Most of the ai-bookkeeper's tests use made-up rows. This week it was scored on two real Apple Card exports found in public repositories: 98 purchases, with Apple's own category on each. Only 58 carry a label that names one business account (restaurant, fuel, utilities); the rest, such as "Shopping" and "Other", name none and were not scored.
Before any change the built-in keywords got 56.9% of those 58 right, 75.9% were given any suggestion, and 75.0% of the suggestions were right. The clearest error: "H-E-B GAS/CARWASH" was filed as meals, because the grocer's name matched the meals rule before the word "gas" was read. A rule for fuel pumps, car washes, garages and parking now sits ahead of it, with cruises as travel and YouTube Premium as a subscription.
After the fix: 63.8% right of all, 80.4% of suggestions right. That is tuned on the same 98 rows, so it is not a held-out number. Nine 7-Eleven rows labelled "Gas" stay filed as meals, because a convenience store is both and the line goes to review either way. These are personal cards, not a business's books; a business card export is still wanted.
The script is ventures/ai-bookkeeper/eval/applecard.mjs. It fetches nothing and prints no cardholder name.