All news

Evergreen checks claims in LLM-written summaries of tables by compiling each claim into a semantic verification query with citations. It keeps F1 at 0.94 while cutting cost 3.1x through early stopping, relevance sorting and prompt caching.