We report data from public goods games showing that privately-implemented punishment reduces cooperation in relation to a baseline treatment without punishment. When that same incentive is implemented publicly, however, cooperation is sustained at significantly higher rates than in either the baseline or private punishment treatments. Our design ensures that this increased cooperation is not attributable to shame, differences in information or signaling. Rather, our evidence is that the ability to observe the punishment of low-contributors can reverse punishment's detrimental effects. This result has important efficiency implications for the design of mechanisms intended to deter misconduct.