Loops HouseChainmind

NEAR HACKATHON
Links
Team
1 member- JIOwner
jimleston osoi
Overview
Problem: smart contracts can't use AI without trusting someone. Today an oracle, an API company or a secure server runs the model, and the contract has to believe its answer.
What I built: Chainmind puts a complete language model inside a NEAR smart contract. The weights (stories260K, 260,032 parameters, int8, 299 KB) live in contract storage and are sealed forever by their SHA-256 hash. When you call generate, every validator runs the full transformer in WebAssembly as part of the transaction and gets the same words. No oracle, no API key, no trusted hardware.
The engine is written for determinism: integer matrix products, its own exp function, precomputed position tables and a fully specified random generator. The NEAR VM, a laptop and a browser produce the same tokens bit for bit, so anyone can re-run any onchain answer and check it.
How NEAR fits in: it's live on NEAR testnet at mind-mwdofvuu.leston.testnet. A real 48-token story cost 194 Tgas (about 0.02 NEAR). Measured in nearcore's VM, one call writes up to about 70 tokens and stops safely before the 300 Tgas limit. Other contracts can call it like a shared library, and the Endless Story lets anyone add a chapter written by the chain.
Try it: https://chainmind-theta.vercel.app runs the same sealed model in your browser and checks its output against the chain's. The README has five do-this, see-that steps for judges.
Next: split bigger models across NEAR shards, and use the engine to verify large off-chain models on chain.