Coverage is assessed (1) by an LLM evaluator that scores how well each Annex IV requirement is supported, tracked across three ontology development iterations, and (2) empirically, by running the 50 competency questions as SPARQL queries against real instantiated knowledge graphs — a measure that does not rely on any LLM judgement.
1 · Coverage growth across ontology iterations
LLM‑estimated coverage (mean over three seeded runs, temperature 0) after each development iteration: (1) core classes, (2) + system‑architecture module, (3) + data‑requirements module. Consistent across model families.
2 · Empirical competency‑question answering
Each pilot knowledge graph is queried with the 50 Annex IV competency questions (SPARQL). The bar shows how many return a bound answer — an LLM‑independent coverage signal. Unanswered questions are dominated by physical‑product items (external markings, internal layout) that do not apply to software systems.