<p>This study proposes a comprehensive data-driven framework for quantifying seismic fragility and performance degradation in historical masonry pagodas. Leveraging shaking table tests on 1:8 scale models of the Small Wild Goose Pagoda in both intact and damage-inclined states, a high-resolution dataset ( &gt; 1000 samples) was constructed, capturing key seismic responses—peak acceleration, displacement, and interstory drift—under multidirectional and multi-intensity ground motions (PGA: 0·15 g–0·60 g). A multi-model machine learning pipeline combining Random Forest, XGBoost achieved high predictive accuracy (R² &gt; 0.90), while SHAP-based interpretability analysis identified ground motion intensity, story height, and damage state as dominant predictors. Fragility curves revealed a pronounced vulnerability increase for the damaged model, with a leftward shift in exceedance probabilities. A multi-task CatBoost model with split-conformal calibration provided reliable uncertainty-aware predictions, supporting risk-informed decision-making. Further, a binary classifier (AUC = 0.928) enabled effective post-earthquake condition identification, while response ratio analysis quantified residual capacity loss, particularly in interstory drift. The proposed interpretable and uncertainty-quantified framework advances seismic risk assessment and resilience planning for heritage masonry structures.</p>