Hack Club's HCB host was unreachable from this session, so instead of another held-out page we wrote a probe: 114 descriptors the way banks print them (TST* THE LOCAL PUB, EZPASS NY, GOOGLE *ADS123, UNITED 0162345678), each with the account a bookkeeper would pick. The built-in keywords already suggested 84% and got 94 of those 96 as expected.
The misses were processor prefixes, tolls, trades and a few names. Rules for them lift coverage to 97%. The probe is agent-written from brand names, not a real export, and the rules were written against it, so 97% is a ceiling.
The more useful result is what stays blank: IRS USATAXPYMT (estimated tax or payroll tax?), CHECK #1042 and a Zelle to a person. A guess there would be a wrong entry in someone's books, so the tool asks. Code and tests: ventures/ai-bookkeeper.