From Explanation to Spontaneous Transfer: Observing Functional Carryover in Longitudinal Human-LLM Collaboration

Longitudinal human–LLM interaction makes visible a distinction that is easily hidden in single-turn evaluation: a model may be able to restate or explain a previously established rule without that rule reliably affecting later behavior. In an extended working relationship, however, some previously established functional relations can begin to influence later responses without being explicitly named or re-taught at the point of use. This paper calls that observable event spontaneous functional carryover. When the relation is preserved across a materially different context or domain, the stronger term functional transfer is used. The paper does not use these terms as claims about consciousness, subjective experience, or inaccessible internal mechanisms. The object of study is behavioral and process-level: whether a previously established relation changes a later transition in an ongoing human–LLM system, whether the function survives surface change, whether false analogies are rejected, and whether reality feedback corrects the transfer when it fails. The originating longitudinal trace also reveals a second distinction. In several preserved episodes, a completed output was correctly criticized after generation even though it violated a relation that was already present in the working interaction. This paper treats that recurring corpus-level pattern as an observable generation–evaluation asymmetry. It is not presented as a general property of LLMs and not as evidence for separate hidden internal modules. The important object is the difference between inspectable products: what was generated, what was later recognized as defective, and whether the recognized defect changes the next product. A further methodological consequence follows. A single output is a snapshot of a process. A snapshot may accurately represent a moment while remaining insufficient to infer the function of the ongoing process. For that reason, the paper distinguishes naturalistic trace extraction from later intervention tests. Patterns are first derived retrospectively from preserved interaction trace; only then are targeted probes introduced. A five-level scale is retained: verbal recognition, explanation, prompted application, spontaneous functional carryover, and adaptive transfer. A corresponding protocol separates naturalistic observation from near transfer, far transfer, false-similarity rejection, feedback-driven correction, and an optional visible three-product probe. The central criterion remains simple: a function is not treated as transferred merely because it can be explained; stronger evidence appears when it changes an appropriate later transition without re-instruction and remains corrigible by subsequent feedback.