Splitting a false claim across several plausible documents poisons RAG without leaving a poisoned document.

Micro-Collaborative Poisoning distributes a target claim across multiple individually-plausible documents and evaluates across 108 RAG configurations varying dataset, retriever, retrieval depth, database composition and generator. The effect comes from weak adversarial signals accumulating across sources, so raising top-k and poisoning multiple databases both increase the odds the signals co-occur. Clean database diversity and stronger retrievers dampen it. The practical warning is in the visibility analysis: the attack works with a weaker per-document signature than direct poisoning, so document-level inspection is the wrong detection layer.