|Abstract:||Three experiments investigated whether graded contextual constraints produce graded on-line representations for contextually-defined categories. In Experiment 1, people read pairs of sentences. In each pair, the first sentence included a category term whose interpretation was constrained by contextual information to either a low, moderate, or high degree. The second, target sentence contained a category exemplar that matched the category better with each increase in constraint. The results showed that target sentence reading times were facilitated by high constraints only. In Experiment 2, when exemplars were presented as targets in a lexical decision task, facilitation increased with each increase in constraint. In Experiment 3, high constraints facilitated decision times of more appropriate exemplars, but inhibited decision times for less appropriate exemplars. Moderate constraints, weaker than those of Experiment 2, affected neither exemplar. Collectively, the results suggest the following about inferences for contextually-defined categories: (a) context must be fairly constraining for an inference to occur, (b) inferences contain information about a specific exemplar, not about a revised general category definition or a reduced scope of reference, (c) once instantiation by an exemplar occurs, its strength varies with the amount of contextual constraint.