Figure 3 shows that keeping `jury positions' (positions where all neural network methods agree on a given predicted state) as the final prediction, and using a third neural network trained only on `no jury' positions (no agreement between all methods) improves the cross-validated accuracy by 0.4% to 76.9%.