Retrieval-Induced Consensus Collapse: Agreement Is Not Independent Evidence in Multi-Agent RAG

Agent agreement is used as a correctness signal in multi-agent RAG systems, but this assumes independent agent failures. The assumption breaks when all agents query the same shared index; they retrieve the same documents and reach the same answer, making a shared error look like consensus. This is called Retrieval-Induced Consensus Collapse (RICC), and it's both common and hard to fix with usual methods. With a clean corpus, different query reformulations per agent still return the same top document 98.9% of the time, and scaling from three to seven agents only reduces this by 10 percentage points. This failure is due to the shared-index retrieval structure, not the way queries are phrased. Agreement-based detection has almost no room to work. SSVP detects only 2.2–7.1% of collapsed trials; Self-Consistency detects 0–2.2%. Both are close to the 3.7% clean-corpus trigger rate, which means there's no useful signal. The independence violation causing collapse also eliminates the divergence signal that detection relies on. Overlap-based remediation turns into unconditional re-retrieval, firing on 98.9% of clean trials. Query diversification recovers 86.1–97.3% of single-hop errors under oracle conditions, but accuracy shifts only 2.2 points with the problem document present ($p>0.05$), and multi-hop repair never exceeds 18%. In shared-corpus RAG, agreement is not a reliable safety signal. The fix needs to address the index, not the vote.