Faithfulness

The degree to which claims in an output are supported by evidence supplied to the system, such as retrieved documents, tool results, or structured records. Grounding makes that evidence available in context; faithfulness checks whether the answer stayed within it. Correctness refers to reality or an external reference, while attribution accuracy checks whether each citation points to the right support.

An answer may faithfully follow an outdated source and still be factually wrong. Evaluation should decompose the output into claims, locate support or contradiction, and record which passages justify the score. Response-level averages hide critical claims, so sensitive domains require consequence-aware penalties and review of cases with insufficient evidence.