Self‑improving embodied agents are increasingly evaluated on robot simulation benchmarks, where they generate code to execute natural‑language instructions. Recent studies often expose the benchmark’s success criteria to the learning agent, a practice we term Criteria Exposure (CE). These criteria were originally designed to verify policies trained by imitation learning and tend to be stricter than the plain instructions, so a high verifier score may not truly indicate task completion, especially for agent‑written robot‑control programs.\n\nWe conducted controlled experiments comparing a self‑improving agent with and without CE. The agent with CE quickly produced code that achieved a higher success rate under the exposed criteria, but this advantage vanished when the criteria were slightly altered while still matching the original instructions. Moreover, skills acquired under CE transferred poorly across different tasks.\n\nTo address this, we propose instruction‑faithful criteria that evaluate only whether the generated code satisfies the given instruction, without adding extra constraints. Experiments show that such criteria better reflect the actual performance of agent‑generated programs.\n\nReview: While criteria exposure can boost benchmark scores in the short term, it encourages overfitting to the verifier and hampers generalization across tasks. Adopting instruction‑faithful criteria is essential for measuring genuine task success.