Red-team your RAG with deliberately poisoned retrieval and measure the parametric-override rate

arXiv:2606.07783·medium signal

This reliability protocol (arXiv:2606.07783, submitted June 5, 2026) tests RAG on factoid questions the model already answers correctly without retrieval, then injects clean, misleading, and mixed evidence to see when retrieved content overrides correct internal knowledge. It introduces parametric-override and confidence metrics to quantify how strongly poisoned context flips outputs. Actionable: before shipping a RAG system, inject plausible-but-false passages into your retriever, measure override rates and confidence degradation, and prefer configurations that hold factual accuracy under adversarial retrieval.