<![CDATA[Tax compliance monitoring increasingly depends on timely detection of anomalous transactions, yet a prediction alone is insufficient when an administrative decision must also be justified against applicable rules. This study develops a compliance-aware anomaly detection architecture for an Integrated Tax Management Information System (ITMIS). It combines a heterogeneous ensemble of isolation forest, LSTM autoencoder, and gradient-boosted tree models with conformal uncertainty estimates, deontic-temporal policy monitoring, a constrained reinforcement-learning threshold governor, and cryptographically traceable decision records. The source manuscript reports an AUPRC of 0.847 on a synthetic deontic tax transaction stream, compared with 0.712 and 0.668 for two adaptive baselines. It also reports a demographic parity difference of 0.021, a gate rejection rate of 6.8%, and mean end-to-end latency of 47 ms. These figures are presented as reported study outcomes rather than independently reproduced measurements. The proposed contribution is to connect policy verification to the decision to accept or reject an adaptive threshold update, rather than reserving compliance checks for post hoc review. The framework offers a design for combining anomaly sensitivity, interpretable decisions, fairness monitoring, and tamper-evident auditing. Its operational and legal guarantees remain conditional on the correctness of encoded rules, monitoring semantics, and the underlying implementation.]]>
Copyrights © 2025
