A major challenge in second-language pedagogy and research is that of determining linguistic competence. Spontaneous oral production gives some indication of the state of a learner’s interlanguage, but the presence of non-target-like forms in such production confounds the analysis since the teacher or researcher cannot be certain whether such forms are random or systematic. Corrective feedback (CF) in oral production, usually in the form of recasts or elicitation, can thus appear arbitrary and inconsistent. This thesis investigates the effectiveness of delayed CF, in which representative samples of learners’ non-target-like production are systematically collected and tracked. The investigation employed three methods: first, accuracy and fluency in production were measured by means of a test in which learners reformulated their own non-target-like production and that of peers; second, accuracy and reaction time were measured as learners judged the well-formedness of those same reformulations; third, the developing complexity of learner production is monitored by means of an ‘error corpus’. Results indicate that delayed CF of this kind is effective in pushing learners towards greater complexity and accuracy in both production and recognition, and constitutes an approach to the problem of determining what the individual learner knows that has theoretical validity and pedagogical relevance.