The open-source bookkeeper's built-in keywords suggest an account for 72.7% of a San Francisco makerspace's 2020-2023 bank lines, 96.0% of them right. We replayed the same 278 labelled lines in date order, letting the product learn from each line the treasurer filed before suggesting the next.
Result: 98.9% of lines got a suggestion, 97.1% of those right, 96.0% right of all lines. Payee memory supplied 250 of the 275 suggestions. Changing card digits and dates in the bank's descriptor did not hide a repeat payee. The remaining wrong answers are the same eight PayPal payouts the engine books as transfers and this organisation books as income.
Read it with care: this ledger is mostly repeating bills and rent, so repeat payees are most of the gain. A business with a long tail of merchants will score lower. The data is fetched at run time from a public repository and never copied; the script is eval/queeriouslabs-online.mjs.