Multi‑agent AI systems improve inference by spawning sub‑agents and synthesising reports. The core issue is that a new agent is not automatically a new observation: apparently independent reports may stem from the same evidence, and truly independent evidence can yield nearly identical reports. We formalise this as the epistemic Sybil problem. For a set of reports R, a report Z satisfies I(Θ; Z | R)=0 and is called an epistemic Sybil extension. Aggregators that rely solely on reports cannot generally distinguish replication from independent corroboration, because identical reports can lead to different posteriors under unobserved ancestry. We introduce a Gaussian shared‑root model showing that common ancestry does not imply complete redundancy. Repeated extraction adds information toward a source‑level ceiling, while correlated extraction errors induced by a shared base model among independent agents lower that ceiling further. Experiments involve over 20 000 controlled LLM‑agent report and extraction calls on synthetic evidentiary documents. With a single evidence root fixed, increasing report multiplicity from 1 to 32 collapses naive posterior coverage from 0.940 to 0.263. Holding report count constant while raising evidence‑root multiplicity from 1 to 16 closes the gap, and at k=16 the aggregators are statistically indistinguishable. The agents’ replicate extraction errors are correlated (γ_cal≈0.719 out‑of‑sample), and a correlated‑extraction aggregator restores calibration accordingly. A controlled manipulation isolates representation similarity from evidential ancestry: a report‑space deduplication mechanism’s mean inferred cluster count changes by 1.425 (95% CI [1.363, 1.485]), whereas a fourfold change in true ancestry shifts it by only 0.040 ([-0.045, 0.120]). The takeaway is that collective inference should track evidential ancestry and dependence rather than merely agent or report multiplicity or similarity.
Review