Safety benchmark scores reflect training policy more than model capability

A study with two labeling policies on the same prompts shows that guardrail performance is dominated by which definition of 'unsafe' was used in training.