Food-System Stress Testing: Prediction Tasks Under Alternative Outcome Definitions

Food-system stress tests can misdirect warning and policy interpretation when current-state screening, prospective entry and post-entry escalation are treated as one problem. This study develops a task-specific framework that fixes each prediction origin, risk set, event window and outcome definition. It provides a unique conservative demand-capped allocation and a sufficient condition for module-input perturbations to preserve event membership. The empirical analysis uses precomputed fields for 218 countries in four six-step scenarios without executing the full operator architecture. Within this constructed system, origin-time disruption, exposure and buffer measures achieved an area under the receiver operating characteristic curve (AUC) of 0.835 for prospective Level-1 entry, versus 0.762 for initial score margin. Post-entry evidence was sparse: 81 of 135 Level-2 arrivals were same-step jumps, while a fixed three-step window contained 26 events from 11 countries. Added origin-time fields showed no stable improvement over Level-2 cutoff proximity. Reallocating 0.05 among component weights changed eligible origins from 156 to 319 and events from 17 to 59; outcome definitions therefore changed the estimand before model fitting. The framework supports internal stress testing, but performance does not establish real-world predictive validity. Operational use requires independent outcomes, externally fixed event definitions, scenario variation and longer follow-up.