Jieyuan Yu, Lun Tao
With accelerating globalization and urbanization, vernacular architecture, as an important carrier of regional culture, is increasingly confronted with the dual fragmentation of its spatial environment and historical information. Conventional reconstruction research remains constrained by incomplete historical records and the lack of a systematic methodological framework. This study integrates a multiple-evidence analytical approach with digital humanities technologies to develop a comprehensive reconstruction method for perished vernacular architecture. The method proceeds through three stages: AI-assisted historical textual criticism, digital translation and inference of spatial information, and parametric modeling with visual representation, thereby enabling both spatial reconstruction and cultural-semantic interpretation. To enhance the reliability and transparency of the digital reconstruction process, this study proposes an Execution-Feedback-Revision (EFR) cognitive recursive model as a theoretical framework for human-AI collaborative reasoning and further reveals an evidence-constrained mechanism for the convergence of uncertain historical spatial cognition. Large language models expand the scope of spatial interpretation by integrating multi-source historical information and conducting semantic reasoning. Expert knowledge and multi-dimensional evidence systems jointly establish constraint boundaries for hypothesis evaluation and refinement. Through iterative recursive feedback, spatial hypotheses inconsistent with historical evidence are continuously identified and revised, progressively narrowing the interpretation space and transforming uncertain initial understandings of historical spatial knowledge into reliable spatial interpretations supported by evidence. Using the Maoshan Academy in Ningbo, Zhejiang, as a case study, the research systematically demonstrates the interactive negotiation among multi-source historical materials, AI-generated feedback, and human verification across key tasks such as site confirmation, spatial-sequence analysis, and roof-type identification. It further employs parametric techniques to generate adjustable digital models capable of expressing uncertainty, thereby reconstructing and analyzing the academy’s architectural form, spatial configuration, and cultural significance. The digital reconstruction methodology proposed in this study provides a systematic technical framework for research on extinct vernacular architectural heritage reconstruction and validates the evidence-constrained convergence mechanism of uncertain historical spatial cognition. It further offers a transferable methodological reference for the digital regeneration of extinct vernacular architectural heritage and vernacular knowledge spaces across diverse regional and cultural contexts.