Verifiable AI on Sui. A language model runs off-chain at near-native speed, and anyone can prove on-chain that it ran wrong, down to a single arithmetic operation a small contract re-executes. Below are four demos. Each shows one thing. They are all real and running on Sui testnet.
A resolver bonded a wrong AI judgment. Re-run the one disputed micro-op on-chain and take its money. This is what the others can't do cheaply.
The real Qwen3-1.7B reads a proven fact and types its verdict, deterministically. Same bits on any CPU or GPU. That is what makes it disputable.
Pick a real API — Bitcoin price, exchange rate, live weather, GitHub stars, a headline, a match result. Generate a real zkTLS proof on the spot, watch it recover the pinned attestor, and let the judge read the proven value.
An outcome decided by counting independent zkTLS proofs against a committed rule, then paying out. No vote. No token holder can change a signature check.
Want everything in one place? The full prediction-market app combines all of it with a guided tour. The engine itself is written up in opml/README.md.