Interrupting the Loop: Periodic Subject Changes Raise Judged Surprise and Connection in Base Language Models