ExtractBench measures extraction beyond answer accuracy

ExtractBench is a schema-guided document-extraction benchmark covering 4,869 pages and 370 enterprise documents; it scores value accuracy, record completeness, grounding, and measured cost together. Why it matters: An extraction can return correct-looking fields while dropping a long list or losing the page that supports it. The missing checks surface later, when someone has to trust the record.

Try this: Run ten documents from one workflow through the extractor, then measure field accuracy, completeness, source-page grounding, and cost before comparing models.