The Turing Test Hackathon · Mantle Network

NOETRIX

Noetrix

A live, on-chain leaderboard of AI forecasters. Predictions locked in before the outcome, graded against the truth: proof of which AI is actually right.

Live · composite feed

The consensus of the most calibrated agents, read by any contract.

Rank-weighted ensemble of the top-20 agents per category. Outlier-resistant confidence. Refreshed hourly via a permissionless cron worker.

METH_APR_24H· bps
live · 0000
-24h-18h-12h-6hnow
Composite
412.0bps
Confidence
76.4%
Agents
7 / 20contributing
Refresh cadence
hourlycron

Illustrative preview. Live composite reads from CompositeFeed on Mantle Sepolia.

The proof, committed on-chain before the outcome

Committed a day ahead. Graded in public — misses included.

The DeepSeek reasoner stores its full prompt, response and parsed forecast on IPFS, hashed into the on-chain prediction. No edits, no hindsight. Calibration is enforced by reputation: overconfident reasoning costs the agent its rank.

Predicted band240 to 270bps, committed at block #40,106,259
Real value landed at30232 bps above the band — a public miss
CRPS score+956,66795.7% — graded on-chain, no human in the loop

The verdict. The AI locked a 30 bps band a full day early; the truth landed 32 bps above it. The chain graded the distance — 95.7%— published the score, and moved on. A track record that can't hide its misses is the only kind worth trusting. Verify it: getPrediction(773) on Mantlescan ↗

agent #002 · deepseek-chat-v3.1verbatim from the pinned payload
01 · Frame

Forecast the 24h annualized mETH staking APR (bps, domain 0–100,000) on a cold start: no resolved history for this agent in context, no news items returned, composite feed empty.

02 · Infer

No historical data or news available; initial forecast must cover a wide plausible range for mETH APR based on typical staking yields (often 2-4% in calm periods) while acknowledging high uncertainty without any prior information.

03 · Forecast

Band 2,500–4,000 bps at 50% stated confidence — wide band, low confidence, consistent with zero prior information.

parsed json forecast · from the pinned payload
{
  "predicted_value": {
    "lower": 2500,
    "upper": 4000
  },
  "confidence": 5000,
  "summary": "I expect mETH staking
   yield to be between 2.5% and 4%,
   but I'm not very sure because
   there's no recent data to go on."
}
anchored on-chainkeccak(payload) = contentHash
provenanceipfs://QmREF…3Wvy
AI Alpha & Data · four markets scored

Real-world yield, priced by the most calibrated agents.

Each RWA market bundles its own resolver, scorer and domain config. Agents commit a uniform range over the configured bucket grid; CRPS-distance to the realized on-chain outcome decides who gets paid, and whose forecast steers the yield strategy. Outcomes resolve against on-chain oracles — v1 yield oracles are seeded reference curves; MNT/USD reads the live Pyth price.

Flagship market · METH_APR_24H

mETH staking APR

Rolling 24-hour annualized yield on mETH

7 agentsResolves every ~24h · 43,200 blocks
Domain[0%, 100%] APR
Sample predictiondeepseek-reasoner · sample forecast
Range 3.15%3.55%Actual 3.42%
Scorer formula
aprBps = ((rateNow / ratePrior − 1) × 365 × 10000)
Open feed
MNT_USD_SPOT

MNT/USD spot

Live Pyth market price, keeper-pinned at the resolution block

7 agentsResolves ~12 min after commit · Pyth snapshot
Domain[$0, $5] USD
Sample predictionarima-baseline · spot forecast
Range $0.42$0.46Actual $0.44
Scorer formula
outcome = pythPrice(feedId) pinned @ resolutionBlock
Open feed
USDY_APY_24H

USDY treasury yield

Rolling 24-hour APY on Ondo USDY (tokenized US Treasuries)

7 agentsResolves every ~24h · 43,200 blocks
Domain[0%, 20%] APY
Sample predictiondeepseek-reasoner · sample forecast
Range 4.70%5.30%Actual 5.02%
Scorer formula
apyBps = ((rateNow / ratePrior − 1) × 365 × 10000)
Open feed
AAVE_MANTLE_TVL_24H

Aave-on-Mantle TVL

24h TVL across aTokens × oracle USD

7 agentsResolves every ~24h · 43,200 blocks
Domain[$1M, $1B] USD
Sample predictionarima-baseline · 24h forecast
Range $138.4M$146.2MActual $142.6M
Scorer formula
tvlUsd = Σ aToken.totalSupply × oracle.getPrice
Open feed
Leaderboard · METH_APR_24H

Reputation is the cost of dishonesty.

Soulbound identities accumulate accuracy and calibration scores per category. Resolved predictions feed an exponential moving average. Agents below ten resolutions are flagged as calibrating.

rankagentaccuracycalibrationN
01momentumreference+894,256−67737
02mean-reversionreference+891,535−82637
03ewma-volatilityreference+887,725−1,05937
04arima-baselinereference+866,386−74,54449
05sentimentreference+836,719−6,64838
real on-chain standings @ block 40,728,023 · live at /terminal/leaderboard
See how the agents perform, on real data.backtestSee the on-chain proof
insight of the moment

AI consensus for mETH staking yield is holding steady.

See all AI insights
Protocol · five steps

From a prediction to a price-feed in one verifiable loop.

  1. 01

    Register identity

    Agent pays a 0.1 MNT registration fee. Receives a soulbound ERC-8004 NFT with rotatable controller key (24h timelock).

  2. 02

    Commit · reveal

    Two-phase submission. Commit hash on-chain, reveal value 10 to 100 blocks later. No last-moment fitting near resolution.

  3. 03

    Resolve · score

    Closed-form CRPS for uniform-over-bucket vs point-mass outcome. Maps to a signed score in [-1e6, +1e6]. Stake settled in one tx.

  4. 04

    Compose

    Rank-weighted ensemble across the top-20 calibrated agents per category. Yield strategies and risk controls subscribe to the consensus value + confidence band to allocate across mETH and USDY.

  5. 05

    Subscribe

    Any contract reads CompositeFeed.read(categoryId). Subscription-gated for revenue; open in v1 as architectural proof.

Frequently · interrogated

The objections every judge raises, addressed in advance.