In this paper, we propose a methodology for collecting naturally occurring data on evidentials and epistemic modals. We use Desano (Eastern Tukanoan) as a case study. This language has a complex evidential system with six evidential forms. The methodology in question consists of having Desano speakers to play a logic game, "Mastermind". In this game one player (the codemaker) places colored pegs behind a screen and the other player (the codebreaker) tries to guess the code, receiving partial feedback from the codemaker through clues after each intermediate guess. In order to offset the unnaturalness of the codemaker's exclusive knowledge of the actual code, we adapt the task to have two codebreakers playing the game jointly and discussing what they know, what the code could/must be, etc. We found that there are several benefits to this method. It provides naturalistic dialogue between multiple speakers, rather than just monologue; utterances naturally vary as to whether speakers in the scenario have access to and interest in what kind of information source the speaker has or simply the conclusion they draw from this information; finally, an important point is that speakers find the task enjoyable. We hope that this study can add the body of literature on methods for collecting naturalistic speech for language documentation and description.