This paper presents an experimental validation of the Miulus Law, an information-theoretic constraint describing the stability of self-referential systems under noise. Using controlled multi-agent simulations, the study empirically demonstrates how epistemic fitness—defined as the ratio of verifiable signal to informational noise, scaled by communicative reach—governs coherence, convergence, and collapse dynamics in distributed belief systems. A series of experiments explores key failure modes in AI-mediated information environments, including recursive amplification, shock-induced destabilization, bounded growth, coalition resistance, and tipping-point behavior. The results show that systems can appear locally stable while undergoing global epistemic degradation, and that increased reach can both stabilize and catastrophically destabilize systems depending on noise coupling. Rather than treating hallucination, polarization, or misinformation as model-specific defects, the paper reframes them as structural consequences of operating below a critical epistemic fitness threshold. The findings generalize across artificial agents, human collectives, and hybrid systems, suggesting that epistemic collapse is a predictable phase transition rather than an anomaly. This work contributes to emerging discussions on AI safety, information warfare, epistemic security, and the governance of large-scale cognitive systems. All experiments are reproducible, and the framework is intentionally architecture-agnostic, applying equally to human, artificial, and socio-technical systems.
Lauri V.N. Korpela (Wed,) studied this question.