The Elliptic dataset is widely used in Bitcoin anti-money laundering research, yet its original anonymized features have limited forensic interpretability. Much of the existing Elliptic-based literature relies on these opaque benchmark variables, leaving insufficient attention to semantically explicit and interpretable graph representations for illicit transaction detection. To address this gap, this article proposes a combined approach that integrates transaction-level feature reconstruction with interpretable forensic descriptor engineering. First, the benchmark’s original feature space is replaced with a semantically explicit reconstructed representation derived from public on-chain transaction data and metadata after resolving benchmark node identifiers to transaction hashes. Second, the proposed approach extends this reconstructed representation with interpretable forensic descriptors that capture local transaction abnormality, outgoing value redistribution behavior, and deviations from upstream transaction history. The empirical design isolates the contribution of the proposed descriptors by comparing the reconstructed representation against its descriptor-augmented variant. Across eight classifiers evaluated under a whole-snapshot train-test protocol that preserves within-snapshot graph structure, the descriptor-augmented representation consistently improves illicit class retrieval. CatBoost achieves the best results, increasing the area under the precision recall curve for the illicit transaction class from 85.10% to 90.27%, precision from 77.49% to 87.04%, recall from 75.57% to 81.11%, and F1-score from 76.44% to 83.90%. The article also discusses how the predictive component can be embedded into a hybrid analytical framework that separates machine learning classification from address-level forensic interpretation. This structure supports explainable prioritization and expert review while preserving the distinction between predictive evidence and forensic interpretation. Overall, the findings demonstrate that semantically explicit and forensically interpretable representations can substantially improve illicit transaction retrieval while supporting transparent post hoc analysis in Bitcoin anti-money laundering research.