The Semantic Subset Principle (SSP) proposed by Crain et al. (1994) concerns what the learner needs to do when UG allows more than one reading for a sentence in the input, and one reading entails the other. It is intended to avoid problems when the learner's language includes only the "strong" (entailing) reading but not the "weak" (entailed) reading; learners initially adopt a setting for a parameter that allows only the strong reading, and may later change it upon hearing direct evidence that both readings are allowed. Yet the motivation for this learning strategy has recently been challenged (Musolino 2006, Gualmini & Schwarz 2009). So far, the empirical evidence and counterevidence for the SSP has come from children's truth-value judgments for VP-associated only or quantifier scope. In this study, however, we investigate children's judgment of a more subtle semantic contrast: the presuppositions of repetitive and restitutive again in English goal-PPs (e.g., walk to the village again). This is a case where the SSP can apply, yet I argue that children do not always employ it. They succeed even when the SSP predicts failure. I further propose that children use more general evidence about the syntax of English (e.g., verb-particle combinations), together with knowledge of the basic meaning of again, to derive the restitutive again with goal-PPs.