We propose a KS-based calibration method for survival models that avoids discretization and improves calibration while preserving predictive accuracy across real-world datasets and models.