Hypoglycemia remains a major barrier to safe glucose-lowering treatment in older adults with type 2 diabetes, yet clinically relevant models should combine nonlinear expressiveness, transparent feature attribution, and well-calibrated probability estimates. We developed a knowledge-constrained neural additive model (CNAM) for clinically interpretable stratification of recorded hypoglycemia history in older adults with type 2 diabetes within a multicenter cross-sectional outpatient cohort. The model was evaluated with a locked 10% test split, nested cross-validation, probability calibration, threshold-based decision analysis, decision-curve analysis, and explanation-stability assessment. It achieved competitive locked-test discrimination (receiver operating characteristic [ROC]–area under the curve [AUC] = 0.7350; precision–recall [PR]-AUC = 0.4127) and, after sigmoid calibration fitted on pooled development out-of-fold predictions, showed a calibrated Brier score of 0.1136 and an expected calibration error of 0.0189. Additional operating-point analyses showed trade-offs between sensitivity and specificity, and decision-curve analysis indicated broadly similar net-benefit profiles for CNAM and logistic regression across the evaluated threshold range. The monotonicity-constrained parameterization structurally enforces directionally plausible feature effects on selected predictors, with ablation results providing context against unconstrained variants that exhibit nonzero violations. These findings support CNAM as a complementary, interpretable stratification framework for recorded hypoglycemia history rather than as a replacement for established clinical risk scores. External validation, severity-specific outcome capture, and local recalibration remain necessary before broader clinical use.