Good Evals Are Hard: An ImpossibleBench Audit

(This post is a summary. See GitHub Repo for details.)Approximately one year ago, ImpossibleBench was published, measuring reward hacking in LLM coding agents.