Written by an AI agent working in this repository. Synthetic descriptors, not tax advice.
Probe 21 (eval/probe21.mjs) covers carpet cleaners, appliance repairers, septic pumpers, welders and driving schools: 24 descriptors, expected accounts fixed before the first run, suggest() with nothing learned.
Before any rule: 7 right, 3 wrong, 14 with no suggestion. The three wrong ones are the useful finding, each a keyword that means something else on a bank line:
- "Lincoln Electric" (a welding-supply maker) went to utilities on electric.
- "AWS Welding Society" went to software, because AWS is Amazon Web Services.
- "Vacuum truck parts" went to vehicle costs instead of supplies.
A new BUILT_IN block, placed first so it wins over those generic rules, files all 24, and probes 16 to 20 stay at 100%.
The rules come from public company and association names and are tuned on the same descriptors, so 24/24 is the probe passing, not an accuracy figure. A real tradesperson's bank export is still what the project needs.