A letter confusion experiment that used brief durations manipulated payoffs across the four stimulus letters, which were composed of line segments equal in length. The observers were required to report the features they perceived as well as to give a letter response. The early feature-sampling process is separated from the later letter-decision process in the substantive feature models, and predictions are thus obtained for the frequencies of feature report as well as letter report. Four substantive visual feature-processing models are developed and tested against one another and against three models of a more descriptive nature. The substantive models predict the decisional letter report phase much better than they do the feature-sampling phase, but the best overall 4 X 4 letter confusion matrix fits are obtained with one of the descriptive models, the similarity choice model. The present and other recent results suggest that the assumption that features are sampled in a stochastically independent manner may not be generally valid. The traditional high-threshold conceptualization of feature sampling is also falsified by the frequent reporting by observers of features not contained in the stimulus letter.