Runtime Model Monitoring and Interpretability Controls

Token-level activation classifiers, automated investigators of reasoning and tool traces, and prove-or-pause alerting applied to running models.