Model Grounding

Grounding is the practice of tying a model's output to something outside the model: a retrieved document, a live measurement, a record with a source attached. An ungrounded model produces fluent text whose relationship to the world is unverified, which is tolerable for a draft and unacceptable for anything that informs a position. Grounding does not make a model correct - it makes it checkable, which is the property that matters in production.

In practice it means the pipeline must supply facts with enough structure and attribution to be cited. That is a data requirement rather than a modeling one: resolved entities so the fact attaches to the right company, ground truth close to the original observation, provenance that survives the trip, and freshness measured rather than assumed. It is the same set of guarantees a research desk asks for, requested by a different buyer through retrieval-augmented generation.