To get final confusion matrices, we have to sum up TNs and FNs from both layers and take TPs and FPs only from the second layer.
2020 November 25, Pavlo Tertychnyi, Ivan Slobozhan, Madis Ollikainen, Marlon Dumas, “Scalable and Imbalance-Resistant Machine Learning Models for Anti-money Laundering: A Two-Layered Approach”, in Benjamin Clapham, Jascha-Alexander Koch, editors, Enterprise Applications, Markets and Services in the Finance Industry: 10th International Workshop, FinanceCom 2020 (Lecture Notes in Business Information Processing), Springer Nature, page 54