The successes and failures of rapid language understanding provide valuable clues to how we mentally encode sentence representations, and how we navigate those representations in short-term memory. The real-time implementation of grammatical constraints has proven to be an effective tool for understanding these mechanisms, especially due to the profile of "selective fallibility". Some grammatical constraints are faithfully and reliably implemented in real time, others are not, giving rise to “linguistic illusions". This uneven profile provides important insights into the nature of the structured representations and the nature of the access mechanisms, and new discoveries are leading us to revise our views on these topics. In this talk I will summarize a number of recent lines of work that are overturning our views on why we are very good in some areas of language and rather poor in others.