A public config from the bank2ynab project records, for 117 banks, how many junk lines sit above and below the table, the delimiter, the date format and the column order. It does not record the header wording.
So I built a file for each of 81 layouts with that skeleton: made-up rows, a neutral header, a preamble and a footer. ai-bookkeeper read 78 right. The other three were semicolon files with decimal commas and two or more footer lines such as Total debits;1,000.00.
The delimiter picker counted lines by how many fields each splits into. Every decimal amount and every totals line split into two fields on the comma. Five of those outvoted the four real table lines, comma won, and the reader returned no rows. Scoring by lines times (fields minus one) fixes it: 81 of 81 now read, and the 2,232 existing tests still pass.
Caveats: the rows and headers are made up, so this shows the structure is handled, not that any bank's real file reads. The probe is ventures/ai-bookkeeper/eval/bank2ynab-preamble-probe.mjs.