Danko Komlen

Four states. Forty test cases. Four days to write the rules. We extended Sheldon, Feather's deterministic tax engine, from federal to California, Illinois, New York, and Virginia. Where it landed on TaxCalcBench, the independent benchmark we score against: ✅ Virginia: 160 of 160 scored lines, exact ✅ Illinois: 150 of 150, exact ✅ New York: 180 of 180, exact ⚠️ California: two of ten cases still open ✅ Federal: 190 of 190, held the entire time 620 of 630 state lines. 98%. The interesting part is those two California cases, because they are not our bug. We stalled at 88% and could not move. The cause was defects in the benchmark data, not gaps in our engine. We could have closed that gap in an afternoon by writing rules that reproduced the expected outputs. We did not, because of two commitments we made on day one: 📐 Every rule must be grounded in our Knowledge Base, traceable to the statute or the form instruction. 🚫 No rule gets added to overfit a test case. So we asked Michael Bock, who maintains TaxCalcBench, about the four cases we could not reconcile. He confirmed they were known issues. We had found them independently. We also surfaced one that was not: a New York case missing $500 of state withholding because of a payer-ID matching bug. The maintainers corrected it after we reported it. Our engine was right and the answer key was wrong. You only discover that if you refuse to overfit. The broader lesson: the best way to use AI in tax preparation is not in-situ. Use AI to build the rule engine, then prepare returns with the engine. That compresses writing state tax rules from the years it took incumbents, with hundreds of engineers and accountants, down to days with fewer than a dozen people. Adding a new state is now a question of days, not quarters. If you practice in a state we have not built yet, reply with which one. That is genuinely how we are prioritizing. Full write-up: askfeather.ai/resources/buil… @michaelcutajar @michaelrbock @RDriscollCPA @jasononfirms