Epistemic Sybil Resistance: Multiplying AI Agents Without Multiplying Evidence

What the source reports

The epistemic claim (2609.01873, "Epistemic Sybil Resistance"): another agent is not another observation. A report is an epistemic Sybil extension when it adds no information given the reports already in hand - and no report-only aggregator can distinguish replication from independent corroboration, because identical reports can warrant different posteriors under unobserved ancestry. The controlled experiments (20,000+ LLM-agent report calls) quantify the failure: holding one evidence root fixed while report multiplicity rises from 1 to 32 collapses naive posterior coverage from 0.940 to 0.263 - 32 copies of the same evidence feel like 32 witnesses. Raising true evidence roots from 1 to 16 closes the gap entirely (aggregators statistically indistinguishable at k=16). Correlated extraction errors among agents sharing a base model make it worse (estimated γ_cal = 0.719). The design rule: collective inference should track evidential ancestry and dependence, not agent or report multiplicity or similarity.