Preprint measures evidence replication risk in AI-agent reports

A preprint formalizes an "epistemic Sybil" problem in multi-agent AI: additional reports can stem from the same evidence rather than provide independent corroboration. In more than 20,000 controlled LLM-agent calls on synthetic evidentiary documents, the authors report naive posterior coverage fell from 0.940 to 0.263 when report multiplicity rose from one to 32 with one evidence root. They also report that a correlated-extraction aggregator restored calibration.