In this paper we analyze negative reciprocity theoretically as well as experimentally. Although the reciprocity model of Dufwenberg and Kirchsteiger (2004) is often cited in the literature to justify why people punish unkind behavior, we show that this model is not able to predict punishments in ultimatum games. We therefore propose several model modifications: Amongst others, we suggest “gradual reciprocation”, meaning that subjects want to reciprocate the level of the others’ kindness. Moreover, we assume that perceived (un)kindness of the proposer towards the responder depends on the belief of the former regarding the punishing behavior of other responders. Accordingly, we measure this belief and find that punishments are significantly overestimated, implying fairer offers. While our model modifications lead to a correct prediction of behavior in the (convex) ultimatum game, a within-subject comparison with behavior in a trust game reveals that decisions in the second-mover roles in either game are not correlated. Robustly, we find no correlation, although our novel, graphical way of eliciting decisions from a continuous strategy space allows subjects to precisely express their preferences. Despite this result, we find a small, but significantly positive correlation between reciprocity parameters in both roles. This is possible because behavior is not only determined by a subject’s reciprocal inclination, but by her individual beliefs as well.