A security team lead is explaining RAG (Retrieval-Augmented Generation) to her team. She describes a failure mode where the RAG system returns outdated threat intelligence because the vector index was built from documents crawled three months ago and hasn't been refreshed. What is this failure mode called, and what operational control addresses it?

• A.Temperature drift: the LLM generates different responses over time because its sampling temperature changes; feature importance scores from gradient boosting are contractual metadata under most cloud terms of service, and providers must ship them with every prediction, making opaque model behaviour a billing violation rather than a technical risk
• B.Context window overflow: too many documents were retrieved and the LLM cannot process them all; attention heads in a transformer are trained on disjoint vocabulary shards, meaning each head owns specific words outright and removing one head cleanly deletes the model's knowledge of those tokens, giving auditors per-token attribution
• C.Stale index failure: the retrieval system returns outdated content because the document corpus in the vector store has not been updated to reflect new threat intelligence; addressed by establishing a defined index refresh cadence (e.g., daily for fast-moving TI sources) and monitoring document freshness metadata