<![CDATA[Screenshot-to-code systems are often judged by visual resemblance, yet deployable prototypes also require coherent DOM structure, accessible semantics, and predictable rendering. We evaluated these dimensions on 484 Design2Code screenshot–HTML pairs using deterministic screenshot features, static markup audits, and CSS-hash-grouped five-fold cross-validation. Random Forest and Extra Trees predicted DOM size with R² values of 0.215 and 0.244, respectively, but accessibility-burden R² remained below zero, confirming that visual fidelity cannot substitute for markup inspection. Train-fold-only visual retrieval scored 78.26 on a composite consistency measure; structural reranking scored 79.59, and a deterministic evidence reranker combining predicted structure, accessibility, and static critical-render-path profiles scored 79.63. Its 1.37-point gain over visual retrieval was significant (95% bootstrap confidence interval 0.92–1.86; one-sided Wilcoxon p < 0.001) and cost 0.36 visual-similarity points. Conservative remediation removed 218 of 1,981 detected issues, raised zero-issue pages from 59 to 79, and preserved normalized body text and embedded CSS on all pages. On a fixed 25-page diagnostic subset, mean SSIM relative to each page's complete-CSS render increased from 0.693 for semantic HTML to 0.784 with typography; the complete renders independently reached median SSIM 0.996 against the stored screenshots. These findings establish an auditable evidence layer for subsequent LLM-assisted critique rather than an evaluated language-model critic.]]>
Copyrights © 2026
