Long‑context models read a novel token by token in narrative order, allocating a fixed attention budget to the entire history. A detective reasons differently: they order events temporally, keep a map of character relations, and let a clue from chapter 1 meet a question at the end. This work asks whether a frozen knowledge graph can grant the same freedom to a small local model. Experiments use a fixed qwen3.5:9b reader on thirty detective novels and 234 multiple‑choice questions under nine conditions: five graph routes, a recent‑window baseline, whole‑book compression, ordinary vector retrieval, and a question‑only control. The best graph route achieves 53.85% (126/234), outperforming the recent window (46.15%), compression (51.28%), vector retrieval (51.71%) and question‑only (40.17%). On the subset where no model can answer without the book, the graph route still reaches 42.86%. None of the fifteen graph‑baseline contrasts survive Holm correction, so the result is presented as exploratory evidence. Two structural findings are more robust: annotated evidence concentrates in the topological core of the graphs (2.35× enrichment, pooled), and the two graph‑building pipelines differ markedly in annotation coverage (16% vs 73% of clue paragraphs), meaning pooled accuracy alone would hide the specific bottleneck.
Review