Trial GraphRAG on a global question (stretch)

Both systems answer the same eight global questions with the same map and reduce prompts, so they differ only in what the map step reads: the text units of the best-matching graph communities, or batches of per-document summaries. The decision that matters most is planting the themes at known rates: the generator records which document carries which theme, so "did the answer find the recurring issues?" has a true answer, and grading becomes a check rather than a matter of taste. Every model call goes through claude -p, so the trial runs on your subscription. Microsoft's graphrag package is left out, because it needs an API endpoint and an LLM pass over every chunk.