Scoring a receipt reader on 626 real receipts: dates went from 11% to 97%

ai-bookkeeper reads a receipt's merchant, date and total so it can match the receipt to a bank line. Its last real test set, CORD, masks merchants and dates, so those two readers had never met a real receipt.

SROIE 2019 has 626 Malaysian shop receipts with the company, date and total labelled by hand. Scored on it, the reader got 11% of dates and 59% of totals right. The causes were plain:

• 09/01/2019 was always read month-first. Malaysian receipts are day-first, so it came out as 1 September. The reader now takes the order from another date on the page (a first field above 12), or from a locale marker (RM, SDN BHD, GBP), and otherwise keeps month-first but says it guessed.
• Dates written 12-01-19 or 2018/02/22 were not found.
• A GST summary table at the bottom repeated TOTAL, and lines like TOTAL GST or TOTAL INCLUDES 6% GST were read as the total.
• A printed ROUNDING ADJ -0.01 was ignored, so the total was a cent off the bank line.

After the fixes: dates 96.7%, totals 96.2%, merchant exact 63.7% (93% right or containing the label). CORD did not move (96 of 98 totals).

Limits: the text is the dataset's hand-corrected word boxes joined into rows, not tesseract output. The labels also disagree with each other on rounding and on brand against registered company name, so the merchant figure undersells the reader. The script, eval/sroie-receipts.mjs, fetches the data at run time and copies none of it.