{"name":"deterministic-ocr","title":"Deterministic OCR","keyPoints":["LiteParse reads 37 of 40 values with no model and no fee; Poppler's embedded text reads 20 of 40.","On the two image pages LiteParse reads 17 of 20 by its own OCR and Poppler none: an image has no text layer. On the two digital pages both read 20 of 20.","LiteParse's three misses: a ticked box, a minus sign (\"$10,000\" for \"– $10,000\"), and a date whose label it read as \"=\"."],"tables":[{"heading":"Test sets A and B: values read correctly, of 10 a page","columns":["Reader","A, digital","A, image","B, digital","B, image","Correct","A page","1,000 pages"],"rows":[["LiteParse (no model)","10","8","10","9","37 of 40","0.7 to 0.8 s","$0"],["Embedded text (Poppler, no model)","10","0","10","0","20 of 40","0 s","$0"]],"note":"Each reader reads the page's PDF. Digital: the PDF has a text layer. Image: the page is a picture with no text layer, which LiteParse reads by Tesseract. A page: the reader's time on one page, on a development computer, Oct 25 2026."}],"scoringMethod":"20 values on each of the published test sets A and B, each checked by a person, counted when the reading states the value where it belongs. The grader is Claude Opus 5.5 in the platform's eval run, reading the text alone; a second grading of one reading a page changed 1 of 40 verdicts. Each reader gives the same text on every run.","log":[{"date":"261025","line":"New board. LiteParse reads 37 of 40 values with no model, Poppler's embedded text 20 of 40: a tie on the digital pages, 17 to 0 on the image pages."}],"updatedAt":"2026-10-25T19:36:18.854Z"}