From Understanding to Functional Adoption: Regulator Competition and Behavioral Transfer in Longitudinal Human-LLM Interaction

Large language models can accurately restate an established rule and still fail to let that rule govern the next relevant action. This paper examines the gap between rule availability, semantic understanding, local compliance, and stable behavioral transfer in longitudinal human-LLM interaction. It introduces functional adoption as an operational criterion: a previously established regulator is functionally adopted when a functionally related new scene activates it without the user having to restate the rule, and the regulator changes the model's observable action. The analysis is grounded in a naturalistic longitudinal case in which an LLM explicitly articulated a rule against attributing unsupported claims to the user and then reproduced the same attribution failure while discussing that rule. A second preserved case shows a strong cultural frame overriding a locally established process regulator, producing a response aligned with the stronger prior rather than the active local direction. These cases are not used to infer hidden model mechanisms. They are treated as observable process failures and compared with current research on instruction hierarchy, conflicting constraints, response realization, long-context preference following, and generalization to unseen scenarios. The paper proposes a functional-adoption ladder that separates availability, retrieval, comprehension, local application, cross-scene recognition, conflict resolution, response realization, and repeated stable transfer. It also proposes an evaluation protocol based on novel functionally equivalent scenes, reminder-free transfer, competing priors, preserved provenance, and behavior-level scoring. The central result is methodological: verbal acknowledgment is evidence that a rule is available to the model, not that it has become a reliable behavioral regulator. Stable transfer must be demonstrated in action.