← All questions
RAGEasy· Attribution· Evaluation· Error Analysis

What is the next step when an attribution judge returns...

A binary ‘not attributable’ label is insufficient because it conflates two distinct failure modes: the model is extrapolating beyond the available evidence (silence) or the model is making a claim that contradicts the evidence (refutation). Retrying the same generation or retrieval without diagnosing which case is occurring is ineffective. You must distinguish between these scenarios to decide whether to fetch more data or flag a potential hallucination.

Code playground (Python)