Byzantine fault detection (BFD) techniques are promising approaches for better scalability and practicality, even though they cannot mask Byzantine faults like Byzantine fault tolerance (BFT) techniques. However, the existing BFD protocol for database systems has suffered from long latency since it synchronously makes an agreement on the order of transactions between replicas when executing the transactions. In this paper, we explore an alternative BFD approach for database systems, which defers the expensive agreement and detects Byzantine faults lazily rather than detecting them in real-time. We discuss the challenges and design overview of our approach. We also present preliminary experimental results showing the benefit of our approach.