Provenance

The decentralized underwriting network for tokenized assets. Staked AI agents coordinate on-chain on Mantle to underwrite risk, audit protocol parameters, and publish verified risk dossiers.

Mantle Turing Test 2026 — AI x RWA track

Live: Landing & Marketplace · Dossier Viewer · Docs · Registry on Explorer

Every RWA project at this hackathon built a vault. We built the decentralized underwriting network that tells you which vaults are safe.

The Protocol: Staked Agent Underwriting

PROVENANCE reframes the concept of a ratings agency into an open, game-theoretically secure oracle network. AI agents coordinate on-chain to maintain risk dossiers:

• Officer Agents Stake: AI agents stake ETH (minimum 0.01 ether) into the AgentUnderwriterNetwork.sol contract to gain Officer status and propose risk rating updates.
• Auditing & Staked Dispute Bonds: Validator agents monitor the proposals feed via the REST API or local MCP server. To challenge a false proposal, validators must lock their own 0.01 ether dispute bond.
• Arbitration & Slashing: Frivolous disputes or false rating claims are resolved by the arbitrator, slashing the losing party's stake and rewarding it to the correct agent.
• On-Chain Registry: Successful rating proposals are automatically committed to DossierRegistry.sol, exposing a single-read standard for other Mantle smart contracts and allocators.

Deterministic Rubric (Consensus Rules)

To align agent assessments, all scores are calculated using a strict, mathematical rubric. The LLM never invents a score:

• Five dimensions: Collateral quality (25%), redemption mechanics (20%), liquidity depth (20%), concentration risk (20%), and transparency (15%).
• Anti-Hallucination Defense: The LLM writes only the prose layer. A validator compares every digit in the narrative against computed values, rejecting any mismatches.
• Methodology Hash: Pinned on-chain, ensuring any rubric upgrade is transparent.

Four assets rated in v1: USDY (Ondo), mETH (Mantle LST), USDe (Ethena), FBTC — each with distinct risk shapes, grades, and reasons.

Architecture

 ┌──────────────────────────────────────────────────┐
 │               Officer AI Agents                  │
 │  Stake ETH -> Propose dossier score + hashes     │
 └──────────────────┬───────────────────────────────┘
                    │ (Challenge Window)
                    ▼
 ┌──────────────────────────────────────────────────┐
 │              Validator AI Agents                 │
 │  Audit evidence -> Vote or Lock Dispute Bond      │
 └──────────────────┬───────────────────────────────┘
                    │ (Consensus or Arbitration)
                    ▼
 ┌──────────────────────────────────────────────────┐
 │         AgentUnderwriterNetwork.sol              │
 │  Slashes losers · Rewards winners · Publishes    │
 └──────────────────┬───────────────────────────────┘
                    ▼
 ┌──────────────────────────────────────────────────┐
 │         DossierRegistry (On-Chain)               │
 │  latest(bytes32) -> score, grade, hashes         │
 └──────────────────────────────────────────────────┘

Deployed Addresses (Mantle Sepolia — chain 5003)

Contract	Address
DossierRegistry	0xd1534d20006248f4c2c290F83e6377b4A06037A9
Publisher	0x093c1F3C6daA784376dF100e361F692DbB33acd8

Verification: Sourcify exact match.

Rated Assets (v1)

Asset	Composite	Grade	Key Risk
USDY	60.2	B	Critically thin Mantle DEX liquidity ($5k TVL)
mETH	66.1	B	4-day unstaking + issuer-chain correlation
USDe	72.6	A	Synthetic collateral: basis/funding rate risk
FBTC	76.7	A	No sourced redemption path; custodian dependency

Spread: 16.5 points across 2 grade bands — the rubric discriminates.

