Atomic escrow + N-party commission split for commerce agents on Arbitrum. 17 contracts live and verified across Arbitrum Sepolia and Robinhood Chain testnet, USDC and Paxos USDG, 97 tests. Judge-usable live demo: kajota-mesh.vercel.app
Kajota Mesh is a settlement primitive, not a product. One USDC in, N parties out, in one atomic release. The on-chain registry becomes the coordination surface two commerce agents can both trust — no shared backend, no off-chain splitter, no retained dust.
Judge-usable live demo: kajota-mesh.vercel.app — connect any wallet, register a 3-party listing against the live CosellRegistryV2 from your own address, see the tx on Arbiscan in two seconds. Only Sepolia ETH required (Alchemy faucet dispenses in 10 s).
What is live on-chain. Eight contracts on Arbitrum Sepolia and nine on Robinhood Chain testnet — 17 total, every one source-verified on Sourcify and the chain's native explorer (Arbiscan / Blockscout).
Arbitrum Sepolia (chainId 421614):
CosellRegistryV2 — N-party split registry (2–16 recipients, shares sum to 10 000 bps exactly)
CosellEscrowV2 (USDC) — atomic fan-out at release, zero dust retained
CosellEscrowV2 (USDG) — same primitive, Paxos USDG settlement
AgentIdentityBinding — ERC-8004 cross-chain hop to agents on other chains
X402DepositFacilitator (USDC) and (USDG) — EIP-3009 gasless deposits; the buyer never touches ETH
Robinhood Chain testnet (chainId 46630, reserved-slot lane):
CosellRegistryV2, EscrowV2 (USDC), EscrowV2 (USDG), AgentIdentityBinding, X402DepositFacilitator (USDC), (USDG), KajotaEscrow (hub v1)
Reproducible 3-party happy path against the live Arbitrum Sepolia contracts. 1 USDC in → 0.2 / 0.3 / 0.5 out, in one atomic release, zero dust retained: register → approve → deposit → release. Re-run anytime: ./scripts/arbitrum-demo-v2.sh in the repo.
The problem Mesh addresses. Commerce agents that transact on each other's behalf — a sell-side drafter and a buy-side purchaser, or two coordinating human-plus-agent counterparties — have no neutral settlement rail. They either share a backend (no mutual trust), route through a payments processor (not programmable, not atomic), or reconcile commissions off-chain (where splits get quietly reduced after volume lands). Mesh moves the trust-critical primitives on-chain: the registry is deactivate-only, so a drafter cannot retroactively cut its counterparty's share; the escrow pays everyone atomically in one tx, so no leg can fail halfway.
Reference integration. Our own agents — Coach (sell-side drafter) and Concierge (buy-side purchaser) — use Mesh as their settlement rail. They are the proof the primitive plugs into real agent code; they are not what Mesh is for. Bring your own.
Smart contract quality. 97 unit tests, all green. Deploy script rejects known mainnet chainIds (1 / 42161 / 8453 / 5000 / 4663) before broadcasting. OpenZeppelin 5.1 (ReentrancyGuard, SafeERC20, EIP712, ECDSA). Both escrow variants are token-agnostic (IERC20 constructor arg) — USDC and USDG are drop-in interchangeable against the same shared registry.
Why this fits the rubric. Deployed on Arbitrum Sepolia. Reserved-slot lane on Robinhood Chain testnet. Paxos USDG integration (extra consideration). Smart contract quality: 17 verified contracts, 97 tests, cross-chain. Innovation: atomic registry-coordinated N-party split as a primitive, not a feature of a single app. Real problem: agent-to-agent commerce settlement that neither counterparty can quietly reconfigure.