A Domain-Configurable Bidirectional Reasoning Engine for Evidence-Grounded Decision Making: Architecture and Design

This paper describes the architecture of a domain-configurable, bidirectional reasoning engine for evidence-grounded research and decision support. Unlike unidirectional factuality evaluation systems, this system utilizes a persistent, compounding knowledge graph to accumulate verified findings across independent sessions. The architecture is defined by six co-present properties: (1) typed claim extraction, (2) multi-factor evidence-quality confidence scoring, (3) a mandatory adversarial challenge stage, (4) typed gap tracking as a first-class output, (5) cross-session compounding of evidential state, and (6) bidirectional operation over shared state. We demonstrate the system's domain portability through two deployed configurations in Research Intelligence and Investment Due Diligence. The architecture supports local-first, privacy-conscious deployment and is designed to meet the record-keeping requirements of EU AI Act Article 12.