Three models — Opus 5, Sonnet 5, Haiku 4.5 — each answering twelve records in each of three corpora over the Proto-OKN federation, three attempts apiece: 540 recorded episodes. Each study is the control for the next, so they are best read in this order.
Can a model query one knowledge graph correctly, with no join and no hierarchy in the way? Opus answers 36 of 36; Haiku misses 14. That is what makes the other two studies readable: a failure there is a failure at what that study adds, not at writing SPARQL. (question by question)
Twelve cross-graph joins under three tool conditions — the full toolkit with the graph named, the graph withheld, and six tools withheld at the proxy. The only study with an ablation axis, and so the only one that can price the server rather than the model. (question by question)
A scientist naming a disease category does not say “and its subtypes”. Six questions
whose answers need a subClassOf* closure, each paired with a control that forbids the
expansion — so the same prompt measures discovery in one arm and obedience in the other.
(question by question)
Each study's questions, what makes one hard, and how to add another: single-graph, cross-graph, ontology. How each study's figures are built and read: single-graph, cross-graph, ontology. The harness itself is described in the repository README.
The harness, all three corpora and every recorded episode are in the repository.