Evidence Intelligence: A Computational Paradigm for Structured Reasoning over Unstructured Data

Large language models (LLMs) have demonstrated remarkable capa- bilities in natural language understanding and generation, yet their ability to support reliable reasoning over complex information remains limited. Current architectures, including retrieval-augmented generation (RAG), operate primarily over unstructured text passages, which constrains their ability to systematically aggregate evidence, detect contradictions, and maintain entity-level consistency across sources. In this work, we introduce Evidence Intelligence, a complementary paradigm that augments retrieval-based systems with an explicit repre- sentation layer for structured evidence. In this paradigm, information extracted from text is organized into evidence units linking entities, pred- icates, values, and source provenance, forming an evolving evidence state that supports aggregation, contradiction detection, and confidence-aware reasoning. Rather than replacing text retrieval, this approach introduces a structural reasoning layer that operates over the outputs of retrieval systems. We validate the proposed framework through claim-layer, a reference implementation applied to a literary corpus of approximately 400 pages (Don Quixote, Part I). The system extracts 7,726 structured facts across 1,402 entities, identifying 410 contradiction groups and enabling entity- level evidence aggregation with full paragraph-level traceability. A manual evaluation on a stratified sample of 175 facts yields a weighted precision of 70.3%, with error analysis showing that most inaccuracies arise from structural representation issues (entity normalization and predicate speci- ficity) rather than unsupported claims. Comparative experiments against a standard RAG baseline demon- strate that retrieval alone is effective for locating relevant passages but insufficient for tasks requiring systematic evidence aggregation and con- tradiction awareness. By contrast, the evidence-based representation en- ables explicit detection of conflicting claims and entity-complete evidence retrieval while remaining compatible with hybrid architectures that com- bine narrative context with structured evidence. These results suggest that Evidence Intelligence systems constitute a complementary class of evidence-native reasoning systems, extending retrieval-based architectures with explicit evidence representations that support traceable, aggregation-aware reasoning over unstructured infor- mation.