Action Substitution Versus Fallback Shields in Constrained POMDPs Under Distribution Shift

Shields check the actions proposed by learning or planning agents against a safety speciffcation before execution. A fallback shield replaces an unsafe action with a ffxed default, whereas a projection shield selects the nearest admissible action. This study shows that these operators are not interchangeable when the default action is itself inadmissible. It formalizes an admissible action envelope synthesized from a probabilistic temporal-logic speciffcation and proves that a null fallback cannot restore admissibility when the speciffcation requires a positive action. The operators are evaluated in a constrained partially observable Markov decision process (POMDP) model of a three-echelon supply network under distribution shift, using 800 seed-paired runs. Under a ffxed robust policy, projection achieved a service level of 0.670, compared with 0.059 for the fallback (\( E^{+}=1.62\times10^{6} \)). Under a tabular Q-learning policy trained online with shielding, the values were 0.649 and 0.063 (\( E^{+}=9.35\times10^{5} \)). The study also examines warm-up contamination in an adaptive conformal shift detector. Excluding early forecasting scores reduced calibration error in the predicted direction, but the effect did not reach the familywise threshold and explained only part of the observed under-ffagging.