We have used a computable model of human spatial vision to make predictions for phase-discrimination experiments. This model is being developed to deal with a broad range of problems in vision and was not specifically formulated to deal with phase discrimination. In the model, cross correlation of the stimuli with an array of sensors produces feature vectors that are operated on by a position-uncertain ideal observer to simulate detection and discrimination experiments. In this report the stimuli are compound sinusoidal gratings composed of a fundamental and a higher-frequency component added in various phases. We compare model predictions with three key results from the literature: the effect of the contrast of the fundamental on phase discrimination, threshold phase difference as a function of the fundamental frequency, and the contrast required for phase discrimination as a function of the frequency ratio of the two grating components. In the first two cases, the predictions capture the main features of the data, although quantitative discrepancies remain. In the third case, the model fails, and this failure suggests additional restrictions on the combination of information across sensors.