Previous work shows that a model’s ability to define domain terms correlates to performance in other tasks related to that domain. For example, in a QA setting, a language model can miss a question in a specialized domain for no reason other than not properly recalling what a term in the question means. We argue that a short glossary definition is the smallest unit of evidence a retrieval system can supply against that failure. Concretely, focusing on the finance NLP domain, we assess the role of definitions in improving performance on XFinBench, and also pairing MMLU-Finance with a 13,609-term Investopedia glossary of financial terms. We experiment with a five-model panel spanning mid-capability open models to a frontier closed model, and show that definition-only retrieval produces a structured gain on these datasets, with the effect concentrating in a mid-capability band rather than scaling with model size. Finally, we connect term-level measurement to individual questions through a human-validated annotation, and find that term knowledge does not locate the questions where a definition changes the answer, with none our implemented internal routing signals approaching the per-question oracle.