- The new OBBBA car-loan interest deduction is above-the-line and reduces AGI, right? Mainstream tax-prep guidance says yes. H&R Block's site says yes. It is below-the-line, and the 20% penalty for filing it wrong lands on your signature, not the AI that drafted it. 🧵
- This is the step nobody is verifying. The industry automated tax drafting and left the highest-penalty question, is this position defensible under the statute, to the same model that wrote it. The failure is not bad grammar but a confident answer on the wrong line.
- You cannot trust an LLM to police an LLM. A self-check runs back through the same weights that produced the error, so it inherits the same blind spot. Confidence is not proof. We built StatuteGuard as a verification layer outside the model, not another drafter.
- What decides the verdict is deterministic policy code, not a language model. Agent advises, code decides. The AI may extract the claim from messy text. Whether that claim survives the encoded statute is decided by readable Rego you can confirm against the law yourself.
- Paste that OBBBA position and it returns BLOCK, then animates the citation chain: §163(h)(1), §163(h)(4)(A), §63(b)(7), §62/§63. QPVLI is below-the-line under §63(b)(7). It does not reduce AGI. Not a probability score. A traced path through the statute.
- Then the part filers miss: a 5-way downstream cascade. Put the deduction on the wrong line and it does not stop there. AGI, state tax, IRMAA Medicare premiums, the medical-expense floor, and student-loan IDR all shift. One misplacement, five errors deep into the return.
- It does not pretend to know everything. A 280A home-office question is a facts-and-circumstances gray area, so it returns NEEDS-REVIEW and routes to a human instead of guessing. Deterministic coverage is stated honestly at 71.4%. The rest is escalated, never bluffed.
- On a 42-case labeled golden set: 100% gate precision, zero false blocks. A correct position is never wrongly flagged. That describes the layer, not the model, so a better model does not change it. It still cannot prove to an examiner which provision backed which position.
- For tax-tech and AI builders: on the one step that carries a 20% penalty, do you trust a smarter model's self-check, or a deterministic rule the model cannot override? Where does your line actually sit? #TaxTech #AICompliance #AIGovernance
- We built this as a demo: synthetic positions, cached-model replay, fully local so nothing leaves the perimeter. If your team is working out how to file AI-drafted positions defensibly, we would like to compare notes. Walkthrough: https://veriprajna.com/demos/tax-compliance-ai
Published on X · July 14, 2026
On social media
See this post on its original platform
In our archive