Evaluating Generative AI Reconstructions of the Bamiyan Buddhas: Computational Similarity and Expert Archaeological Assessment

Evaluation of generative AI reconstructions remains a central methodological challenge in digital heritage research. This study evaluates the operational performance of the Heritage-Aligned Reconstruction Framework (HARF) through empirical application to the Western Buddha of Bamiyan. A corpus of 245 reconstruction candidates was generated using Dynamic Prompt Blueprint prompts across two generative platforms and evaluated through a staged pipeline combining computational similarity analysis, interdisciplinary screening, and specialist expert assessment. Structural similarity and feature-matching metrics identified candidates with geometric correspondence to archival imagery, while expert reviewers assessed archaeological plausibility using a structured evaluation rubric. Results show a consistent divergence between algorithmic similarity scores and specialist judgement. Computational metrics effectively filtered candidates by geometric correspondence but remained insensitive to stylistic micro-grammar and iconographic conventions in Gandhāran sculpture. Proportional coherence and niche integration were the most reliable indicators of plausibility, whereas facial morphology and drapery articulation remained persistent instability domains. These findings demonstrate that generative systems cannot be treated as autonomous reconstruction tools but as mechanisms for producing exploratory visual hypotheses requiring evidentiary constraints and expert validation. The study clarifies the complementary roles of computational filtering and domain expertise within AI-assisted heritage reconstruction and identifies the methodological limits of generative models in culturally specific archaeological contexts. Preprint notice: This is the author manuscript/preprint version of a manuscript currently under review at Journal of Cultural Heritage. It has not yet been formally accepted or published as the Version of Record. If accepted, this Zenodo record will be updated to link to the final published article.