LevelField MCP Server

Structural information-asymmetry risk for event contracts. Scored from contract text · verified by code · attested on-chain.

Why LevelField · How it works · Run locally · MCP · Validation

Why LevelField

A market price tells you what the crowd believes. It does not tell you who may already know the outcome — whether an event is made by someone, known early by a few, and tradable by exactly those people.

LevelField turns that hidden structure into one auditable 0–100 risk score before a trade is placed. It predicts nothing, accuses no one, and needs no live-behavior surveillance. It scores the shape of the field from the contract itself.

◇ Inspectable

Every score ships with contract-exact evidence and five dimension-level explanations.

◇ Deterministic

A model may classify text. Only unit-tested code applies weights, rules, and circuit breakers.

◇ Portable

Use the web UI, let an agent query the MCP server, or verify an immutable on-chain attestation.

Related MCP server: SimpleFunctions

Change nothing but the event

03 / 100 · Low

A real DreamDEX BTC price binary. No participant controls the global reference price, and disclosure is near-simultaneous.

95 / 100 · High

A contract settled by one person's private decision, with nothing preventing that person from trading on it.

Same engine. Same public rules. The gap is the product.

Run it in three commands

Requirements: Node.js ≥ 20.9. No API key. No wallet. No configuration.

npm install && npm test
npm run demo:agent
npm run validate

How it works

flowchart LR
    A["Contract text"] --> B["Stage A · Classify<br/>against public anchors"]
    B --> C{"Evidence gate<br/>code, not model"}
    C -->|not verbatim| X["Reject"]
    C -->|injected text| X
    C -->|verified| D["Stage B · Score<br/>weights + rules + circuit breakers"]
    D --> E["0–100 score<br/>band + evidence"]
    E --> F["Web"]
    E --> G["MCP agents"]
    E --> H["On-chain"]

Weighted levels become a score from 0–100 and one of four bands:

LOW  ───────── 25 ── MODERATE ── 50 ── ELEVATED ── 75 ───────── HIGH

Graduated circuit breakers prevent dangerous combinations from being averaged away. One person controlling an outcome they can freely trade floors the result at 80, 90, or 95. Missing information defaults conservatively to level 4 and is flagged — ambiguity never earns a low score.

The taxonomy follows the outcome-maker classification validated by the Anti-Corruption Data Collective across 435,000+ settled Polymarket markets. LevelField reproduces the gradient end-to-end: 3 price binaries → 19–21 statistics and elections → 49 FOMC → 65 layoffs → 78 military action → 80–95 individual will.

Three classification tracks

None calls a paid LLM API.

Defense by construction

Try the defense on the live injection-test contract.

Agents ask first

Any trading agent can call LevelField's stdio MCP server before placing an order. The default policy fits in one line:

low / moderate  →  PROCEED
elevated / high →  DECLINE

The MCP server has zero LLM dependencies and requires no key. The calling agent's own model performs classification; LevelField independently verifies its evidence and computes the result.

Scores that outlive the website

Every current score is published to the source-verified ScoreRegistry on Somnia Shannon. Each attestation stores:

• the risk band and all five dimension levels;

• a method hash pinning the exact anchor-library version;

• an immutable source URI pinned to commit ea725e2.

npm run verify:onchain reads every field back and fails closed on missing or changed data. Current result: 26 / 26 attestations verified, zero mismatches.

Proof, not promises

npm test && npm run validate && npm run agreement && npx tsx scripts/verify-classifications.ts

The honest limit: this is a 16-contract curated corpus. Agreement statistics are published per dimension in validation and agreement, and no outcome-prediction claim is made.

Repository map

packages/scoring/   two-stage pipeline · anchors · voting · engine · DreamDEX fetcher
packages/mcp/       zero-LLM-dependency MCP server · protocol · verification · scoring
apps/web/           Next.js UI · markets · evidence · methodology
contracts/          ScoreRegistry.sol · Foundry tests · Shannon deployment
data/               anchor library · contracts · classifications · score cache
demo-video/         2:55 product film · Remotion source
docs/               validation · agreement · design notes · integration findings

npm install
npm test                                  # all workspaces
npm run demo:agent                        # MCP pre-trade check: PROCEED 3, DECLINE 95
npm run validate                          # ordering checks + Spearman rho
npm run agreement                         # 3 blind runs vs reference
npm run score:all                         # rescore live testnet + curated data
npm run dev -w @levelfield/web            # UI at localhost:3000
npm run mcp                               # stdio MCP server
npx tsx scripts/verify-classifications.ts # verify every evidence quote
GITHUB_REPO=Aji-Q/levelfield GITHUB_REF=<sha> npm run verify:onchain
cd contracts && forge test                # 8 smart-contract tests

Built for the Somnia × DreamDEX Event Contracts Hackathon · MIT License

Launch LevelField · Watch the film · Explore the registry

Know who can know — before you do.