Who Holds the Function? Bidirectional Function-Allocation Failures in Longitudinal Human–LLM Interaction

Repeated failures observed in a longitudinal human–LLM collaboration appeared at first to belong to different categories: a tool was not called when needed; a tool was called when it was not authorized; an unresolved scene was acted on too early; a resolved scene was unnecessarily reopened; a language model shifted into an unsolicited caretaker role; and externally packaged authority was given more weight than an already verified local process trace. This paper asks whether these events can be usefully viewed as manifestations of a common process problem: the wrong function being activated, inhibited, deferred, or carried by the wrong system component. We relate the cases to established work on human–machine function allocation, misuse and disuse of automation, omission and commission errors, three-way decisions, trust calibration, algorithm aversion, sycophancy, and algorithmic paternalism. We then propose Functional Position Misassignment (FPM) as an operational descriptor rather than a claim about hidden model mechanisms. FPM occurs when a function is performed by the wrong carrier, at the wrong time, in the wrong direction, or with the wrong measure. The central methodological contribution is a bidirectional diagnostic rule: every identified failure should be paired with a search for its opposite failure before a corrective rule is stabilized. We conclude with a compact process-control procedure - the Functional Position Check - that separates function, carrier, state, action gate, boundary, implementation trace, and reality feedback. The aim is not to eliminate uncertainty or autonomy from human–LLM systems, but to keep functions in the right place and make role drift observable before it becomes a stable interaction pattern.