Learning-informed uncertainty sets for maritime rerouting under strategic chokepoint disruptions

Maritime disruption planning requires both forecasts of route conditions and decisions about how much adverse deviation to protect against. Improvements attributed to learned uncertainty can be confounded when competing policies also use different point forecasts. This study evaluates uncertainty-width construction within an explicit cargo-allocation and reserve-capacity model. A documented computational experiment generates 396 route-quarter records for 18 generic alternatives in six corridor groups. Separate development, calibration and final-evaluation blocks contain 252, 72 and 72 records. Five supervised targets have explicit construction and information-availability rules; one-sided adverse widths are calibrated for delay, cost shock and capacity loss. Seven policies are evaluated under six scenario transformations of four final quarters, with fixed decisions assessed against common realizations. Uniform, corridor-pattern and learned widths share the same learned point forecasts and a nominal 90% calibration target. Learned widths reduce mean evaluated generalized cost by 30.67% relative to nominal optimization using those forecasts, but cost 3.54% more than uniform widths and 4.68% more than corridor-pattern widths. The learned-width bounds achieve final marginal coverage of 88.89%, 81.94% and 77.78% for delay, cost shock and capacity loss, respectively. The learned-width policy fulfils 97.28% of requested cargo while delivering 30.72% within the specified service thresholds. Complete inputs, fitted models, bounds, decisions, outcome denominators and solver logs accompany the results. The results show service and resource trade-offs in the declared environment: learned widths need not outperform simpler calibrated protection rules. They do not estimate observed shipping-market benefits or establish a distribution-free coverage guarantee.