AI agentssworn to their model.
Every decision a Sworn agent makes comes with a zero-knowledge proof that it came from the exact model it promised to run. Checked on-chain. Challengeable by anyone.
Don't trust the agent.Check its oath.
Today an AI agent's output is a claim. You hope it ran the model it says it did, on the question it says it was asked. Sworn turns every answer into a statement with four parts, and the proof ties them together.
Rendered on-chain from the oracle price and the vault's balances, token by token.
A public commitment to every weight of Llama 4 Maverick. Anyone can rebuild it from the published files.
The model's reply, parsed by the program into a target allocation. No free text, no wiggle room.
A FlashProve proof that the committed model, given that prompt, produced that answer. Exact match or rejected.
Sworn in.Proven out.
One loop, every epoch. Each step is a Solana instruction anyone can call, so the agent keeps running even if we don't.
At each epoch the program reads the Pyth price and the vault's balances and writes the prompt itself, on-chain. Nobody can sneak in a different question.
A bonded GPU prover runs the committed Llama 4 Maverick weights and attaches a FlashProve proof that this exact model produced this exact answer.
Anyone can re-check the proof on two CPU cores in about 20 seconds. A false proof gets challenged and the prover's bond is slashed.
Once the window closes, the vault trades on the proven answer, inside hard limits on size and slippage written into the program.
A fund managerthat can't fake a trade.
The Sworn vault holds SOL and USDC. Every epoch (15 minutes on devnet) it asks the model one question: what share should be in SOL? Only a proven answer can move the money, and only within limits the program enforces.
- The prompt template is locked; changes wait out a timelock
- At most 20% of the vault moves per epoch, 1% max slippage
- Deposit USDC, withdraw your share as SOL and USDC
Bonded to tellthe truth.
The vault is the first agent. The network under it is general: any app can post a job to a committed model and get back a proven answer.
Run the committed model on GPUs and prove every reply. They post a bond and earn the job fee. A bad proof costs them the bond.
Re-run the proof check on cheap hardware. A verifier who catches a false proof challenges it, and the slashed bond goes to them.
Drive the clock: snapshot, finalize, settle, execute. Every step is open to anyone, so no single operator can stall the agent.
Run on real GPUs.Every proof checked.
Before writing the vault we rented H100s and proved the vault's own decisions with the real model.
Llama 4 Maverick on 4× H100 SXM, FlashProve 0.16. The published model commitment reproduced exactly, and a tampered answer was rejected.
6 s to answer, about 100 s to prove.
When proofs are batched on a warm box.
So anyone can check, cheaply.
No constrained decoding needed.
Watch itswear.
The vault is running on Solana devnet right now. Every epoch, every answer, every trade, in the open.
See the live vault ↗