finding as real, wrong or unsure.
- About 4 in 100 Korean summaries and 7 in 100 English ones said something their cited verses do not. The most common slip was naming a speaker who is named only in the opening verse of a speech or a law, a verse the entry does not cite.
- 31 entries are corrected. Nine now also cite that opening verse, such as “The LORD spoke to Moses, saying” at the head of a Leviticus chapter. The others now say only what their cited verses say: in Job and Proverbs, for example, a name that only an uncited verse gives was taken out.
- Claude Opus 5.5 found nearly all the real problems with few wrong findings, which makes it the candidate for checking the whole collection.
This is a check by models, not a review: every entry still awaits human review.
Commits: 2c79aa7 (the code the footer’s version label names)