Hierarchical Shrinkage Baselines with Independent Absolute Thresholds for Context-Adaptive Evaluation of Commercial Driver Behavior

A driver’s history serves as a reference for flagging unsafe behavior until it absorbs the behavior. Under a shrinkage profile with pseudo-count κ0 = 30 segments, the expected gap between the profile and a stable braking-composite mean of 2.82 (population mean 2.35) falls from 0.29 at 18 accumulated segments to 0.043 at 300, and the deviation test loses the signal. To prevent this degradation, the proposed architecture couples an empirical Bayes shrinkage baseline with an invariant absolute threshold, evaluating behavioral features against both individual history and safety envelopes simultaneously. A deviation triggers an alert whenever either baseline condition is met. This dual-layer hierarchy directly informs the composite safety and efficiency indices, governs directional conflict tests, and calibrates the automated feedback cadence based on observed driver responsiveness. In one simulated twelve-week fleet of 42 drivers, 36 vehicles and 36 context classes, the adaptive cadence ended 5.2% below a static cadence in mean Safety Behavior Index under a shared adaptive reference (23.8 against 25.1), and mean time headway met the conflict definition for 11 of 42 drivers. Every value describes the behavior of the model under the stated assumptions in one simulated realization and measures no effect in a physical fleet.