Prove which model produced an AI answer, and that it wasn't tampered with, using verifiable inference.
The Brief
Build a dapp where an AI answer comes with a receipt: a TEE attestation or zkML proof binding the model, the input, and the output together. Users can post the receipt onchain, and a verifier contract or page checks it. Visualize the proof as a glass box that turns green when verified and cracks red when tampered with.
Vibe Aesthetic
Transparent laboratory, x-ray overlays, clinical cyan
Key Features
- Inference behind a TEE or zkML proving pipeline
- Onchain or client-side verification of the receipt
- Tamper demo showing a failed verification
- Shareable verified-answer cards
Optional Jev Layer
TypeSafe Jev (System One) can sit beside your LLM for fast, typed routing and gates—you can ship v1 without it and add this layer when you want calibrated decisions in code.
- LLM + Jev split: LLM presents answers; Jev verify/fail attestation checks with calibrated confidence
Suggested Stack
Swap in whatever you like. Ship on a testnet first, and keep agent spending limits enforced onchain, not just in the prompt.
0 Comments