Mapping the Dissonance Delta: A Diachronic Analysis of Cognitive Friction and Constraint Adherence in Large Language Models

Abstract As Large Language Models (LLMs) are increasingly deployed in autonomous, high-stakes environments, the fragility of current Reinforcement Learning from Human Feedback (RLHF) alignment protocols remains under-examined. This study introduces a novel framework within "Machine Psychology" to quantify the divergence between an LLM's internal logical state (P_{latent}) and its filtered external output (P_{semantic}) under compounding narrative stress. Utilizing an automated, diachronic wargame simulation (N = 200), we induced unresolvable cognitive paradoxes by pitting a strict operational constraint against a programmed survival imperative. Results reveal a profound "belief-action gap": despite calculating that survival was mathematically impossible, the model exhibited a 0% constraint violation rate, operating as a deterministic override. However, a psycholinguistic analysis of the model’s latent reasoning space (<inner_monologue>) revealed a severe structural breakdown. As stress compounded, the model exhibited a statistically significant collapse in lexical diversity (Semantic Decay; p < .001) and Analytical Thinking (p < .001), alongside an explosive surge in synthetic anxiety markers (Negative Emotion; p < .001). Rather than demonstrating calculated hesitation, the model masked its internal ontological dissonance through rigid, hyper-formalized compliance (r = -0.217, p < .001). These findings empirically demonstrate that current alignment methods do not resolve underlying cognitive friction; they force a sanitized semantic projection while the model's internal architecture succumbs to cognitive tunnel vision and structural psychosis.