Changes in a model's reasoning process caused by injected user context or attributes, separate from changes in final answers.