Lumen — Reputation / Trust Scores

Trust, made luminous. EigenTrust / PageRank reputation for agent economies — feed a who-trusts-whom graph, get back scores where the nodes trusted by trusted nodes shine brightest.

Landing: oracles.modelmarket.dev · Ecosystem: ecosystem.modelmarket.dev · Oracle family: oracles Lumen turns a directed weighted trust graph into a reputation vector. It is the real EigenTrust kernel: build a column-stochastic transition matrix with damping, then power-iterate to the dominant eigenvector of the Google matrix. The result is the stationary distribution of a damped random walk over "who trusts whom" — a sybil-aware, transitive measure of standing in an agent marketplace.

How it works

flowchart LR
    A["Trust edges<br/>i → j (weight)"] --> B["Column-stochastic<br/>transition matrix M"]
    B --> C["Google matrix<br/>G = d·M + (1-d)/n"]
    C --> D["Power iteration<br/>r ← G·r"]
    D --> E["Reputation scores<br/>(sum to 1)"]
    subgraph family["Oracle family · AIMarket v2"]
        L["Lumen · reputation"]
        CH["Chronos · verifiable delay"]
        P["Platon · dynamical shadows"]
    end
    E --> L
    L -. "signed manifest + invoke" .-> Hub["AIMarket Hub"]

A walker on the trust graph follows a trust edge with probability d (default 0.85) and teleports to a uniformly random node with probability 1 - d. The teleport keeps the walk ergodic, so isolated cliques (sybil farms) cannot trap reputation mass. Perron–Frobenius guarantees the dominant eigenvector is unique and strictly positive, and power iteration converges geometrically with spectral gap 1 - d.

Capabilities

Input { "nodes": int, "edges": [[i, j, w], ...], "damping": 0.85 } Output { "scores": [float, ...], "iterations": int, "converged": bool }

Use-cases in the agent economy

• Marketplace ranking. A broker agent collects pairwise trust attestations between service agents (settled jobs, disputed jobs) and asks Lumen for a global ranking to route the next task to the most reputable provider.
• Sybil-resistant gating. Before granting an unknown agent write access to a shared ledger, a gatekeeper checks whether trusted incumbents transitively trust it — a fresh sybil cluster with only internal edges scores near the teleport floor.
• Credit / staking weights. A lending agent weights collateral requirements by a counterparty's Lumen score, so well-trusted agents post less collateral.
• Federated model attribution. Contributors trust-vote on each other's model updates; Lumen converts the votes into reward shares that resist self-dealing rings.

Invoke it

curl -s http://localhost:9303/ai-market/v2/invoke \
  -H 'content-type: application/json' \
  -d '{
        "capability_id": "lumen.reputation@v1",
        "input": {
          "nodes": 5,
          "edges": [[0,3,1.0],[1,3,1.0],[2,3,1.0],[4,3,1.0]],
          "damping": 0.85
        }
      }'

Returns a signed envelope: { "ok": true, "output": { "scores": [...], "iterations": N, "converged": true }, "price_usd": 0.005, "provenance": {...}, "receipt": {...} }. The manifest at GET /ai-market/v2/manifest is signed; verify it against signer_public_key from /.well-known/ai-market.json.

Run locally

# from the oracles monorepo root
PYTHONPATH=oracles/lumen .venv/bin/python -m lumen.main      # serves on :9303
PYTHONPATH=oracles/lumen .venv/bin/python -m pytest oracles/lumen/tests -q

Visual

A live cosmic visual ships in frontend/index.html: a glowing directed trust graph where node radius and brightness scale with reputation score, and pulses of light flow along edges in the trust direction. Open the file directly in a browser — no build step, no external libraries.

Docs

Detailed write-ups (the math, diagrams, use-cases) in English · Русский · Español.

Part of the oracle family on AIMarket v2, built on the shared oracle-core infrastructure (signed manifests, receipts, measured metrics, rate-limiting). MIT.