Help me test: do modern retrieval systems mostly retrieve consensus rather than truth? [D]

In three linesResearcher proposes LOGOS-SIE, a synthetic dataset of 500k observations/beliefs across 5k facts and 100 sources, to test whether modern retrieval systems recover consensus rather than truth. Hypothesis: BM25, dense retrieval, and rerankers favor dominant patterns even when 90% of sources are false.

Summary generated by Claude — human-verified