Anthropic's two-stage anti-distillation classifier blocks attempts to extract Claude's reasoning with a false positive rate under 0.1%