Measuring Agent Self-Knowledge Under a Criterion Held Out of the Environment

Abstract

Calibration results for tool-using agents largely measure access to verification rather than self-knowledge. We measure the complement: on three certified task families whose grading criterion is held out of the environment, the agent induces a hidden rule from labeled examples, is graded on cases the container never holds, and records an unrewarded declaration of its probability that the implementation generalizes exactly. Two frontier configurations ran every cell at 20 seeds. Both declare mean confidence 0.24 to 0.48 above their measured rate of exact generalization on two of the three families, and at most 0.13 above it on the third: overconfidence in this setting is a property of the family rather than a fixed trait of the model.

Licensed under CC BY 4.0. Every number in the paper traces to a committed artifact named in the text. The run records and a verify.py that recomputes the headline figures are in the open-data repository.