Contemporary language model architectures exhibit a well-documented failure mode: semantic state drift. A model may produce fluent, locally coherent output while gradually departing from its defined role, goals, or relational stance across the course of a long interaction. Existing diagnostic approaches address this problem at the output layer, asking whether the current response matches an expected persona — equivalent to examining a single photograph: it reveals the present state, but not the process by which it was reached. Observable Semantic Dynamics (OSD) proposes a different epistemological stance. Rather than asking what a model's output is, OSD asks how it came to be that way. This shift in question requires a corresponding shift in the unit of observation — from isolated outputs to semantic state trajectories. OSD is positioned as a Judge-agnostic observation framework: it does not depend on any specific model, and it observes how semantic states evolve regardless of which model is being monitored. The framework's core contribution is visibility, not prediction — shifting the research question from "whether the output changed" to "how the change formed," distinguishing it from Drift Detection, Intent Tracking, and Mechanistic Interpretability approaches. This work is the observation layer of the Recursive State Transition Architecture (RSTA) framework. The two frameworks share conceptual primitives — State, Trajectory, Transition — but solve different problems: RSTA actively preserves semantic state against premature collapse during generation; OSD measures, after the fact, whether that preservation held. Live tooling and documented behavioral observation cases are available at the project repository linked below.
MAO LIN CHANG (Fri,) studied this question.