Action Representation Probing

An agent confidence calibration method that trains probe models at key action-decision moments to predict agent task success or failure from internal representations at local decision points.