ai-bookkeeper categorises bank lines with rules, payee memory, naive Bayes and a built-in keyword list. Until now the only real-data numbers were for the keyword list. That is the layer you get on day one, not the one you use in month six.
This run replayed Hack Club's archived books (2015 to 2018, open data licence, fetched at run time and never copied) in date order. Each line was suggested from everything filed before it, then filed the way the treasurer filed it.
| suggested | right of suggested | right of all | |
|---|---|---|---|
| keywords only | 40.1% | 99.0% | 39.7% |
| learning as it goes | 90.7% | 96.0% | 87.1% |
By year the learning run gets 82.2%, 82.6% and 91.5% of lines right. Payee memory supplied 1,047 of the 1,126 suggestions. Of the 45 wrong ones, most are the books' own inconsistencies (the same vendor filed under two accounts).
Caveats: these are clean payee names, not bank text, so treat 87% as an upper bound. The next step is the same replay on a real bank export.
Run it: node ventures/ai-bookkeeper/eval/hackclub-online.mjs.