Architecture of Intelligent and Explainable Risk Analysis for Post-Release Customs Control

The relevance of the study is driven by the critical need for Russian customs administration to transition from automating individual operations to intelligent decision-support frameworks, in accordance with the Development Strategy until 2030. At the same time, the high cost of official errors persists, while requirements for the justification of inspections and targeted control continue to grow. A contradiction arises between the need to intellectualize post-release control based on artificial intelligence, on the one hand, and the high responsibility of inspectors, time constraints, the growing number of weak risk signals, and mistrust of opaque «black box» models, on the other. The aim of this work is to develop an architecture for an intelligent and explainable risk analysis system for post-release customs control, designed for safe pilot testing in a low-risk environment. The research method is based on designing a four-loop microservice architecture that integrates open-source intelligence extraction, deterministic artificial intelligence interpretation, context-aware action orchestration, and an intelligent document interface, as well as on economic-mathematical modeling of nonlinear risk scoring and minimization of the expected operational loss function. The paper introduces, for the first time, the concept of built-in explainability, in which contextual data components (source, confidence, status) are processed synchronously with numerical features, ensuring a 100 % reproducible traceable decision trail. An adaptive routing system based on an attributed multigraph is presented, which translates analytical output into parameterized response scenarios, along with a cognitive multi-agent architecture that simulates the work of an experienced inspector with a metacognitive verification mechanism. The effectiveness of a three-class decision-making logic is demonstrated, whereby an intermediate class of analytical uncertainty is processed automatically, reducing the proportion of unjustified in-depth inspections. The practical significance lies in providing a conceptual model for seamless integration into existing Federal Customs Service of Russia systems via secure APIs, as well as in regulating safe piloting within a closed loop based on quantized small language models. Experimental validation on a synthetic dataset confirmed the internal logical coherence of the architecture, the operability of end-to-end routing, and the mathematical guarantee of additivity of deterministic artificial intelligence interpretation explanations. The proposed approach makes it possible to transform standard risk scoring into procedurally justified semi-automatic control, ensuring the legal significance of artificial intelligence outputs. Beyond customs administration, the architecture can be applied to build trusted decision-support systems in other regulatory bodies requiring verifiable analytics and secure data handling.