Our flood model passed its own fairness test. Then the audit flagged a priced variable that would have skewed premiums against the wrong neighborhoods. We built FloodProof on a real book of 170,803 NFIP flood claims from Harris County, Texas. On 40,615 held-out claims, its composite factor captured 1.69× the top-decile loss dollars the FEMA flood zone flagged, and the score itself clears the 80% fairness rule at an adverse-impact ratio of 0.938. Most tools stop there. We didn't. The audit checks every priced variable against real census-tract demographics, and five failed the 80% rule on this book. The worst was something as innocent-sounding as "number of floors," at a ratio of 0.36. Left in the rate, it would quietly push premiums against certain communities. So the code refused to file. The gate marked the package NOT FILING-READY, named the offending variable, and routed it to a human actuary to justify or drop. Agents advise, code decides. It blocks itself before the examiner does. A passing model is not a passing filing. Fairness isn't a property of the score, but of every variable you price on. Once your model drafts a rate, does anything re-check the finished factor for disparate impact before it files, or does it go out on the model's word? We'd genuinely like to hear how your team handles it. #Actuarial #InsuranceTech #AIGovernance #AlgorithmicFairness #Underwriting