definition file
Definitions that can survive a review meeting.
LLM Terms Desk starts every definition with a practical question: what would change if this word appeared in a user notice, procurement checklist, product requirement, or public explainer? A term is accepted only after it separates a capability from a promise and a preference from an obligation. The goal is language that remains clear when quoted by a search result, copied into a policy memo, or read aloud during a product approval meeting.
Memory
A named product behavior that stores or retrieves user-specific information beyond the immediate request.
Grounding
A link between an answer and source material that is available for review, not a mood of confidence.
Model improvement
A separate use of submitted material to train, tune, evaluate, or monitor a system after the current service is delivered.
Human review
A stated responsibility to compare output against a policy, source record, or decision threshold before action.
How a definition is drafted
The desk writes in plain English first, then adds the review shape: actors, permitted data, duration, evidence, and exception. That shape matters because AI language often hides several promises inside one sentence. A useful definition lets the reader identify the exact behavior being named and the decision that depends on it.
What gets rejected
Terms are rejected when they rely on hype, pretend a probability is a guarantee, or describe a governance practice without naming who performs it. The desk also avoids definitions that are correct only for one vendor but presented as universal. Precision beats broadness when language affects consent, compliance, and user trust.