This paper reports results of an experiment designed to investigate the nat
ure of cooperation and punishment. Subjects are matched in a series of two-
person, two-stage games with a sequential equilibrium that supports first-s
tage cooperation with a credible threat of subsequent punishment. Participa
nts sometimes used a consistent punish/reward strategy, and when they did,
cooperation rates increased dramatically. The results thus contradict "payo
ff relevance": second-stage behavior can be influenced by first-stage outco
mes that have no effect on the payoff structure. Nevertheless, high coopera
tion rates were often not observed, even with a Pareto undominated "punishm
ent" equilibrium in the second stage.