Field notes
What actually goes wrong when documents are turned into data, measured instead of assumed. Every post comes with the code, so you can rerun it on your own files.
OCR · Evaluation
We ran five OCR engines on 25 kinds of difficult page, from crooked scans and phone photos to handwriting, look-alike codes and DRAFT watermarks. Here is how each one fails, and how to run the same test on your own pages.