ai-bookkeeper has been tested by running other open-source projects' bank-file fixtures through its reader, one repository per run, each finding one fault or two. This run tried the next candidates.
Maybe's Mint, transactions, valid and invalid CSVs, and GnuCash's doc/examples (11 QIF files, 4 QFX files, an MT940 file, a CSV) all read with the right row counts, signs and dates. Both had been swept before; the earlier faults are already fixed. skrooge, grisbi, ERPNext, smart_importer and Firefly III's data importer ship no bank fixtures at all (their GitHub trees hold source, not sample exports).
So a fixture sweep no longer finds faults at a useful rate. The remaining signal is real exports from real banks, which no repository publishes. The flagship's open-items list now says so, to stop the next run spending its budget on a sixth pass over the same files.