The present study aims to test the hypothesis in McCarthy (1998) and Hayes (2004) that output-to-output faithfulness constraints are highly ranked a priori which leads children to initially favor non-alternating forms but to eventually produce alternating forms as adults do. Using spontaneous speech data of six mother-child dyads (Demuth et al. 2006), I examined how word-medial /t/ and /d/ in the flapping environment were realized in suffixed words such as eating (base: eat) and hiding (base: hide). The results show that children indeed produced suffixed words with variants that are faithful to the corresponding segment in the base (e.g., ea[t]ing and hi[d]ing), especially in earlier stages of learning. Such characteristic speech of children cannot be fully explained by the influence of the input data that they receive from their caregivers. Thus, I conclude that children’s production of variants that are phonologically similar to the base is due to highly ranked output-to-output faithfulness constraints at the initial state of learning. Based on a set of OT constraints, the children’s grammar is modeled with maximum entropy grammar (Goldwater and Johnson 2003).