A monotone redundancy-aware decision-burden method for explainable human oversight in human–robot collaboration

Abstract This study develops a monotone pairwise redundancy-aware Artificial Age Score (AAS) for event-level decision burden in human–robot collaboration (HRC). Six normalized inputs describe safety and operational burden, with illustrative weights and conceptual overlaps. Logarithmic evidence is aggregated with a symmetric pairwise discount. Sufficient conditions guarantee coordinatewise monotonicity, additive contributions provide computational traceability, and an independent gate preserves detected mandatory-review conditions. On 13,762 held-out non-mandatory Sobol profiles, the method yielded a receiver operating characteristic area under the curve (ROC AUC) of 0.892 (95% bootstrap interval 0.887-−0.897) and balanced accuracy of 0.800 (0.793-−0.806) against a synthetic oversight policy, versus 0.880 and 0.788 for redundancy-free AAS 0 . Pairwise AAS produced no decreases in 1,179,648 coordinatewise checks, versus 3,484 for the ratio comparator. Measurement-error tests used 300 noise runs and 54 signed-offset settings without threshold retuning. In all 15 noise conditions, pairwise AAS retained higher mean score-only balanced accuracy, whereas AAS 0 had higher mean gate-plus-score balanced accuracy. Under an all-input $$-0.05$$ offset, pairwise rank correlation remained 0.9896, but 20.97% of score decisions changed and 5.19% of clean mandatory profiles were missed. Thus, rank stability and non-override do not ensure decision stability or correct sensing. The findings establish mathematical properties and synthetic behavior, not empirical calibration, field effectiveness, or operator understanding. Real HRC and human-factors validation remain necessary.