The Fractal Graph Rewriting Framework: Dynamic Activation Geometry and a Runtime Thermodynamic Gate for Transformer Language Models

We extend the FGRF scale-dependent rigidity probe from static weight matrices (V3) to dynamic activation spaces, and convert the framework’s thermodynamic gate into a layer-specific runtime anomaly detector. A consolidated, registry-driven probe hooks all 24 attention output projections of Qwen2.5-0.5B, SmolLM2-360M, and GPT-2-medium, extracts the normalized squared-spectrum entropy Hσ and the two-peak decomposition of the complexity profile C(k) per layer, and applies a scaled Landauer gate to consecutive layer transitions.