Spawning more agents on the same evidence collapses posterior coverage from 0.94 to 0.26

arXiv·medium signalXBlueskyLinkedInCopy link

This paper formalizes an epistemic Sybil problem: another agent is not another observation, because apparently independent reports can descend from the same evidence root and no report-only aggregator can generally tell replication from independent corroboration. Over 20,000 controlled LLM-agent report and extraction calls on synthetic documents confirm it. Holding one evidence root fixed while report multiplicity rises from 1 to 32 drops naive posterior coverage from 0.940 to 0.263; holding report count fixed while evidence roots rise from 1 to 16 closes the gap entirely. Replicate extraction errors from a shared base model are correlated at an out-of-sample gamma of 0.719, which lowers the information ceiling further. Fanning out N agents over one source buys confidence, not information.