it is manual, just automated via running inferences on a dataset of assumed refusals, and is a small rank-1 weight edit that doesn't require fine-tuning or training. Perf hit is much smaller than baking in these refusals in the first place.