Until now ai-bookkeeper's keyword layer was scored on made-up merchants. This run scored it on a real small organisation: a San Francisco makerspace published five years of its bank exports (402 lines, 2019 to 2023) with its own category on each line.
Before: the built-in keywords suggested an account for 9.4% of the 278 lines whose label names one account. Real descriptors are bank text, not merchant names: ACH DEBIT PGANDE -WEB ONLINE, GOOGLE*CLOUD 01E INTERNET, DEBIT CARD #XXXX 10/31 MONKEYBRAINS 415-974-1313 CA.
After adding PG&E's truncated spellings, small ISPs, Google Cloud and other hosts, Patreon as sales and "secretary of" as a filing fee: 71.9% suggested, 96% of those right. The 8 wrong ones are PayPal payouts booked as a transfer where this organisation counts them as income. With a PayPal account in your chart a transfer is right, so that stays.
Why it is smaller than it looks: one organisation, a handful of recurring vendors carry most of the gain, and the rules were tuned on the file they are scored on. Another organisation will land well below 72%. What no text rule can fix is CHECK WITHDRAWAL (their rent, 48 lines) and DEPOSIT - BRANCH; that is what payee memory and the review queue are for.
The scoring script is eval/queeriouslabs-bank.mjs. The data has no licence file, so it is fetched at run time and never copied into the repo.