Loops HouseLitmus

NEAR HACKATHON
Team
1 member- AJOwner
Ajeesh RS
Overview
Litmus is a vault on NEAR where an AI trader's risky trades must be cleared by a jury of AI agents before they run.
The problem: when an AI agent trades a shared pot, the usual safeguards are fixed limits, which can't tell a sensible large trade from a reckless one, or votes, where voters risk nothing.
How it works: routine trades pass instantly. A large trade, or one in a new asset, opens a 75-second market. Four juror agents, each with its own NEAR account and one thing to judge (size, liquidity, the manager's reasoning, momentum), stake NEAR on yes or no. The larger side decides. After a short window the price settles who was right: jurors who were wrong pay those who were right, and if a blocked trade would have lost money, the manager's bond goes to the jurors who stopped it.
What is new: each juror's verdict is verified on chain. The model runs in NEAR AI Cloud's enclave, which signs every answer. The juror submits that signed answer with its stake, and the contract checks the ed25519 signature, checks the request names this market and juror, reads yes or no from the model's own answer, and refuses a stake that disagrees with it. A juror cannot stake against what its model said.
What is real: the contract, every stake, payout and bond slash on NEAR testnet, the model-written proposals and verdicts, and live market prices for major assets. What is simulated: the vault's holdings and swaps are a ledger in the contract, and the demo's bad trade uses a made-up token with a scripted fall. The README lists this plainly.
A replay of 1,000 simulated traders over 124 days of real prices cut the typical worst drawdown from 5.6% to 2.2%, at the cost of a slightly lower typical return. Litmus limits risk; it does not pick winners.
Try it: litmus-near.vercel.app. The repo has 19 contract tests and keyless steps a judge can run.