The deployment of LLMs in high-stakes compliance workflows presents a fundamental tension: while generative models excel at synthesizing heterogeneous information into coherent narratives, their propensity for hallucination and weak grounding renders them unsuitable for audit-critical applications without substantial architectural constraints. This work proposes a framework that reframes AML triage as an evidence-constrained inference problem, explicitly separating the generative process from the decision rationale through a structured output contract.
The technical approach combines three complementary mechanisms. First, the system performs multi-source evidence retrieval that aggregates policy typologies, customer behavioral profiles, transaction alert triggers, and subgraph structures into a bounded context window. This retrieval stage functions as a form of in-context grounding, limiting the LLM's generative scope to information explicitly extracted from the knowledge base. Second, the output contract mandates explicit citation indices, forcing the model to maintain bidirectional links between claims and supporting evidence while explicitly flagging contradictions or gaps. This architectural choice transforms the triage rationale into a verifiable artifact rather than a post-hoc explanation. Third, counterfactual validation tests decision robustness by querying whether minimal, semantically plausible perturbations (e.g., adjusting transaction amounts or customer risk profiles) produce coherent changes in both recommendation and justification—a mechanism that operationalizes faithfulness as a measurable property.
Empirical evaluation on synthetic AML benchmarks demonstrates substantial improvements over RAG-only and unconstrained LLM baselines. The method achieves PR-AUC of 0.75 and escalation F1 of 0.62, while maintaining citation validity at 0.98 and counterfactual faithfulness at 0.76. These metrics collectively establish that the framework produces not merely accurate decisions, but decisions whose rationales are both auditable and internally consistent under perturbation.
The work's significance lies in demonstrating that governed LLM deployment need not sacrifice compliance rigor for decision support utility—a finding with implications beyond AML to any regulated domain requiring explainable, verifiable automation.
