Document reading
GPU OCR path, layout recovery and scan quality on declared corpora. Failures stay in the record.
PRODUCTDOCUMENT UNDERSTANDING
Reading is the first compile step.
Files arrive as PDFs, scans, spreadsheets, pages and repositories. The compiler has to recover text, layout and structure before anything can be compiled into a world.
Pages without a text layer
A scan is read as an image, not skipped. Uncertainty is reported rather than filled in.
Structure recovery
Headings, tables and regions have to be recovered before a fact can be attached to a place on the page.
Doubt is part of the read
Low-confidence regions stay marked. A reader that never reports doubt cannot be believed later.
Measured results and failures live on /research and /evidence. This page does not publish a closed accuracy number.