Grounding

Tying a model's claims to a verifiable source, like a document or a tool result, instead of letting it answer purely from its own trained knowledge.

In practice: This is the fix for a model that sounds confident but is just pattern-matching from training data rather than checking anything real.

Debated: Some practitioners use it narrowly to mean citation-backed answers with a traceable source; others apply it to any case where a tool result shapes the response, even without a citation.