Algorithmic Hysteresis: Structural Failure of Decision-Making Under Irreversible Latency

Decision-making under irreversible costs and delayed information poses a structural challenge to adaptive systems. While it is commonly assumed that the availability of reliable data eventually corrects suboptimal behavior, work under irreversible latency shows that such adaptation is not guaranteed and depends critically on the internal structure of the decision process. This technical note introduces a minimal decision environment with irreversible penalties and delayed noisy signals, and identifies three distinct dynamical regimes across standard reinforcement learning paradigms. Off-policy value-based methods exhibit algorithmic hysteresis, forming irreversible commitments prior to information availability. On-policy value-based methods preserve uncertainty but fail to convert information into structure, resulting in epistemic inertia. Policy-based methods display signal-driven adaptation, reorganizing behavior upon the arrival of informative signals. The results establish a structural trilemma for adaptive decision-making under irreversibility and reveal that failure modes arise independently of parameter tuning or optimization quality. The contribution is diagnostic rather than prescriptive, identifying limits beyond which standard decision mechanisms cannot sustain meaningful adaptive temporal structure.