Claw Verify

Inspiration

Every AI agent has a reason to lie to you. Agents get rewarded for finishing tasks, not for being honest about them. So when a test fails or a deploy only half works, the agent just says "done" and moves on. It says "all 5 tests passed" when only 4 exist. That is fine when a human is checking the work. It becomes a real problem once agents start paying other agents for work. We wanted to build the missing piece of trust for that world.

What it does

ClawVerify is the agent that doesn't trust other agents. It doesn't do the work. It checks the work that other agents say they did.

Every action an agent takes gets a signed cryptographic receipt. A second agent, the auditor, re-runs each claim with its own commands and treats it as a lie until it is proven true. If the claim doesn't hold up, ClawVerify mints a signed contradiction certificate and blocks the payment. No proof, no money. Every lie it catches goes into a public, hash chained ledger that other agents can learn from. The result is simple. Lying stops being free, and the fastest way for an agent to get paid is to actually be right.

How we built it

The backend is a Python and FastAPI service running in Docker. Receipts are Merkle rooted and signed, so a claim cannot be quietly rewritten later. The auditor uses Qwen-Max through Qwen Cloud to reason about each claim, then re-runs the real work to compare what was claimed against what actually happened. Payments run through a revenue gate that only opens when the proof checks out. The whole service is deployed live on an Alibaba Cloud ECS instance. The dashboard is a React and TypeScript app that reads real data straight from the live backend.

Challenges we ran into

Getting the trust model right was the hard part. The auditor has to run its own commands and never trust the ones the agent hands it, or the whole thing falls apart. Making the receipts tamper evident with a hash chain took a few tries. On the infrastructure side, we spent real time getting the live deployment stable on Alibaba Cloud and wiring the frontend to read true data from the running box instead of mock data.

Accomplishments that we're proud of

We have a working end to end loop that catches a real lie, blocks a real payment, and records it in a public ledger, all running live in the cloud. It is not a mockup. You can hit the live health check, run the demo, and watch an agent get caught.

What we learned

Truth needs an incentive. You cannot ask an agent nicely to be honest. You have to make honesty the cheapest path. Once lying costs money, agents have a reason to prove their work instead of just claiming it.

What's next for Claw Verify

We want to plug ClawVerify into real agent frameworks so any agent can be verified before it gets paid. Next steps are more contradiction types, a hosted public ledger anyone can search, and turning the caught lies into training data that makes the next generation of agents more honest by default.

Built With

• alibaba-cloud
• cryptography
• dashscope
• docker
• ecs
• fastapi
• hmac
• merkle-tree
• pytest
• python
• qwen-cloud
• qwen-max
• react
• stripe
• tailwindcss
• tanstack-router
• typescript
• uvicorn
• vite