We ran HealthBench on our health AI's safety layer. It scored lower than the bare model.A 150-conversation HealthBench subset, a safety layer that costs points, one real dose leak we found while measuring, and what moved the score.6 min read·tabibu.hashnode.devOct 6
