EOS 1.0-3.0: Experimental Validation of Coherence Phase Transitions in Large Language Models

This paper provides experimental validation of the EOS (Emergent Order Systems) framework through controlled perturbation studies on a GPT-2 class model. We demonstrate reproducible phase transitions from ordered to disordered representational regimes, with coherence functioning as an operational order parameter. Results show smooth decay in low-energy regimes and catastrophic collapse at a critical threshold, consistent with second-order phase transitions in statistical physics. This work validates the theoretical predictions made in the EOS framework papers (DOI: [your earlier Zenodo DOIs]). Keywords: Large Language Models, Phase Transitions, Representation Stability, Coherence, Critical Phenomena, Machine Learning