Key components of the multiple constraint satisfaction framework are explored in a series of experiments set in complex and ambiguous domains. All cases show the prevalence and importance of a purposeful structuring of the information by the participants. The participants gradually generate coherence even without increasing information. In accordance with multiple constraint satisfaction predictions, the assessments of inferences increasingly spread apart. Also, the correlations between the dependent variable (the decision) and the independent variables, as well as between the independent variables, consistently grow stronger as the participants progress through the decision stages. The information structuring - a gradual simplification of the component structure— is captured as principal components associated with the various decision stages. Neural networks predict the judgments in the various decision stages relatively well. Finally, the role of the ongoing structuring of the underlying information was explored through the application of trained networks to data in other decision stages.