In each subsequent research round, the engine feeds previously extracted facts as a numbered list into the continuation prompt to force the model to identify novel details and avoid redundant retrieval.