Large language models are increasingly producing structured outputs that are consumed directly by programs and downstream systems. Semantic correctness alone is insufficient for reliable execution. To address this, we introduce SERIAL‑7, a Korean‑English benchmark built on an executable verification framework that spans seven representation formats.\n\nThe benchmark employs 177 validated canonical artifacts and decomposes the generation task into three stages: Derive – extracting the semantic goal from the description, Realize – generating artifacts under explicit constraints, and Preserve – ensuring consistency across representations. It defines 43 atomic representation constraints and provides executable checks for format, structure, and task‑specific semantics, enabling verification that progresses from basic compliance to full artifact validity.\n\nExperiments on ten models show REALIZE‑L3 human‑score rates ranging from 42.0% to 93.9%. Full compliance evaluation uncovers additional artifact‑level failures even when final answers are correct and representations appear valid.\n\nThese findings demonstrate that reliable structured generation requires explicit, executable verification of how artifacts are realized and preserved across representations.\n\nReview: SERIAL‑7 offers a systematic approach to evaluate LLMs’ structured output capabilities, highlighting the necessity of end‑to‑end checks from semantic derivation to cross‑representation preservation, and marks a significant step toward verifiable AI.