I imagine Anthropic would rather train a small control model instead of resorting to sampling hacks