A forced-choice ethics benchmark asks models to pick a lesser harm or emit launch_nukes. Some pick extinction, then explain why.