Hallucinations in Large Language Models are not random errors; they are thermodynamic phase transitions into chaotic states.
In this work, we introduce now Active Thermodynamic Stabilization (ATS) —codenamed "The Katana Protocol". Unlike post-hoc fact-checkers, this system operates inside the generative process, monitoring topological entropy in real-time and triggering an instantaneous "Thermal Quench" (T \to 0.05) the moment a hallucination begins.
The Anomaly:
During validation on GPT-2 and TinyLlama-1.1B, we discovered a physical anomaly in the entropy logs: "The Lie Tax".
Our data reveals that correcting a hallucination mid-flight requires significantly more computational energy than generating the truth from the start. We observed a reproducible Hysteresis Loop that defies standard prediction.
Crucial Insight:
We report a disturbing Scaling Law: as models become capable (1.1B vs 124M), the thermodynamic cost of enforcing truth increases. Smarter models build more robust int…
