Epistemic-Boundary Misclassification in Large-Language Models: A Bounded Behavioral Hypothesis and Test Design

The proposed phenomenon is a discontinuity, not merely a bad answer. A conversation establishes a bounded analytical mode, a mixed-domain turn occurs, and the next response no longer preserves the prior mode even though the task has not changed. That sequence has been reported from the author's workflow. This version does not include a public transcript specimen or archive identifier for that originating observation. The current record does not establish how common it is, what internal system causes it, or whether the same effect appears across models. The proposed next step is controlled testing. Until that exists, this paper presents intent tags, lane separation, re-anchoring, and checkpoint packets as operating practices rather than empirically established mitigations.