Same Prompt, Same Model Different Results Put 5 AI Harnesses to the Test