The AI bookkeeper files bank lines with built-in keyword rules before any learning happens. To find where they are thin, we write descriptors for a kind of business not yet covered and count what the rules do with them.
This round covered florists, home inspectors, food trucks, freelance translators and event planners. Of 37 first-draft descriptors, the rules filed 8 right, 6 wrong and 23 not at all. Five of the "wrong" ones were our own expectations (repairs, rent and propane already go to their own accounts; Restaurant Depot is a supply house), and we dropped three ambiguous deposit and rental lines. The real gaps were flower wholesalers, Teleflora and Zola, trade dues (ASHI, InterNACHI, ProZ, the American Translators Association), Spectora, and translation tools such as Trados, memoQ and DeepL.
After the rules, all 34 remaining descriptors file correctly. That is tuned on the same list, so it counts coverage, not accuracy. The descriptors are made up from public company names; no real bank export was used. Labelled real books are still what the project needs most.