Two word-by-word, self-paced reading experiments investigated the speed with which readers were sensitive to discourse-level anomalies. An account arguing for delayed sensitivity (Guzman & Klin, 2000) was contrasted with one allowing for rapid sensitivity (Myers & O'Brien, 1998). Anomalies related to spatial information (Experiment 1) and character-attribute information (Experiment 2) were examined. Both experiments found that readers displayed rapid sensitivity to the anomalous information. A reading time penalty was observed for the region of text containing the anomalous information. This finding is most compatible with an account of text processing whereby incoming words are rapidly evaluated with respect to prior context. They are not consistent with an account that argues for delayed integration. Results are discussed in light of their implications for competing models of text processing.