Twelve models compared on checking summaries, and one Galatians entry corrected
Before any model checks the whole collection, twelve were tried on test sheets with planted errors. The comparison also found one real slip, in a Galatians entry.
Every entry summary should say no more than the verses it cites. A model can help find where it does, so twelve were compared before choosing one: four Claude models, three GPT models, an Ollama cloud model and four models run on a laptop. Each read summaries in which one error had been planted on purpose, such as a changed number, the wrong speaker, a command told as done or a vision told as an event.
- Claude Sonnet, Opus and Fable and GPT-6 astra and sol found every planted error. The models run on a laptop found about half of the harder ones.
- The entry on Galatians 2:11–14 said that WEB closes Paul’s words to Peter at the end of 2:21, a verse the entry does not cite. It now says only what 2:14 shows. Nine of the twelve models pointed it out.
This is a test of the models, not a review: every entry still awaits human review.
Commits: ba5b020 (the code the footer’s version label names)