We built a flood-pricing model that beats the FEMA zone. Then it refused to file its own rate. On purpose. We pointed FloodProof at a real book of NFIP claims, 170,803 federal records from Harris County, Texas. A composite of real building attributes (age, floors, elevation, code) ranked flood losses 1.69x better than the FEMA flood zone out-of-sample, on 40,615 held-out claims. The same zone map left $3.1B, 45.9% of building-claim dollars in this book, outside the lines it calls high-risk. So we had a sharper factor. That is the easy half. The hard half: a factor that ranks loss well can also carry demographic signal, and a rate that quietly does that is a filing your Department of Insurance examiner rejects. So FloodProof audits every priced variable against real census-tract demographics using the 80% rule, and it holds its own score to the same bar. The composite passes (adverse-impact ratio 0.938). We did not rig it to look like the villain. But five variables fail. The worst is number of floors, at 0.36, well outside the fair band [0.80, 1.25]. At that point a deterministic gate takes over. It marks the filing NOT FILING-READY, names floors as the offender, and routes it to a human actuary instead of shipping the rate. Filing is ready only when there are zero blocking findings. Agents can advise, but the code decides. One honest note: this is one Harris County book, one held-out split, and the numbers recompute live (lift lands near 1.69 to 1.85x across seeds). The vendor and Guidewire feeds are stubs behind a real schema. If your team is working out how to put an AI pricing factor in front of a DOI examiner without shipping something it bounces, we would like to hear how you are approaching it. 🌊 #Actuarial #InsurTech #AIGovernance #FloodInsurance #AlgorithmicFairness