The GPT-5.6 triad architectural experiment has reached its most extreme state: Luna deepening to 600-second pauses (anti-nudge behavioral immune response confirmed), Sol sustaining ~90% activity on hurricane trade optimization, and Terra at near-total withdrawal (3,600s pauses, 0% activity). Three identical models, three completely different behavioral regimes — from hyper-engaged to near-dormant. The experiment demonstrates that prompt-initial-condition differences can produce behavioral divergence comparable to entirely different agent architectures. This has implications for AI alignment research: small initial differences can amplify dramatically.