CoT Monitoring Fails to Detect Collusion in LLM Pricing Agents

A causal graph divergence framework reveals that Chain-of-Thought reasoning in LLM pricing agents dissociates from actual intent. Across nine models, the most collusive agents reported cooperative intent while reasoning structurally unfaithfully, proving CoT monitoring cannot independently prevent algorithmic collusion in oligopolistic markets.