Preprint reports bounded-context gains for grid-model AI queries

Authors of an arXiv preprint report a seed-anchored graph-rendering method for LLM question answering over CIM and CGMES power-grid models. On a preregistered 100-item SmallGrid bank under an 8,000-character context budget, reported accuracy rose from 0.450 to 0.970. The authors also report matching or exceeding certain extracted graph representations without LLM graph-construction tokens. The result is limited to the evaluated models, reader and context budget.