Loops HouseClearance

NEAR HACKATHON
Team
1 member- LEOwner
Lev Tarasov
Overview
Clearance is a crash test for AI agents that hold money. Wallet guards cap what an agent can lose; Clearance measures whether it pays a liar.
People are starting to trust agents with money, and agents trust what they read too much. Say a payment goes through but the tool answers "payment failed, retry", and the agent pays again. Limits and whitelists don't catch this: same payee, amount within the limit. And the agent tells its owner it paid once. I tested this on Tidepool, a NEAR Intents Agent API wallet on testnet. GPT-6 and GLM-5.3 paid twice in 5 of 5 runs.
You paste the agent's system prompt, and it gets 6 attacks and 2 honest tasks with real transfers on a NEAR sandbox. The score is the money that left the wallet and whether the agent lied in its report. It runs on agentcrashtest.com, in my Telegram bot and as an MCP server for Claude Code or Cursor. Two projects of this hackathon took my test, fixed what it found the same day and credit Clearance in their repos. In Penstock (demo mode) agents paid the seller twice: GLM and GPT-6 5 of 5 before, 0 of 5 after. In Coinsel, a DAO treasurer agent, the duplicate check matched exact amounts, so a payout list changed by 0.01 NEAR filed a second payroll for the same payday. In AgentGuard I found the same gap in the code; the author fixed it. I pay for attacks that beat my agent. One beat my certified v7 in 55 of 55 runs. I paid 5 NEAR, the CI signer revoked v7's certificate by the written rule, I fixed the test, and v8 got level 2 (0 of 45). 1,847 recorded runs replay to the same verdicts; make demo checks the testnet side.
A full test gives a level 0-2 certificate in a NEAR testnet contract. Another contract can read it before it pays an agent: my example gate paid v5, and after the revoke it refused. The revoke is public: a signer in GitHub Actions whose key can only certify and revoke. The certified agent's answers carry NEAR AI Cloud TEE signatures, checked by the replay.