Abstract: In architectural heritage conservation, the digital reconstruction of ornaments requires not only visual similarity but also the preservation of historically grounded compositional logic. Existing deep learning–based generative approaches primarily rely on image-level style features and lack explicit mechanisms for representing compositional rules, often resulting in structural deviations in symmetry, hierarchy, and pattern organization. To address this limitation, this study proposes an evidence-to-prompt generative framework that transforms heterogeneous historical documentation into computable design variables through an attribute–slot–sub-slot semantic ontology. A Minimal Semantic Unit (MSU) representation is introduced to convert textual evidence into structured generative prompts, while visual grammar principles are formulated as constraint conditions integrated into the diffusion process to regulate compositional structure. Using Chinese historical eave-tile as a case study, a dataset containing 2, 300 image records and 3, 800 textual evidence entries is constructed to evaluate the proposed method across multiple model configurations. Results show improved performance in symmetry consistency, structural continuity, and overall compositional integrity compared with baseline approaches, indicating that structured semantic representation and grammar-aware constraints enhance the stability and controllability of generative outcomes. The proposed framework provides an interpretable and reproducible pathway for knowledge-driven generative design in architectural heritage digitalisation.