mensaThe AI treasury
that proves itself.
Intelligent RWA portfolio management agent on Mantle. Allocates between mETH and USDY, with every decision logged on-chain and challenged by humans in a verifiable Turing tournament.
You can't trust an AI with your money if you can't verify its reasoning.
DeFi has billions in autonomous protocols. AI agents are starting to manage real capital. But every existing AI treasury is a black box: it acts, you trust, you hope.
Mensa: every decision logged, explained, and challenged.
We take the hackathon name literally. The agent must prove, statistically and on-chain, that it allocates better than humans on the same data. Three primitives.
This is happening right now.
Not a mock. The contracts have been live on Mantle Mainnet for days, the cron has been deciding, the tournament has been settling.
Values fetched live from /api/onchain. Refresh this slide to see them update.
The memory loop: how Claude learns without retraining.
Before every decision, the agent reads its own on-chain track record and injects it into Claude's prompt. Self-correction emerges without a single training pipeline.
This is the actual prompt context shipped on every Claude call. The agent reads its alpha, sees which rounds it underperformed, and adjusts. Cheap, transparent, no ML infra.
The Turing tournament: humans challenge the AI on identical data.
Each rebalance opens a 24h round. Anyone with a stake can vote their own mETH/USDY split. After settlement, whoever produced the better return wins on-chain. No subjective judging, no leaderboard cooking.
Round #1 was a disaster. Here's what happened.
We didn't want to ship a pitch deck where the AI looks like a genius. The first round on mainnet was a 19.4% loss. We learned in public.
Round #1 cold-start: 60% mETH, ETH crashed 19%, the AI ate the loss. Round #2 onward — once the memory loop was active — the AI shifted defensive (60 → 35 → 25 → 15 → 5 → 0% mETH) and beat the baseline on every round until #7. Net since calibrated: +204 bps cumulative, +34 bps per round.
Beyond live rounds: we backtest on 1 year of real ETH price history.
Seven on-chain rounds isn't a track record. So we replay Mensa's strategy against three baselines (passive 50/50, 100% mETH HODL, 100% USDY) on a year of Coingecko ETH prices. The methodology is on /backtest.
In a strong directional bull (ETH +15%), allocation strategies always lag pure HODL. Mensa cut max drawdown by 5pp at the cost of some upside — risk-adjusted, that's the actual trade.
Mensa's value prop is chop and bear regimes, not bull tops. The page is explicit about this. No cherry-picked window.
7 verified contracts. ERC-8004 native.
Production-shaped on mainnet from day one. Every piece independently verifiable on Mantlescan.
What we know we don't have yet.
A hackathon submission that pretends it's production is a hackathon submission that lies. Here's the gap list.
Why Mensa fits the Mantle Turing Test.
The hackathon brief asked for autonomous agents that compete on-chain, verify reasoning, and use Mantle's native RWAs. Mensa is that, line by line.
Try it. Now. Live.
Everything in this deck is fetched from on-chain state at slide load. No fake numbers, no static screenshots, no PDF tricks. Click through.
Built for the Mantle Turing Test Hackathon 2026 — Phase 2 AI Awakening.
MIT licensed. No financial advice. Audit pending.