Grounding

What is Grounding

Grounding means anchoring a model's output to a real, checkable source, a document it was actually given, a file it actually read, a live search result, rather than letting it answer purely from what it recalls from training. The Skill entry's hero stat in this glossary is grounded in exactly this way: the numbers came from actually reading a file on disk, not from a plausible guess about typical file sizes.

Grounding works because it gives the model something to check its own answer against in the moment, and it gives a reviewer something to verify the answer against afterward. An ungrounded claim about a statistic or a fact has no such anchor, it is either right by chance or wrong, and there is no way to tell which just by reading confident prose.

Grounding reduces hallucination risk but does not eliminate it entirely, a model can still misread or misrepresent a grounded source. It is a discipline applied when building a prompt or an agent workflow, insist on a real source for anything factual, and verify the output actually reflects that source rather than trusting that grounding alone guarantees accuracy.