How do I tell a stale RAG index apart from a truncated context window?

An answer from a stale index can pass faithfulness and groundedness because it accurately repeats an old chunk. Truncation is a context-presence failure: the right chunk was retrieved, then cut before the model saw it.

So compare what retrieval returned with what reached the model. If the needed chunk was never retrieved, the miss is in the index or the retriever, not proof that the context window cut it. If it was retrieved and then disappeared from the prompt, check context assembly and the window limit. If the model saw an old chunk and repeated it faithfully, compare that chunk with the current source and check when the index was last built.

Neither one shows up in the trace as an error. A tool returning stale data reads the same way, for the same reason: every span reports success.