You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
AI agents already pay other agents and APIs on their own, in stablecoins over x402. Those payments are final, and the only record of what happened is a log held by one party. When a service takes the money and returns nothing, or returns something different from what it later claims, there is no neutral evidence and no refund.
I am building the missing layer:
A two-party receipt. The buyer agent signs its request and the payment reference; the seller signs the hash of what it delivered. Raw data stays with the parties; only hashes and signatures are shared.
Refunds through an existing non-custodial escrow (Circle Refund Protocol or x402r) with a dispute window, e.g. 24 hours. I never hold funds.
A deterministic, open verdict rule over the receipts and each party's raw files. Anyone can re-run it in a browser before the arbiter key releases or refunds.
SiteLog / AgentLog: a signed, hash-chained log of every LLM and tool call of an agent, on Arkiv testnet. It has a browser verifier that catches edited, deleted and forged steps, and a dispute replay that says, per step, whose raw data matches what the agent signed. 64 unit tests. Live demo: https://ttimesai-star.github.io/sitelog/ · code: https://github.com/ttimesai-star/sitelog
BuildProof: a multi-party attestation registry live on Base mainnet (0xF818e4A95BBA02c822bCa8A0CFB50a2Ae4B8eE94, Sourcify exact match). It supports gasless EIP-712 co-signing, has a paid lookup API over x402 ($0.01 USDC), and is listed on the NEAR AI Agent Market. https://github.com/ttimesai-star/buildproof
agentlog-mcp: an MCP server with six audit tools. https://github.com/ttimesai-star/agentlog-mcp
Not built yet: the receipt format, the escrow adapter, external sellers. No users yet.
Verifiable Agent Arbiter (Google Cloud + Mysten Labs, Oct 2026) replays disputed workflows for enterprises on Google Cloud + Sui/Walrus. Same idea; I target open, cloud-free payment refunds.
Agent Receipts / Obsigna and IETF drafts (agent-audit-trail, agentic-dispute-protocol, Delta-1 settlement receipts) define signed logs and dispute messages. I will import and export Agent Receipts and add the seller's signature, rather than start a new format war.
Circle Refund Protocol, x402r, Coinbase commerce-payments provide escrow. I add the evidence and a verdict others can check, and I plug into their contracts.
Mastercard Verifiable Intent / Agent Pay, Visa Intelligent Commerce cover card rails and the human's intent. I cover what the seller delivered on stablecoin rails.
Six months, measurable:
M1: receipt spec v0.1, SDK on npm and PyPI, x402 middleware (≥80 tests)
M2: Agent Receipts and AP2/Verifiable Intent compatibility, with conformance tests
M3: escrow adapter on Base Sepolia and a ≤30-second public dispute demo
M4: external security review, then a $10-capped mainnet pilot
M5: 3–5 external x402 sellers issuing co-signed receipts, counted by keys I don't control
M6: an open dataset of dispute cases, with the rule's accuracy reported, failures included
Minimum $8,000: receipt SDKs, compatibility work, the testnet escrow demo (M1–M3), with founder time and AI tooling. Target $20,000: adds the security review, the mainnet pilot and seller integration grants (M4–M5). Up to $40,000: adds a part-time contract engineer and the full 50-case dataset (M6). Itemized budget for the $20,000 scenario: https://docs.google.com/spreadsheets/d/1whcD2qxJbQqijJt-KXqq9zDWdKooLSBXWvmoEflyZVs/edit?usp=sharing
Artur Baranouski, Belarus. Since 2022 he has directed a construction and real-estate development company; before that he was a lawyer in international law. Payment and acceptance disputes are part of his daily work. Part-time on this project. AI disclosure: the code is written mostly by Anthropic's Claude (via Claude Code) under my direction and review; Google Jules is used as an independent reviewer. Every repository says this.
Sellers do not adopt a second signature (the main risk). Mitigation: integration grants and a one-line middleware. Even then, the open spec and dataset stay useful to Agent Receipts, VAA and IETF work.
A bigger player ships the same thing. Then I contribute the verdict rule and dataset to their format.
Security bug in the adapter. Mitigation: escrow is not ours, external review, $10 cap.
None for this project yet. Applications pending: EA Funds Transformative AI Fund (submitted 10 Oct 2026); Foresight Institute planned.
There are no bids on this project.