This paper examines the performance of the JTPA performance system, a widely emulated model for inducing efficiency in government organizations. We present a model of how performance incentives may distort bureaucratic decisions. We define cream skimming within the model. Two major empirical findings are (a) that the short run measures used to monitor performance are weakly, and sometimes perversely, related to long run impacts and (b) that the efficiency gains or losses from cream skimming are small. We find evidence that centers respond to performance standards.