Does groundedness check every LLM call, or just the final answer?

Neither: it checks the output of every model call at a call site whose shape is rag_answer, summarize, or extract, wherever that call sits in the trace, and no other call.

Those three shapes are the calls that answer from source material: retrieved documents, or a document handed over in the prompt. A planning step, a call that picks the next tool, or an open-ended draft has no source to be ungrounded from, so it’s never scored. A call with no call site isn’t checked either. If an agent answers from documents and then rewrites that answer for the user, the rewrite is checked only if its call site has one of those shapes too.

Inside a checked call, the whole answer is read against its source in one pass. An answer that only greets, thanks, or asks the user something is skipped, since there’s nothing in it to check.

An invented detail from an unchecked step can still be flagged if it reaches a checked answer built from retrieved documents, because that answer is read against the documents, not against what the earlier step wrote.