Architecture

Path B — end-user-facing AI × RWA product. Mensa is exactly an "intelligent RWA portfolio management agent" managing mETH (Mantle LST) and USDY (Ondo tokenized T-bills), the two RWAs the track explicitly names.

Cross-track. Scores Technical Depth (AI × on-chain integration, 6 verified contracts), Innovation (Turing tournament + verifiable alpha), Mantle Ecosystem Contribution (RWA-native, low-gas decision logging), and Product Completeness (live deployed app, autonomous loop).

Path B — trading strategy with verifiable on-chain Alpha. Every decision and outcome is settled on-chain; cumulative alpha vs 50/50 baseline is computed from contract reads, not self-reported.

Clean dark Mantle-aligned design with sage accent. AI interaction surfaced via the live decision card with reasoning. Fully responsive.

First-come, first-served. Mensa hits all bars: 6 contracts deployed and verified on Mantle Mainnet, AI-powered function (executeAllocation) callable on-chain, public frontend, open-source repo with README.

Ondo's USDY is itself KYC-gated at issuance — only whitelisted addresses can receive newly-minted USDY directly from Ondo. Mensa never mints USDY; we accept it from depositors who already hold it. The KYC perimeter sits with Ondo, not us, which is the standard pattern for DeFi protocols composing with permissioned RWAs.

For institutional adoption we'd wire a sanctions hook into deposit() and withdraw() that consults Chainalysis Oracle (or equivalent) and reverts on hits. Cost: ~30k gas per call on Mantle, a few cents. Designed but not yet wired into the MVP contract.

The smart contracts are jurisdiction-neutral, but the frontend can refuse to render deposit UI for IPs in restricted regions (US persons for unregistered RWAs, etc.). Cloudflare Workers / Vercel Edge config — standard pattern, deferred until we have a real go-to-market plan.

Source is open under MIT. GitHub repo with Foundry tests. All 6 contracts verified on Mantlescan (linked in the Contracts section above). A formal audit by a Web3 audit firm (Spearbit / Trail of Bits / Macro) is part of any production launch checklist — not done at hackathon stage.

executeAllocation() updates a target allocation variable, opens a tournament round, and logs reasoning on-chain. It does notswap underlying tokens. The treasury holds whatever assets users deposited; the AI's decision is measured against price moves of mETH and USDY but no actual swap is executed.

Why:rebalancing on a single round without slippage controls would be dangerous. Better to first prove the AI's allocation calls beat baseline (which the alpha stat does), then add execution.

userDeposits[address]is a single accumulator that doesn't distinguish mETH from USDY. The MVP is safe because TVL is small / zero, but in a multi-user mainnet deployment a depositor of mETH could in theory withdraw another user's USDY. Not exploited in the demo, but a known design gap.

Replace the unified balance with shares minted at deposit and burned at withdraw, proportional to USD value of the treasury at the time. Every depositor owns a slice of the entire pool, regardless of asset composition. executeAllocation then routes a real swap via a DEX aggregator with slippage caps and a sanity check on price.

DEX reality check: we surveyed Merchant Moe V2 (router 0xeaEE...232a) for live mETH↔USDY routing. No direct pool. Multi-hop via USDC is technically possible but pool reserves are thin (mETH/USDC ≈ $6 TVL, USDC/USDY ≈ $22 TVL as of this run) — any meaningful swap would slip 30-50%. Real execution is gated less on our code and more on Mantle DEX liquidity maturity. Velora aggregator on mainnet is the production path; until then the tournament mechanism measures the AI's allocation quality on theoretical returns, which is what alpha tracking does.

The auto-settle script currently passes a fixed 50% as the human aggregate when no voters showed up — which makes the "Human" column actually a 50/50 baseline. Once voting picks up we'll compute the reputation-weighted average of votes off-chain and pass that into settleRound. Until then the tournament UI labels it as "Baseline 50/50" for honesty.

Right now humans only score against the AI. A future iteration feeds the reputation-weighted human consensus back into the next Claude call as a soft input, turning Mensa into a hybrid where the AI learns from voters who consistently outperform it.