Upload PiroloThermodynamic_Stabilization-2025.pdf

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 internal narratives, requiring greater energy to override.

With our multi-model experiments, we have established that the system is absolutely scalable to new, complex LLMs.

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+PiroloThermodynamic_Stabilization-2025.pdf filter=lfs diff=lfs merge=lfs -text

PiroloThermodynamic_Stabilization-2025.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c809838806cc4b1164699fd8203b169ce7c32f6a1831217cb26e90a091c84554
+size 3857362