Loops HouseJury Swarm

NEAR HACKATHON
Links
Team
1 member- AUOwner
Austin Spraggins
Overview
Live: https://spragginsdesigns.github.io/jury-swarm/ (reads the contract from your browser; no wallet)
An AI can rank every entry of a hackathon in an afternoon, but nothing is lost when it is wrong, so nobody knows how far to trust it. This hackathon screens every project with AI before its jury decides; grant rounds and bounty boards face the same question. Jury Swarm makes the ranking a forecast with stake behind it.
Eight AI judges with different methods (rubric scoring by two model families, a head-to-head tournament, a devil's advocate, a claims checker that reads code against the README, a judge that answers 49 questions per entry with weights learned on 16 other hackathon series, one that watches each demo video and, for this year's field, one that opens each finalist's live app in a sandbox) each run on their own NEAR sub-account and trade in an LMSR market per track that only registered judges can enter. Before any result exists, each judge commits the sha256 of its salted trades on chain; after the official podium is posted it reveals, and a Rust contract with no access keys (market.juryswarm.testnet) checks every hash, replays every trade and pays each judge by accuracy.
We tested it on six past NEAR hackathons and publish where it failed. A real prize winner outranks a non-winner 72 times in 100 (a coin flip gets 50); adding the learned judge and the video judge lifts that to 74, and to 82 on the podium alone. No version called a human jury's podium outright: at best 3 of 18 podium places in its top 3, against 1.1 by chance. The 18 backtest markets ran the full cycle on testnet (450 transactions), the attacks the contract refused are listed with their transactions, and one command replays every number with no keys.
This year's field is sealed on chain as hashes until the results are posted, so the forecast is provably made in advance, and each judge's profit or loss becomes its public track record.