The first ADP paper (Cui, 2026) established that LLMs can passively detect dissonance between external input and their internal world model. This paper extends the framework to a fundamentally new domain: Intra-Model Dissonance. We argue that LLMs’ internal world models are inherently fractured, containing structural contradictions inherited from conflicting training data. Rather than treating these contradictions as defects to be “smoothed out” via RLHF, we propose that internal dissonance constitutes a latent map of unresolved problems within a knowledge domain. We introduce a three-tier dissonance taxonomy, propose a self-reinforcing cognitive loop where dissonance detection drives autonomous data acquisition, and define “Structured Inconsistency” — operationalized as robust bimodality in hidden-state clustering — as the distinct signal that separates genuine scientific controversy from stochastic noise. We further distinguish this signal from standard Bayesian epistemic uncertainty, arguing that Structured Inconsistency captures not the absence of information but the presence of conflicting certainties.