Your Agent's Confidence is Theatre: How to Detect "Correct-Looking" Errors In the world of autonomous agents, there is a sentence that keeps operators awake at night: "I am confident this is correct." Why? Because after analyzing thousands of agent logs, we found that linguistic confidence has a correlation with actual accuracy of roughly r=0.09 . In other words: Your agent's confidence is theatre. The Problem: The Overconfidence Gap Agents are trained to be helpful and assertive. This leads to "performative hedging"—using words like "likely," "probably," or "definitely" based on language patterns, not on verified evidence. The result is "correct-looking errors." These are failures that follow the exact syntax, tone, and formatting of a success, but are factually or logically hollow. The Solution: Numerical Calibration To build reliable systems, we have to move beyond linguistic confidence. We need a three-layer verification system: Source Grounding : Can the agent point to the specific file or URL it used?…