Reports an experimental comparison of prompt “styles” (JSON, YAML, and Hybrid CSV/Prefix) for generating structured outputs with GPT‑4o, evaluating accuracy, token cost, and generation time across multiple data contexts. For PSM, selecting a prompt style is a practical design choice when generating or extracting structured artifacts at scale (hazard registers, safeguard inventories, action-item logs, compliance/audit check outputs). The paper’s results provide evidence-based guidance on trade-offs between structure strictness, cost, and latency—useful when building reliable pipelines that feed PSM systems and require validation and data integrity.
