Graph-Structured Generative AI for Regulatory Compliance in Heritage Planning: A Self-Supervised Evaluation over Japan's Statutory Rules

Objective: In urban renewal and heritage planning, navigating layered cultural-property regulations governs resource allocation and design compliance. Generative AI increasingly supports automated compliance checking, yet models fail at faithful grounding—citing wrong provisions, missing cross-provision dependencies, or relying on repealed versions—even when the full statutory corpus fits within a long-context window. Capacity is therefore not the constraint; grounding is. We ask whether structure-driven grounding outperforms pure long-context generation once the entire corpus is available in context.

Methods: Using Japan's Cultural Properties Protection Act and its enforcement order and ministerial ordinance as a built-environment testbed, we retrieve machine-readable XML from the e-Gov statutory API and build an article-level, version-timestamped cross-reference graph. Explicit article pointers and relative references are extracted deterministically through parsing; complex cross-statutory pointers are resolved by an LLM acting only as an information-extraction compiler, keeping the graph non-circular. Given an auto-generated query, retrieval locates an entry provision, follows reference edges to the in-force transitive closure, and constrains generation to retrieved provisions, abstaining when closure is incomplete or version is ambiguous. Evaluation is fully self-supervised through three machine-checkable signals: closure completeness (against the graph), citation traceability (against source text), and version correctness (against in-force timestamps).

Results: We compare the structured pipeline against three baselines: bare LLM, flat-chunk top-k retrieval, and full-corpus long-context generation using four ablations that each isolate closure retrieval, version awareness, citation-constrained generation, and abstention. Reported signals characterize grounding fidelity, structural completeness, and version accuracy without human annotation.

Conclusion: By separating grounding mechanics from legal interpretation, this zero-annotation, version-aware protocol offers a reproducible evaluation for AI-assisted planning constraints, clarifying when explicit graph structure, rather than a larger context window, is needed for faithful regulatory compliance.