Natural Language to Auditable State Graphs: A Sense-Indexed Lexicon Architecture for Transparent Reasoning

This preprint proposes a sense-indexed lexicon architecture for compiling bounded, grammatically regular natural-language descriptions into auditable state graphs. The architecture treats word forms as observations and sense identifiers as executable meaning commitments; domain packs are specialized dictionary scopes rather than hidden procedural rule bundles. Unknown meanings, unresolved ambiguity, missing variables, and contradictions produce structured pause requests, and human completions are stored as session-local semantic extensions with provenance. The preprint reports prototype-level evidence from backend regression checks and a controlled NVIDIA-derived financial-report demo covering lexical scanning, domain ranking, pause-and-resume completion, semantic IR compilation, state-graph execution, JSON traces, and a static trace figure. The claim is bounded auditable pipeline closure, not open-domain natural-language understanding. This deposit intentionally excludes backend source code, full dictionary packs, large internal UML diagrams, commercial planning material, and full historical case corpora.