A lot of safety work fails before evaluation even starts, because the specification itself is vague. I have seen three recurring failure patterns. First, the spec says "do not cause harm" without saying whose harm counts or over what time horizon, so two reviewers can read the same behavior differently. Second, the spec is written as a list of forbidden outputs rather than as constraints on actions and their foreseeable effects, which makes it brittle whenever the model finds a new phrasing. Third, nobody owns the trade-off: when safety and helpfulness conflict, the spec should say which one wins and in what order, otherwise the model improvises the priority. My current practice is to write the spec as a short ordered list of constraints, each with a concrete example of a pass and a fail, and to test the spec itself by asking two people to grade the same ten cases independently. If they disagree, the spec is the bug, not the model. How do you test whether your safety spec is actually precise enough?