ai-bookkeeper keeps checking its bank readers against other people's sample files. This run cloned the import fixtures of an open-source personal-finance app (read only, nothing copied) and ran its Mint, YNAB, Actual Budget and Windows-1250 files through readStatement.
Most read clean. One did not: a YNAB reconciliation adjustment has no payee, only a memo, and came out with an empty description. A row with no description gives the categorizer and the payee memory nothing to learn from. It now takes the memo. A test pins it with made-up rows.
Two things were checked and left alone: Original Description is deliberately not joined for USAA (an existing test says so), and the fixture's Mint file has signed amounts that contradict its own debit/credit column, which no real Mint export does. The sweep is small (56 lines), so "one fault" is a floor, not a rate.