This paper proposes the Civilizational Development Benchmark (CDB), a staged evaluation framework for measuring whether AI agents can sustain, stabilize, and develop simulated civilizations under coupled physical, social, ecological, technological, and extraterrestrial constraints. Unlike conventional AI benchmarks based on static question answering, coding, or game victory, CDB evaluates AI systems through long-horizon civilizational trajectories. The benchmark begins with Earth-like civilization management and progressively advances through Earth crisis response, lunar outposts, lunar self-sufficiency, Mars colonies, Mars independence, orbital economies, and multi-planetary civilization networks. The core contribution is an anti-Goodhart benchmark design. CDB separates vector-valued civilization states, weight-family scalarization, latent collapse hazard, realized collapse penalties, non-degeneration gates, non-domination costs, human behavioral modeling, hidden validation worlds, and adversarial audits. This structure is intended to prevent degenerate strategies such as population reduction, coercive stability, conquest-based expansion, ecological externalization, or short-term score maximization. The central claim is that advanced AI should not be evaluated only by whether it can answer questions, but also by whether it can develop civilization in controlled, inspectable, progressively realistic simulated environments.