Dyslexia is a deficit of the identification of words, which is thought to be a consequence of different possible cognitive impairments. Recent data suggest that one of these might be a specific deficit of the visual attention span (VAS). We are developing a remediation software for dyslexic children that focuses on the VAS and its training. A central component of this software is the estimation of the performance of a given participant for all possible exercises. We describe a preliminary probabilistic model of participant performance, based on Bayesian modeling and inference. We mathematically define the model, making explicit underlying generalization hypotheses. The model yields a computation of the most probable predicted performance space, and, as a direct extension, an exercise selection strategy.