Retrieval‑augmented generation (RAG) grounds a language model in retrieved documents, which reduces hallucinations but opens a new attack surface: poisoned documents can mislead the model. This work measures the robustness of the quantized Llama 3.1 8B model when a fraction of its retrieved context is tampered with. Three poisoning strategies are examined—entity swap, number swap, and negation. Each strategy is applied to zero, one, two, or three of the three retrieved passages, yielding a factorial sweep of 588 runs on a fact‑checking task built from FEVER. Accuracy drops from 77.9% with clean context to 43.5% when all three passages are corrupted. Entity swap flips the largest share of previously correct answers; number‑based corruption leaves accuracy flat while poisoned passages are a minority and drops sharply once they become the majority. The model mostly abstains rather than inventing new falsehoods, and a lexical‑overlap proxy for unsupported generation decreases under attack instead of rising. This is a small‑scale measurement with coarse automated labels, so the observed patterns are suggestive pending controlled decoding and stronger adjudication.
Review