In the answer-until-correct mode of multiple-choice testing, respondents are directed to continue choosing among the alternatives to each item until they find the correct response. There is no consensus as to how to convert the resulting pattern of responses into a measure because of two conflicting models of item response behavior. The first suggests that partial knowledge allows the subject to eliminate some distractors immediately, and then assumes essentially random guessing among the remainder. The second proposes that the first error made by the subject results from misinformation, but that guessing comes into play after that. The paper considers three latent trait measurement models from each of these perspectives. Each is an extension of the Rasch one-parameter logistic model. The first, which is most relevant to the partial knowledge viewpoint, is based on a count of the error choices before the correct response is identified. The second calibrates the difficulty of each step in each item. The third calibrates the difficulty of each distractor. It is argued that the second model provides the best context for distinguishing between the misinformation and partial knowledge approaches. (Author/PN)