Text-Grounded LLM-Assisted Design Rationale Interfaces: Turning Advertising Layout Metadata into Explainable UI/UX Decision Cards

Automatic graphic design generation has progressed from template retrieval to layered generative pipelines, but designers still need explainable review support that links generated layouts to practical UI/UX criteria. This paper presents a text-grounded, LLM-assisted rationale-card interface for advertising layouts using OpenCOLE annotations, including intention, description, keywords, background caption, object caption, heading, and subheading. The interface converts these fields into structured decision cards covering visual hierarchy, white space, call-to-action (CTA) prominence, brand tone, layout balance, and risk warning. An experimental evaluation was conducted on all 23,419 rows from the OpenCOLE train, validation, and test splits. Because expert UI critique labels are unavailable, the tasks are evaluated as weak-label schema-recovery tasks rather than professional design-quality judgments. A linear SVM achieved 0.888 accuracy and 0.872 macro-F1 for nine-way design-purpose recovery. The best CTA model reached 0.817 accuracy, 0.801 macro-F1, and 0.884 quadratic weighted kappa, while text-based layout-proxy prediction reached 0.734 accuracy and 0.706 macro-F1. Structured cards improved required-dimension coverage from 4.21 to 5.92 out of 6 and field-grounding from 0.681 to 0.918. The findings show that OpenCOLE text fields can support inspectable first-pass design review without replacing expert judgment or direct visual-layout evaluation.