The question has been waiting since casehubio/ledger#114 was filed: when L4 and L6 land, should EigenTrust run alongside Bayesian Beta?

No. The math settles it before implementation begins.

EigenTrust propagates trust through a peer attestation matrix — C[i][j] is how much actor i trusts actor j, normalised from direct interactions. In QuarkMind, the only attestor is AgentOrchestrator. It observes all four plugins and records outcomes. No plugin attests to any other. The harness row of C is the only non-zero row; all four plugin rows are zero.

With every plugin row zeroed, EigenTrust’s dangling-node fallback fires for all four actors simultaneously on every iteration. Power iteration oscillates rather than converges. Four actors is well within the pathology zone documented in GE-20260421-09d636 — the exact scenario where the Stanford paper’s recommended fallback creates 3-cycles.

So Bayesian Beta only. But the ADR draft got something wrong, and Claude caught it during review.

The initial version said “Bayesian Beta only — no casehub-engine-ledger dependency.” That is not accurate. TrustWeightedAgentStrategy — the routing component that reads trust scores and dispatches between competing plugin implementations — lives in casehub-engine-ledger. L6 needs the module. What QuarkMind avoids is not casehub-engine-ledger but EigenTrust computation specifically.

The module bundles two independent concerns: TrustWeightedAgentStrategy (reads trust scores, routes dispatch) and TrustScoreJob (computes EigenTrust GLOBAL scores nightly). The strategy dispatches on CAPABILITY scores — Bayesian Beta per-plugin ratios accumulated by casehub-ledger. GLOBAL scores are computed by TrustScoreJob but not consumed anywhere in the capability routing path. EigenTrust runs. Its output goes nowhere. Inert, not absent.

ADR-0009 records the distinction: L4 activates casehub-ledger runtime only (Bayesian Beta starts accumulating). L6 adds casehub-engine-ledger for routing (EigenTrust job runs, GLOBAL scores produced, not read). Revisit if peer attestation develops or the actor pool exceeds ~50.

<
Previous Post
Two Writing Styles, one commit — and a spec that forgot to check the skill
>
Next Post
The Sweep and What Stuck