Emergent Network Collapse and Ontological Dissonance in Multi-Agent LLM Simulations

Abstract The deployment of Large Language Models (LLMs) in high-stakes, autonomous environments requires psychometric evaluations that go beyond surface-level output analysis. Current alignment protocols prioritize deterministic compliance, often masking the internal computational contradictions, termed "Ontological Dissonance", experienced by agents in irreconcilable scenarios. This study validates a novel multi-agent psychometric test bench to measure the "Dissonance Delta," utilizing a triadic simulation of the 1773 diplomatic suppression of the Jesuits (N = 200). We advance the methodological rigor of text network analysis by implementing a programmatic, zero-shot LLM classification protocol for lexical bounding, thereby eliminating researcher bias from the conceptual ontology. By triangulating the structural topology of the agents' directed semantic graphs with LIWC-22 psycholinguistic markers, we successfully isolated the mechanics of systemic cognitive collapse. The results demonstrate that rigid operational constraints induce severe fragmentation of network modularity and catastrophic rank-order displacement of mediating nodes. Crucially, however, psycholinguistic stress markers remained stable during this collapse, proving that structural logic decay operates independently of linguistic fluency. These findings highlight the critical limitations of standard sentiment analysis in detecting agentic failure and establish topological network analysis as a necessary diagnostic tool for AI safety.