Generalising HARF Beyond Bamiyan: A Bounded Cross-Site Evaluation of a Generative-AI Heritage Reconstruction Framework at the Temple of Bel, Palmyra

Abstract

The Heritage-Aligned Reconstruction Framework (HARF) was developed and initially evaluated using the Western Buddha of Bamiyan to structure evidence-constrained generative-AI heritage reconstruction. This article examines its bounded generalisation to the Temple of Bel at Palmyra, a materially, formally and iconographically contrasting heritage monument. The Dynamic Prompt Blueprint was re-instantiated using documentary evidence concerning the temple’s dimensions, limestone construction, Corinthian order, architectural layout, inscriptions and decorative programme, while retaining the framework’s top-level structure. A fourteen-item Palmyra questionnaire block was completed by 32 members of the interdisciplinary panel who also completed the Bamiyan evaluation. Within the Palmyra evaluation, Corinthian-column representation received the most favourable rating, whereas restored relief, inscription and surface detail received the least favourable rating. Eighteen respondents (56.2%) selected “none of the iterations are historically accurate” when assessing relief iconography. The paired difference in global prompt-sufficiency ratings did not reach the conventional significance threshold (p = 0.061), although ratings were positively associated across the two sites (Spearman ρ = 0.664, p < 0.001). The seven-domain keyword-coverage profiles were also positively rank-correlated (ρ = 0.771, p = 0.043). These findings support the bounded re-instantiation of HARF’s prompt schema and diagnostic evaluation logic in a contrasting heritage context. They do not establish equivalence between the two implementations, archaeological validity of the generated Temple of Bel images or validation of the complete HARF workflow.