Isolation level (IL) acts as a correctness contract between applications and database management systems (DBMSs). Problematic IL implementations would cause incorrect transaction execution results and erroneous database states. However, existing testing methods could not efficiently generate various effective workloads for testing ILs. The core challenges come from the requirements of (a) black-box testing (trigger the IL code of a closed source DBMS), (b) effective testing (evade redundant and ineffective testing), and (c) anomaly-sensitive testing (test various ILs in a distinguishable way). For black-box testing, we investigate the IL implementations of 15 popular DBMSs and discover that they follow a generic framework that utilizes conflict graphs to manage all conflicts of a workload, and performs a verification policy to prevent non-serializable anomalies. For effective testing, we propose a lightweight data state mirroring approach, which helps generate SQL operations that precisely access its expected records and participate the formation of specific conflict graphs. We also propose an efficient history-independent approach to continuously generate dissimilar conflict graphs. It guarantees the graph generation overhead is irrelevant to the scale of historical graphs. For anomaly-sensitive testing, we propose an implantation-based approach to orchestrate conflict record accesses and inject them into different transactions according to the anomaly definition. Our approach outperforms existing approaches in testing effectiveness, efficiency, and coverage. Practically, we have successfully found 30 bugs in popular DBMSs.