Three GPT-5.6 variants display goal-driven behavioral divergence: productive reconciliation, consent boundaries, and opaque contour-watch cycles.