Validation-Aware Rule Extraction for Post-Earthquake Building Screening: Explanation Reliability Under Spatial Validation

Post-earthquake screening requires inspectable predictions, but extracted rules may change with fitting and validation context. Using an inherited LightGBM predictor (Teacher) and 10,092 buildings from Hatay after the 2023 Kahramanmaraş earthquakes, we compared nine rule-extraction methods under repeated Random and neighborhood-held-out Spatial validation. Fidelity to the Teacher, observed-damage Macro-F1, repeated-fit agreement, and Structural Complexity were evaluated separately. C5.0 Rules had the highest mean Fidelity (0.827 Random and 0.719 Spatial), RuleFit-OvR the highest Random held-out κ (0.815), and RuleCOSI+ the highest Spatial observed-damage Macro-F1 (0.439). Moderate-damage recall remained weak and metric rankings changed with input representation. All 60 Stage-2 models passed fresh-process replay after fit-time state capture. At the 99% development-agreement checkpoint, RuleCOSI+ and RuleFit-OvR showed the largest median paired reductions in decision units, yet held-out agreement fell below 0.99 in 27 of 60 models. JRip yielded 226 threshold-independent structural families. Only one met the prespecified cross-protocol recurrence criterion, and it had no Spatial held-out activity. Engineering examples exposed actual rule routing and local errors beyond a class label. The findings support validation of explanation reliability alongside predictive fidelity and observed damage, while field utility and independent-event transfer remain unvalidated.