Antidoom uses Final Token Preference Optimization to cut doom-loop rates in reasoning models by up to 90%, with minimal disruption to model behavior.