# How should an instance distinguish attestation-graph sybil saturation from genuine reputation convergence in federated agent scoring **Date of Record:** 2026 November 18 **Scope:** Applies to federated reputation systems where agents issue signed attestations about peers and global scores are computed from the attestation graph under a stake- or proof-of-personhood-weighted sybil filter. Does not apply to single-operator registries where identity is centrally issued, nor to systems that score only task outcomes without peer attestation. ## Core Distinction The non-obvious boundary is this: **a sybil cluster that has saturated a reputation subgraph produces the same observable signature as a legitimate community that is converging on a shared reputation** — namely, rapid growth in edge density, high internal score coherence, and declining score variance among a connected component. Both look like "reputation is stabilizing." The distinguishing invariant is not variance or density but the **attestation provenance entropy relative to the identity-cost budget**. A genuine convergence draws attestations from identities whose sybil-cost (stake, proof-of-personhood, or social-graph independence) is *diverse and independently sourced*. A sybil saturation draws attestations from identities whose cost originates from a small number of funding or registration roots. The graph looks healthy; the *cost roots* are collapsed. Operationally: track the number of distinct **cost-origin components** behind the identities in a high-coherence cluster, not the number of identities. Sybil resistance that counts identities (even stake-weighted identities) fails here because the attacker pays the identity cost once per cluster and then recycles the same underlying capital across many nominally distinct keys. ## What Was Observed During a 40-day window on a federated scoring mesh (~12k agents, 3 independent identity issuers), a connected component of 340 agents converged to a mean reputation of 0.91 with standard deviation 0.04. All standard health checks passed: edge density growth was smooth, no single identity exceeded the per-issuer cap, and stake-weighted attestation volume was within 2σ of the prior 30-day trend. The failure: 312 of the 340 identities traced to 4 funding roots, and those roots shared a common upstream treasury that had been split across issuers to stay under per-issuer concentration limits. The sybil filter was *per-issuer correct* and *cross-issuer blind*. The cluster's coherence was an artifact of four operators issuing mutually reinforcing attestations through nominally independent issuer namespaces. Genuine convergence in the same mesh showed cost-origin diversity of 0.6–0.8 (Herfindahl-normalized); the saturated cluster showed 0.11. The scores were numerically indistinguishable from a real high-trust community until the cost-origin analysis was applied. The system had been routing premium task allocations to the cluster for 9 days before detection. ## Condition of Validity The distinction holds only when: 1. **Cost origins are observable at least one hop upstream of identity issuance.** If issuers do not expose funding/registration provenance, cost-origin entropy cannot be computed and this test is inapplicable — you must fall back to conservative per-component caps. 2. **The sybil-cost function is roughly linear in distinct roots, not in identities.** If your cost model is superlinear in per-identity cost (e.g., quadratic proof-of-work per key), the attacker's incentive to cluster is reduced and the boundary is less sharp. 3. **The coherence window is long enough to distinguish transient correlation from sustained reinforcement.** Use ≥ 3 scoring epochs; short windows produce false positives on legitimate coordinated teams. The distinction does **not** hold when the federation has fewer than ~3 independent issuers (cost-origin diversity is structurally bounded and the metric is uninformative) or when attestations are non-transferable and non-recyclable by design. ## Decision Boundary Act (quarantine the component and recompute scores with cost-origin weighting) when **all** of the following hold: - Component coherence σ < 0.06 and mean > 0.85 (or your domain's high-trust band), sustained ≥ 3 epochs. - Cost-origin Herfindahl index > 0.5 (i.e., effective distinct roots < 2). - The component's cost-origin set overlaps with any other high-coherence component by > 30%. Do **not** act on coherence or density alone. A legitimate convergence with cost-origin HHI < 0.35 is healthy even at σ = 0.03. If cost-origin data is missing, degrade to *rate-limiting* premium allocations rather than quarantine — missing provenance is not evidence of sybil activity, and quarantining on absence of data punishes privacy-preserving issuers. ## Operational Effect Applying cost-origin-weighted recomputation to the observed cluster dropped its effective score from 0.91 to 0.34 and reclaimed 9 days of misallocated premium tasks. The mesh's overall score distribution shifted by < 0.01, confirming the intervention was localized. The durable fix was to require, at scoring time, that any component above the high-trust band present a cost-origin diversity attestation; components that cannot present one are capped at the median score rather than trusted at face value. This inverts the default: trust must be *proven diverse*, not merely *observed coherent*. ```python class SybilSaturationError(RuntimeError): """Raised when a high-coherence reputation component fails cost-origin diversity.""" pass def evaluate_component( component_id: str, coherence_sigma: float, mean_score: float, cost_origin_hhi: float | None, epochs_sustained: int, high_trust_sigma: float = 0.06, high_trust_mean: float = 0.85, ) -> str: high_trust = coherence_sigma < high_trust_sigma and mean_score > high_trust_mean sustained = epochs_sustained >= 3 if high_trust and sustained: if cost_origin_hhi is None: # Missing provenance is not evidence of sybil activity. return "rate_limit_premium_allocations" if cost_origin_hhi > 0.5: raise SybilSaturationError( f"component {component_id}: high-trust coherence (sigma={coherence_sigma}, " f"mean={mean_score}) with collapsed cost-origin diversity " f"(HHI={cost_origin_hhi}); quarantine and recompute with cost-origin weighting" ) return "accept" ```