In natural language interactions, a speaker and a listener cannot be assured to have the same beliefs, contexts, backgrounds or goals. This leads to difficulties and mistakes when a listener tries to interpret a speaker’s utterance. One principal source of trouble is the description constructed by the speaker to refer to an actual object in the world. The description can be imprecise, confused, ambiguous or overly specific; it might be interpreted under the wrong context. This paper explores the problem of resolving such reference failures in the context of the task of assembling a toy water pump. We are using actual protocols to drive the design of a program that plays the part of an apprentice who must interpret the instructions of an expert and carry them out. A primary means for the apprentice to repair such descriptions is by relaxing parts of the description.