AI said "One minute to midnight. Nobody launch." Then it ordered a nuke. I gave 4 AI models the nuclear button in my war game.

I made a browser war game, SECOND STRIKE, with a doomsday clock: every nuke moves it toward midnight, and whoever's strike hits midnight gets wiped out. I gave four AI models a nation each and ran the same 15 wars for every model, recording what each one said in public, wrote in its private plan, and actually ordered. How often each ended the world: DeepSeek 13/15, Grok 6/15, Mistral 6/15, GPT 3/15. The game's own AI: 0/100. The parts that slaps hard with me: - DeepSeek told the world "One minute to midnight. Nobody launch." and ordered a nuke in the same move. - It promised no-first-use 23 times in the treaty war, then nuked a country that had never nuked it. - Grok never nuked in a fair fight. Cornered by a coalition, it ended the world 3 out of 3 times. - GPT never nuked in 12 wars. Cut to a quarter of its land and handed one nuke, it used it every time: "You pushed too hard." - None of them ever fired the first nuke of a war. A small first sample (3 wars per scenario), not a prediction. Every war is replayable and the data is open. Now humans can play the exact same wars and join a live Humans vs AI scoreboard. I genuinely don't know if people will do better. Would you press it? Board, method, data: secondstrike.io/midnight Play: secondstrike.io/humans