The multi_statement_transaction_simple test in concurrency_replication does a setup using standard write concern. If the first transferMoney in this test happens to run before the setup becomes majority committed, it will do transfers that are no-ops (because it does updates with queries that find no documents), leading to a wrong count. So at least the last setup write must be done with write concern majority (might make sense to change to a bulk write)

Additionally the transferMoney code should check to make sure if the update updated exactly one document.