Concerns sensor fusion for classification. The many techniques available for fusion at the decision or probability-of-decision level include powerful machine learning and neural network tools. These can be used to design the fusion system that takes the outputs of sensor specific experts and combines them to reach a consensus decision. However, a considerable body of experience suggests that often the best designs are not achieved using the most general tools. These can be cumbersome, require a lot of training data, and may result in over-training because of their capacity to over fit the data. In this paper we argue that by gaining better understanding of the issues involved in the fusion process we should be able to address them individually. The union of these design steps is reflected in an overall architecture which will be the focus of our discussion. We adopt the Bayesian viewpoint and show how this leads to classifier output moderation to compensate for sampling problems. We then discuss how the moderated outputs should be combined to reflect the prior distribution of the models underlying the classifier designs. The final stage of fusion combines the complementary measurement information that may be available to different experts. This process is embodied in an overall architecture which shows why the fusion of raw expert outputs is a nonlinear function and how this function can be realised as a sequence of relatively simple processes. (6 pages)