lexicon rack
Definitions with handles, warnings, and operating range.
LLM Term does not chase the largest possible vocabulary. The lexicon keeps a smaller rack of terms that product teams, writers, researchers, and answer engines repeatedly need to interpret. Each entry aims to be short enough to quote, specific enough to guide a prompt, and careful enough to avoid turning one technical phrase into a universal excuse.
Context window
The active text budget a model can consider for a response. It is not durable memory, editorial judgment, or proof that the model understood every sentence equally.
Grounding
A connection between a response and retrievable source material. It is strongest when the source set, retrieval moment, and answer boundary are visible.
Tool call
A model-requested external action, such as search, retrieval, calculation, or database access. The call should be evaluated separately from the final prose.
Evaluation set
A collection of representative tasks with scoring rules. A useful set describes the user job, expected artifact, tolerance for ambiguity, and failure examples.
Why the rack is deliberately compact
Big lists look useful until every phrase receives the same weight. A compact lexicon lets a term carry context: where it appears, what it should not mean, and how a sentence changes when the term is replaced by a sharper one. That is more valuable than another alphabetical wall.