Openclatura: Verification-Guided Agentic Construction of Molecular Nomenclature Rules

AI systems that construct scientific tools or generate scientific artifacts require verification mechanisms that are substantially more reliable than language-model self-evaluation. We introduce Openclatura, an open-source, deterministic system for generating IUPAC names from molecular structures, whose rule set is constructed through executable verification. Candidate nomenclature rules proposed during an agentic development workflow are tested by generating chemical names, parsing those names back into molecular structures with the independent OPSIN parser, and checking round-trip consistency against the original molecular graph. Verification failures therefore provide concrete counterexamples for revising the symbolic rule set rather than relying on model judgments alone. The resulting system is deterministic and inspectable: naming decisions can be traced to molecular-graph elements, nomenclature rules, and their verification outcomes. Across QM9, PubChem, and ZINC22, Openclatura produces structurally valid names for $100\%$, $99.0 \pm 0.0\%$, and $96.5 \pm 0.1\%$ of molecules, respectively, outperforming the neural STOUT baseline while requiring only a single CPU core. We further expose these structured traces through Robomoleculographer, which converts molecular graphs and naming decisions into natural-language descriptions. Openclatura demonstrates how independent, executable domain verifiers can be incorporated into agentic construction loops to produce transparent and reproducible scientific software.