We present a model aimed at accounting for learning of predictive and causal relationships involving stimulus compounds, by means of a mechanism based on a normative-methodological analysis of causality that goes beyond the traditional associative/rule-based controversy. According to the model, causal learning is attained by computing the validity of each stimulus in a given learning situation. The situation is determined by the assumptions, objectives, and aims held by the learner or demanded by the learning context. Hence, validity computation depends on task demands: causal, predictive, or diagnostic according to a general principle of normative contextualization that allows learners to adapt a between-cues competition principle in a flexible way. Validity is computed using the Relevance Relativization mechanism, a linear model, based on the balance between the probability of stimulus combinations and the probability of each cue. Thus, cue interactions occur mainly when the combination of stimuli shows predictive changes in relation to the same cues considered individually. This model makes novel predictions concerning variations of the competition principle as a function of the type of procedure, including blocking, simultaneous blocking, and relative validity. In addition, our model also integrates top-down and bottom-up processing levels, including individuals' assumptions or previous beliefs.