Introduced on day 51
Faithfulness (groundedness)
Whether every claim in a generated answer is supported by the retrieved context; it measures support, not whether the context itself is true or current.
In detail
Faithfulness is usually computed by splitting an answer into claims and checking each against the context, by people or by a model judge that needs its own validation. A faithful answer built on an outdated passage is wrong yet scores perfectly, so it is reported next to correctness against a reference and next to retrieval metrics.