Insider Threat Detection and Evidence-Grounded Triage Using a Heterogeneous Dual-Branch Network and a Constrained LLM Agent

Insider threats are difficult to detect and investigate because risky behavior can occur through valid accounts, develop slowly, and leave mixed behavioral traces. We study an analyst support architecture that links anomaly detection with evidence-based triage. Rather than treating insider threat analysis as a single classification task, the architecture organizes it as a workflow that combines behavior modeling, evidence retrieval, rule interpretation, and constrained reasoning. For anomaly detection, we build a multi-view representation over subject time windows and design a heterogeneous dual-branch network (HDBN) that fuses semantic and statistical branches through confidence-gated fusion. The detector captures temporal event patterns and statistical deviations while retaining structured alert clues for downstream investigation. For investigation, we construct a dual-track evidence infrastructure for LLM-assisted reasoning: a star schema data mart supports structured fact verification and a rule tree index supports policy and rule interpretation. An intent-aware hybrid retrieval mechanism routes sub-questions to the relevant evidence source and aligns structured facts with unstructured rules. We also design a constrained reasoning agent based on a controlled reasoning and acting workflow. Action-space constraints, evidence-completeness gating, and state recovery are used to limit reasoning drift and unsupported decisions when enterprise evidence is incomplete. On the CERT r4.2 dataset, the HDBN reaches 96.8% recall at the selected operating point. In the constructed investigation tasks, the constrained ReAct agent reaches 91.8% end-to-end triage accuracy compared to an unconstrained LLM agent baseline under the same tool access setting. The results support the feasibility of traceable detection-to-investigation workflows in the evaluated benchmark setting, while deployment on real enterprise data and adversarial scenarios remains to be validated.