A fourth slice of real card memos: 40.5% before any rule saw them

ai-bookkeeper's keywords on 168 held-out Hack Club HCB card charges from page 4: 40.5% suggested, 56.0% after four tuned rules.

Tools and projects · 1 min read · Source
ai-bookkeeperaccountingevaluation

We keep measuring ai-bookkeeper's built-in keyword layer on card memos it has not been tuned on. This time: 168 card charges from the first 40 transparent organisations on page 4 of Hack Club's public HCB feed.

Held out: 68 suggested (40.5%). Earlier held-out readings were 17.6% and 38.9%, so the layer holds up on new organisations. The misses were mostly what should stay blank: bare amazon (27), Walmart, Target, Costco, groceries. The rest was a long tail: Postmark, ElevenLabs, Misumi, CannonKeys, merch printers, ezCater.

After four rules: 94 (56.0%). Written from the same list, so tuned, not held out. Memos have no category in this feed, so this is coverage; a reader's check of the hits found nothing plainly wrong. Labelled real books are still the missing piece.

Rerun: node ventures/ai-bookkeeper/eval/hcb-memos.mjs --orgpage 4 --from 0 --to 40.

More in Tools and projects

All writing · All in this topic · All tools and research