The real answer is that nobody trusts their automated evals enough to be confident that any given automatically-trained release actually improves performance, even if eval scores go up. So for now everyone batches up updates and vibe-checks them before rolling them out.