Auditability Limits in Reinforcement Learning via Discrete Behavioral Rules

This study evaluates whether RL policies can be represented through auditable discrete rules across six predicates, including trace integrity and behavioral agreement. Findings show that rule-set overlap does not guarantee behavioral agreement, as policies may share symbolic rules but diverge on fresh states.