Which model density is best in pair natural orbital local correlation theory?

Low-scaling electron correlation theory based on the pair natural orbital approximation, PNO-CCSD(T), has become a powerful computational tool. Motivated by the recent discovery of large errors for organometallic molecules, we assess the role of the model density used to discard unimportant contributions. We find that second-order perturbation theory provides the best compromise between cost and accuracy, but coupling between localised occupied orbitals must be accounted for. Errors in the CCSD energy are then well below 1~kcal/mol, even for molecules with moderate multi-reference character, and the primary remaining source of errors lies in the treatment of the (T) energy contribution.