Does stated commitment to democracy anticipate citizens' willingness to defend it? We leverage a unique opportunity to answer this question. In 2022, we conducted a two-wave panel survey in Peru, measuring explicit support for democracy through survey questions and implicit support through votes for undemocratic hypothetical candidates in a conjoint experiment. Later that year, former president Pedro Castillo unexpectedly (and unsuccessfully) attempted to shut down Congress, and we conducted a third wave among the same panelists. Our analysis of responses to Castillo's move reveals that Peruvians do not unanimously condemn Castillo’s self-coup, and, unsurprisingly, Castillo voters were more indulgent of it. Castillo voters further adjusted their democratic standards to accommodate Castillo’s move. Using differences-in-differences, we also show that Castillo voters defected from democracy after the self-coup. Then, using machine learning techniques and regressions, we find that explicit survey responses about democratic values are the strongest predictors of condemning Castillo's undemocratic attempt, similarly predictive than having voted for Castillo. The implicit measure from our conjoint also predicts such condemnation, but very weakly. Harnessing this rare setup, we contribute to explaining and anticipating citizens’ commitment to democracy when the time comes.