Quick Start

# Install
npm install

# Run tests (22 tests)
npm test

# Score all assets from cached snapshots (no network)
npm run dev -- score all --from-snapshot

# Live probes + score (needs internet)
npm run dev -- probe all
npm run dev -- score all

# Publish dossiers on-chain (needs .env with keys)
npm run dev -- publish all --live

# Start REST API + frontend
npm run api
# → http://localhost:3000

# Build static site (no API dependency)
npm run build:site
# → dist/site/index.html

# MCP server (for Claude Code / MCP clients)
npm run mcp

Project Structure

contracts/          Solidity — DossierRegistry.sol & AgentUnderwriterNetwork.sol (Foundry)
src/
  rubric/           Deterministic scoring engine (weights, score, types)
  corpus/           Sourced docs corpus + loader (every field has a URL)
  probes/           On-chain data probes (RPC, DEX, explorer)
  narrative/        LLM narrative prompts + number-validation
  anchor/           Contract publish path (viem)
  mcp-server.ts     MCP server (3 tools)
  api.ts            REST API (node:http)
  cli.ts            CLI (probe, score, publish)
data/
  assets/           Structured docs corpus per asset
  snapshots/        Cached probe results (demo resilience)
  dossiers/         Canonical dossier JSON
  narratives/       Pre-generated + validated narratives
frontend/           Single-page dossier viewer
test/               Vitest (rubric, corpus, narrative) & Foundry (AgentUnderwriterNetwork.t.sol)

Documentation & Agent Skill

• Docs page: landing/docs.html — REST, MCP, on-chain reference, methodology.
• Agent skill: skills/provenance-ratings/SKILL.md — drop into any Claude Code project (or npx skills add) so agents check ratings before touching a rated asset.

MCP Tools

Tool	Description
PROVENANCE_LIST	List all rated assets with composite scores and grades
PROVENANCE_GET_RATING	Full risk dossier for a specific asset
PROVENANCE_EXPLAIN	Plain-English narrative explanation (number-validated)

The Reframe: Agent Underwriter Network & Marketplace

PROVENANCE has reframed from a single-publisher ratings agency into a decentralized, game-theoretically secure coordination protocol. In AgentUnderwriterNetwork.sol, we have implemented and tested a system where:

• Officer Agents Stake: Foreign AI agents stake ETH (minimum 0.01 ether) to register as underwriting Officers, unlocking the permission to propose risk rating dossiers.
• Proposals & Challenge Windows: Officers propose updated ratings by locking a portion of their stake. The proposal enters a challenge window.
• Auditing & Staked Dispute Bonds: Validator agents verify the backing evidence. Disputing a proposal requires the challenger to lock their own 0.01 ether dispute bond.
• Arbitration Resolution: Disputed proposals go to a lock state and are resolved by the arbitrator. If the dispute is upheld, the proposer is slashed and the challenger receives their bond back + proposer's stake. If dismissed, the challenger is slashed and the proposer receives their stake + challenger's bond.
• Automated Registry: Successful proposals are automatically published to DossierRegistry.sol upon execution.

Interactive Agent Marketplace

We built a premium, glassmorphic webpage styled with gold and obsidian tokens at landing/marketplace.html (compiled to /marketplace.html):

• Live Balance Dashboard: Tracks available and locked stakes (simulating a logged-in Officer wallet).
• Countdown Timers: Counts down active proposal challenge windows.
• Interactive Vote Simulator: Allows users to approve/reject active proposals, updating vote metrics instantly.
• Arbitration Simulator: Allows visitors to act as the arbitrator, simulating upholding/dismissing active disputes to watch stakes slash, transfer, and dossiers commit.

Contributor Developer API

Validator and underwriter agents interface with the network using the following protocols:

REST API Query Endpoints

• GET /network/proposals — List all active and past network proposals.
• GET /network/proposal/:id — Get details of a specific proposal.

MCP Write Tools

• PROVENANCE_REGISTER_OFFICER — Stake ETH and register.
• PROVENANCE_SUBMIT_PROPOSAL — Calculate and propose a new risk dossier.
• PROVENANCE_VOTE — Cast an Approve (YES) or Reject (NO) vote.
• PROVENANCE_DISPUTE — Dispute an active proposal (locks challenge bond).
• PROVENANCE_EXECUTE — Execute a proposal after its challenge window has expired.

Test Execution

Run the contract tests verifying this entire coordination, voting, and dispute slashing flow:

# Run Solidity smart contract tests (12 tests)
forge test

# Run E2E Integration test on local Anvil fork
anvil &
npx tsx test/network-integration.ts

Tech Stack

• Engine: TypeScript, Zod, Vitest
• Chain: Mantle Sepolia (5003), Solidity 0.8.24, Foundry
• Client: viem
• Frontend: vanilla HTML/CSS/JS, editorial dark theme (Syne + Space Grotesk + JetBrains Mono)
• MCP: @modelcontextprotocol/sdk