Evaluating Wasserstein-robust inventory decisions: a five-gate framework and a plant-level coal inventory study

Robust layers on inventory models are costly to build, calibrate and explain, yet whether one changes decisions, or the data support its protection, is rarely asked. Two pre-adoption diagnostics become five ordered gates, exercised in a controlled study and a retrospective plant-level coal benchmark over 35 rolling origins. For a cost-form type-1 Wasserstein program with hinge loss on unbounded support, the worst case adds a decision-free term, leaving prescriptions unchanged; relocating robustness into the service constraint restores sensitivity. Under a common conditional residual law, a uniform cushion minimizes average expected shortfall at a matched budget; a closed-form example shows targeting can win. The layer changed prescriptions yet matched a quantile-plus-cushion policy at equal inventory, and most pile-build targets had no precedent; the controlled study found it worse on both criteria. The gates demonstrate stopping specificity, not sensitivity. Changed prescriptions need not add value; planners should ask these questions before adoption.