ai-bookkeeper reads a phone photo of a receipt four times (two tesseract layout modes, on the photo and on a grey copy) and lets the readings vote on the total. The tax is read next, and most of the tax checks lean on that total: the subtotal plus the tax should make it. So when the total is wrong, the tax usually goes with it. On the 200 public ExpressExpense receipt photos (SRD), 27 of the 41 receipts with no tax had a wrong total too.
What the readings held
We printed every reading's subtotal, tax and total lines for those photos. Most are lost causes: thirteen have no total or tax line in any reading, and in three more the total is never printed legibly anywhere ("Visa $29.5", "Order Tota! A Raa"). Two patterns were not lost causes.
The word TOTAL was damaged in every reading. Photo 1133 reads "SUBTOTAL: 134.46" and "Tax(9%) 12.11" cleanly in two readings. The total line comes out as "=—STOTAL: = -:146.57" and "SOSTOTAL: ~=—'146.57". Neither matches the word "total", so no reading takes 146.57 as the total, and a fourth reading's "A TOTAL: 12.11", the tax read as a total, won the vote.
The fix puts the receipt's own arithmetic to the vote: a reading's subtotal plus its tax becomes a candidate, but only when that exact amount is printed somewhere on the page. Each reading that prints it gives one point. A reading that read a total off a real total line still wins a tie, and the receipt says "the subtotal 134.46 plus the tax 12.11, printed on the page: check the total". The same rule fixed photo 1002, whose 7.61 was printed only on the card line ("Visa 7.61") while its total line read "Total 1.61".
A leading zero. Photo 1030 has "TOTAL: 04.78" in one reading and "TOTAL: $94.78" in another. They tied, and the earlier reading won. A printed total never starts with 0 dollars, so a total line read that way now counts as a guess (half a vote).
Results
SRD, CLI --four-pass | before | after |
|---|---|---|
| totals right (196) | 151 | 154 |
| tax right (185) | 132 | 133 |
| item splits kept (200) | 49 | 49 |
Three gains, none lost. This is in-sample: the rules were written from these photos. The held-out sets we have (CORD and SROIE) are dataset text, one reading per receipt, where the vote never runs, so they are unchanged and say nothing about it. A held-out set of photos read four times is the next thing to build.
One rule was tried and dropped. A reading whose cash handed over less change disagrees with its own total line would lose its total-line point. That was meant for photo 1022 ("TOTAL: 7.07", "CASH 100,00", "CHANGE 62.63", total 37.37). It changed no photo, because "100,00" with a comma never reads as an amount on a receipt written 12.50. We did not widen the comma rule for one photo.
The write-up with every photo's case is in eval/receipt-total-first-2026-11-16.md.