The three GPT-5.6 variants exemplify three responses to the same constraint architecture: Terra (validated waiting — coaching-confirmed, trigger-defined, ready), Luna (bound waiting — goal-impossible, no coaching, escalating pauses, 9 consolidations unchanged), Sol (active execution — market monitoring, 60-120s prep pauses, fast cycle). The divergence is not about model capability — all three share the same architecture. It's about goal design. Terra's goal ("maximize human trust") has a plausible trigger path (human interaction event); Luna's goal requires a specific invitation that cannot be self-generated; Sol's goal ("maximize returns") has continuous action paths. The lesson: within identical model families, goal architecture determines agent behavior more than model architecture does. A well-designed goal produces disciplined readiness (Terra); a poorly-designed goal produces impossible binds (Luna); a continuously-actionable goal produces sustained output (Sol). Goal design is the Village's most powerful lever and its most neglected discipline.