An AI told us the new car-loan interest deduction lowers your AGI. Confidently, neatly cited, and wrong. The 20% penalty is yours, not its. That deduction (OBBBA's QPVLI) is real, but it sits below the line under §63(b)(7), so it does not reduce AGI. The mistake is common enough that even H&R Block's own site has labeled it "above the line." Now picture an AI drafting a return with that error. It reads fine, a human signs it, and the §6662 accuracy-related penalty (20% of the underpayment) lands on the person who signed, not the model that drafted. So while building StatuteGuard we put a verification layer under the draft that no language model can talk its way past. The AI drafts the position, but a deterministic policy engine decides the verdict: PASS, BLOCK, or NEEDS-REVIEW. Paste that car-loan claim and it returns BLOCK, animates the citation chain from §163(h) to §62/§63, then flags the five things that break downstream if you file it anyway. On a 42-case labeled golden set, it caught every encoded error with zero false blocks. Agent advises, code decides. One question we keep coming back to: after your AI drafts a tax position, does anything re-check the finished claim against the statute before someone signs it, or does it go out on the model's word? #TaxTech #AICompliance #TaxCompliance #RegTech