Operational Interpretability: Intercepting Catastrophic Misalignment at Runtime
> What should we measure about an agent whose free-form reasoning we cannot trust and whose internal weights we cannot inspect? > > This post outlines Operational Interpretability: transforming internal deliberation from unconstrained text into typed, first-class runtime primitives, monitored in real time by an independent sidecar model to intercept...
Sep 240