“The practice of designing Reinforcement Learning (RL) environments inspired by evaluation benchmarks contributes to the disconnect between high eval scores and poor real-world performance in AI models.”