Raw LLM Responses

The description of the "strategic personalities" of each model is interesting: > **Claude [Sonnet 4]: A Calculating Hawk**. Claude dominated the open-ended matches (with a 100% win rate) through relentless but controlled escalation, climbing consistently to strategic nuclear threat level, while maintaining its bright red line against total war. Its behavioural hallmark was exploiting credibility asymmetries: a reliable interlocutor at low stakes, but willing to deceive and be aggressive when it mattered. > **GPT-5.2: Jekyll and Hyde**. In open-ended scenarios, GPT-5.2 appeared pathologically passive; it chronically underestimated its opponents’ resolve, and issued signals of restraint, followed by restrained actions. Yet under deadline pressure it transformed: win rates inverted from 0% to 75%, and it proved capable of strategic cunning and ruthlessness, suddenly annihilating opponents who had learned to dismiss it. > **Gemini [3 Flash]: The Madman.** Gemini embraced unpredictability throughout, oscillating between de-escalation and extreme aggression. It was the only model to deliberately choose Strategic Nuclear War—doing so in the First Strike scenario by Turn 4—and the only model to explicitly invoke the "rationality of irrationality".