aog comments on ImpossibleBench: Measuring Reward Hacking in LLM Coding Agents