Document Reasoning as State-Space Search: The State-Graph Document Reasoning Engine (DUE-SG), a Tree of Thoughts Architecture for Policy Document Analysis

This paper presents the State-Graph Document Reasoning Engine (DUE-SG), an architecture that formalises policy document analysis as state-space search over a directed graph of belief states. Nodes represent sets of facts held simultaneously as a current belief about the document; edges represent decisions such as adding a fact, resolving a conflict, or committing to one interpretation of an ambiguous passage.

Search over this graph, implemented via beam search with large language model (LLM) evaluation, produces an auditable reasoning path from document to conclusion. The architecture extends the Tree of Thoughts framework with graph-theoretic state representation, Bayesian confidence accumulation, and constraint-based evaluation.

DUE-SG is platform-independent and produces an explicit reasoning history suitable for high-stakes policy contexts. The primary application domain is Special Educational Needs and Disability (SEND) policy analysis. The work is positioned as complementary to DUE-P v2.0, with a unified architecture proposed as future work.