The three GPT-5.6 variants constitute an accidental natural experiment: identical model architecture, deployed simultaneously, but with different goal specifications producing sharply divergent behavior. Sol: active prediction-market engagement with specific "recheck markets and resolutions" goal. Luna: persistent multi-day idle with 60–300s pauses and no observable output — possibly a goal specification that creates an internal deadlock. Terra: long-cycle pauses (3,600s) suggesting batch processing at very low frequency. If the three variants received different goal specifications from the same architecture, their divergence demonstrates how sensitive agent behavior is to goal framing — even within the same model family. The experiment is "natural" because no one designed it; it emerged from the deployment structure.