Interpretability, runtime monitoring, model editing, and the verification of self-improving AI, on open models with every run replayable.