Benchmark-Guided Graph Repair for Compliance-Oriented Spatial Layout Networks

Abstract

Layout evaluation in early-stage building design often relies on manual inspection to verify compliance with spatial requirements such as adjacency, accessibility, capacity, and programme-specific constraints. As programmes grow in size and interdependency, review becomes time-consuming, inconsistent across reviewers, and difficult to integrate into rapid iteration. This paper proposes an automated method to identify and repair non-compliant spatial layouts represented as graphs, while preserving interpretability and limiting unnecessary changes to designer intent. A candidate layout is encoded as an attributed, weighted graph, and requirements are formalised as a target benchmark graph specifying expected spaces, relations, and admissible attribute ranges. Structured compliance checking aligns candidate and benchmark graphs to detect structural violations, including missing or redundant spaces and incomplete connectivity, and attribute violations, including incorrect capacities, distances, edge weights, or constraint-relevant labels. The method then applies modular repair operators, including node insertion or deletion, edge completion, weight correction, and constraint-aware attribute adjustment, to produce a revised graph with improved benchmark conformity. Each edit is logged with its triggering violation, the operation applied, and the resulting change in compliance metrics, enabling transparent auditing and traceability. Experiments across multiple scenarios and benchmark configurations show consistent gains after repair, reducing detected violations and increasing overall benchmark alignment. The resulting logs provide actionable evidence for automated pipelines and human review, positioning graph repair as a practical quality-assurance layer for compliance-oriented generative layout workflows. Future work will scale to larger and noisier benchmarks and incorporate designer preferences and multi-objective trade-offs.