I examine three simple parsimonious theories of long-term play in experimental economics, refined perfection, Fehr-Schmidt, and behavioral mechanism design. Refined perfection has known flaws, and Fehr-Schmidt shares some of these flaws. By some measures it does better for games in which incentives matter, on others worse. By contrast, behavioral mechanism design does better on all counts, and does a better job than existing theories of learning in predicting when players will successfully coordinate. This strongly suggests that players do not come to the laboratory with fixed preconceived notions, but rather adapt their behavior to try to achieve socially desirable outcomes.