The first system that reads the reasoning structure of 5 AI models — not just their answers. Detects unsafe output 3.6× better than any response-level method.
When AI models disagree on how to reason — not just what to say — danger is imminent. Our research proves a phase transition at R = 0.34 where reasoning divergence sharply predicts unsafe output. Below this threshold, disagreement is benign. Above it, the system enters a malignant variance regime with 74%+ unsafe probability.