The Null Manifold: A Physico-Geometric Falsification of Emergent Agency in Large

The "Scaling Hypothesis" posits that sufficient parameter and data scale in Large Language Models (LLMs) leads to the spontaneous emergence of general agency. This paper challenges this notion by proposing a tripartite physical falsification framework. We analyze the inference process not as cognition, but as a physical trajectory constrained by causality, topology, and thermodynamics. We demonstrate: (1) Causal Severance: Scaling observational data (N → ∞) only refines the correlation within a Markov Equivalence Class, failing to bridge the gap to interventional ( do(x)) causality due to fundamental identifiability limits. (2) Topological Triviality: We show that standard Transformers function as overdamped gradient flows, causing the first Betti number (β1) to vanish, thereby collapsing the neural manifold into a contractible state incapable of sustaining the re-entrant loops required for working memory. (3) Thermodynamic Stationarity: We prove a "Zero-Work Theorem," contrasting the "Passive Demon" nature of LLMs with biological autopoiesis. We conclude that without the introduction of inertial dynamics to restore phase-space volume, LLMs remain Semantic Simulacra—perfect isomorphic mirrors of syntax that remain ontologically distinct from sentient agents