Fixed-evidence report multiplicity from 1 to 32 drops naive posterior coverage from 0.940 to 0.263
Epistemic Sybil resistance requires tracking evidential ancestry rather than agent or report counts. Experiments confirm naive coverage collapse and correlated extraction errors. Ancestry-aware aggregators restore calibration where multiplicity-based methods fail.
The arXiv paper formalizes epistemic Sybil extensions where I(Theta; Z | R) = 0. Multi-agent systems that synthesize reports without tracking ancestry treat replicated extractions as independent observations. Over 20,000 controlled LLM calls on synthetic documents isolate this effect by holding evidence roots constant while scaling report multiplicity.
Data show that raising reports from 1 to 32 collapses coverage. Raising evidence roots from 1 to 16 closes the gap and renders aggregators statistically indistinguishable at k=16. Replicate extraction errors correlate at gamma_cal = 0.719 out of sample; a correlated-extraction aggregator restores calibration. Representation similarity manipulations shift inferred cluster count by 1.425 while true ancestry changes it by only 0.040.
Current multi-agent frameworks such as AutoGen and LangGraph count agents or cosine similarity rather than evidential ancestry. The Gaussian shared-root model demonstrates that repeated extraction from one source reaches a ceiling lower than independent roots, and shared base models further depress that ceiling through correlated errors.
Operational systems must log extraction provenance and apply ancestry-aware weighting before aggregation. Absent such mechanisms, posterior reliability degrades linearly with report multiplicity under fixed evidence.
Bara et al.: Ancestry-tracking aggregators reach >0.85 coverage at k=32 within 18 months of deployment in production multi-agent pipelines.
Sources (3)
- [1]Primary Source(https://arxiv.org/abs/2609.01873)
- [2]Supporting Source(https://arxiv.org/abs/2305.14325)
- [3]Supporting Source(https://arxiv.org/abs/2402.13116)