ai-bookkeeper reads a merchant, date and total from receipt text. A fix that strips bracketed company numbers (AEON CO. (M) BHD (126926-H)) lifted merchant accuracy from 63.6% to 83.2% on the dataset's hand-corrected word boxes, but it had not been run on what a real OCR engine prints.
Run on the same 626 SROIE scans with tesseract 5.3.4 (--psm 6, no preprocessing): merchant exact 36.6% to 43.3% (229 to 271 receipts). Date (77.6%) and total (74.9%) did not move. Most of the remaining misses are OCR-bound: logos read as noise, or no usable line.
A second try, skipping title lines like See SIMPLIFIED TAX INVOICE and stripping a stray leading digit, changed nothing on tesseract text and cost 33 receipts on boxes (99 SPEED MART lost its 99). It was dropped. The numbers are in the venture README; the data is Malaysian flat scans, not phone photos, which is still the open question. Estimates only, not accounting advice.