Evaluates AI coding agents as autonomous researchers that improve state-centric world models under fixed compute budgets - View it on GitHub