This chapter explains how to design experiments that make AI claims credible. It begins with framing clear experimental questions, then shows how to choose evaluation tasks that actually test the core difficulty your work targets. It discusses selecting baselines and metrics that make comparisons meaningful and exposes trade-offs rather than hiding them. The chapter emphasizes careful analysis of stochastic results, including variability across runs, and warns about pitfalls such as overfitting to benchmarks or rewriting the story after seeing outcomes. Evaluation is presented as an iterative cycle that also generates new ideas.