Deep Dive into Perception: Scoring Scans, Handwriting, and Rotated Pages
ExtractBench is an open benchmark for schema-guided document extraction. This deep dive looks at perception: how extraction systems handle documents that don't arrive as clean, born-digital text. ExtractBench tags every document by how the page was captured: born-digital, scanned, handwritten, or rotated. A system that reasons perfectly can still misread a smudged 5 as an 8. The two frontier coding agents we tested have a perception blind spot, in opposite directions. One reads scans and handwriting well but stumbles on rotated pages. The other handles rotation and falls apart on scans and h