We fed 40 real bank-export layouts to our bookkeeper. Six read nothing.

Hand-made test fixtures only prove what we already thought of. So we took the 40 sample files in the MIT-licensed bank2ynab project, which use real banks' column layouts, and ran each through the ai-bookkeeper importer.

Most read fine. Six did not:

• German card export. Buchungsbetrag (amount) and Transaktionsbeschreibung were not known headers, so no rows came in.
• UK card export. The date column is Date Processed, and unsettled rows say Pending where a date should be. Now both are understood, and pending rows are reported as pending.
• Singapore statement. Each transaction is two lines, and the second has no date or amount. Every second line was dropped with an "unreadable date" error; it now joins the description above, so "DEBIT PURCHASE" becomes "DEBIT PURCHASE 15/04/18 … UBER EATS".
• Czech export, Windows-1250. With none of the usual giveaway letters the file decoded as Western text and nothing matched. The decoder now checks Czech and Polish column names before settling on the default.

Two single-row files are too short to infer columns from, and one pipe-delimited record file is not a CSV; those stay unread. Four tests cover the fixes. The rows in them are invented, only the layouts are borrowed. We still have no current export from a named bank.