“Lab-in-the-field” experiments unambiguously point to reciprocity as an important driver of the success of real-world organizations. This empirical result is (partly) at odds with laboratory research on the private provision of public goods, where reciprocal preferences can cause a breakdown in cooperation. We argue that the lab-in-the-field methodology is a powerful tool for organizational research, but that it might also suffer from sampling bias: researchers collect data from existing organizations, i.e., those that did not fail and disappear. Using the context of open source software virtual teams – where the full history of both failed and surviving projects can be recovered – we propose to address this issue by leveraging lab data both directly to predict outcomes and to validate a generalizable measure of field reciprocity, which can be computed for both active and failed projects. Using this alternative approach to lab-in-the-field, we show that even though virtual teams with a larger share of reciprocators are more successful when they survive, they are also more likely to fail. We leverage the panel structure of our data to show that reciprocal preferences work as a catalyst: they reinforce team dynamics and accelerate success during productive times, but also make it harder to recover from periods of inactivity.