arXiv:2605.14381v5 Announce Type: replace-cross Abstract: Recent advancements in generative AI facilitate large-scale synthetic data genera- tion for model evaluation. However, without targeted approaches, these datasets often lack the sociotechnical nuance required for sensitive domains. We introduce NodeSyn, an evidence-grounded methodology that generates socially relevant syn- thetic queries by leveraging a fine-tuned taxonomy generator (TaG) anchored in real-world evidence. Evaluated against four mainstream LLMs, NodeSyn elicited failure rates up to five times higher than human-authored benchmarks. Ablation studies confirm that our granular taxonomic expansion significantly drives these failure rates, while independent validation reveals critical deficiencies in prominent guard models.