This project studies illicit Bitcoin transaction detection on the Elliptic dataset under a stable pre-shutdown split (train 1–32, val 33–37, test 38–42). We compared strong tabular baselines, GNNs, graph-aware non-neural models, compressed graph–tree hybrids, directed residual GNNs, and combination models. Generic GNNs improved over weaker graph baselines but remained below the best tabular model. A graph-aware ET stack using directed neighbor-risk aggregates reached 0.905 test PR-AUC, while compressed hybrid models showed that GNN embeddings help more when constrained through low-dimensional bottlenecks, including Matryoshka-style designs, before integration into trees. The best standalone graph models were directed residual GNNs (up to 0.916), and the top result, 0.9187, came from a preserved-head combination model integrating GraphAgg ET with SIGN/stack components. Overall, graph information helps most when integrated with tabular models rather than used through a standalone GNN.