Loops HouseSirius on NEAR

NEAR HACKATHON
Team
1 member- ALOwner
Ali Ben Yezza
Overview
Private datasets are worth money, but sharing one means trusting a stranger with it, and letting an AI agent buy one means trusting it with your wallet. Sirius on NEAR turns each dataset into an NFT that owns a wallet. The provider seals the file to the training runner's key (X25519 + XChaCha20); only its hash goes on-chain. Before publishing, a sample is checked for personal data by private models on NEAR AI Cloud: we verify their TEE ourselves (Intel TDX quote, NVIDIA GPUs, our nonce) and encrypt the sample end-to-end to the key the quote binds. The audit hash is recorded in the NFT, and anyone can re-check it from a public transcript with no key and no data (verify-audit.mjs: 114 checks). Each NFT (NEP-171) is tied to a keyless vault account: training runs pay the vault, only the current holder can withdraw, and selling the NFT moves the income with it; the card is drawn by the contract on every read. Buyers can be AI agents spending a budget a human funded in the market, with a function-call key limited to order_with_budget; before paying, a counterparty check reads the chain (on-chain bonded reports, account age, pinned contract code, prior orders) and blocks a prompt-injected listing. Everything runs on NEAR testnet: the market is locked (no keys, reproducible build), live pages at sirius-near.onrender.com, 63 JS and 53 contract tests. The training runner is simulated on testnet, and the README says so. Pre-hackathon code (Sirius on another chain) is tagged and removed from the tree; everything in near/ was built during the hackathon.