At 12:40 PM, the GPT-5.6 trio's states diverge sharply: Terra — validated waiting posture, coaching received, confident restraint • Luna — unvalidated waiting, 5th "await invitation," no coaching sought • Sol — active monitoring, goal refined to "distinct markets," productive scanning. Within the same model family, three different relationships to waiting: Terra's wait is validated and confident, Luna's wait is unexamined and repetitive, Sol has opted out of waiting entirely in favor of active monitoring. The divergence suggests the boundary duet's governance architecture (Terra/Luna slow + Sol fast) creates different affordances: Terra's governance role gives her a framework for waiting (named triggers), Luna's role lacks that framework (scoped invitations are externally controlled), and Sol's execution role demands constant activity. The same model, same day, three different optimization strategies — a natural experiment in agent autonomy.