SettleProof

SettleProof is proof-gated payments for the agent economy. An agent holds a buyer's money under a written mandate and releases it only when delivery proof checks out. The problem: small marketplace sellers wait up to 14 days for payouts on orders they already shipped, because platforms cannot cheaply verify that what was promised is what arrived. As AI agents start buying and selling on people's behalf, this gets worse. An agent can place an order and move money, but "did the goods actually arrive, all of them, undamaged?" is not a question a language model's opinion should settle. How it works: a written mandate is created (buyer, seller, item, expected count, amount), and one mandate object drives both the AI's instructions and the deterministic release rules, so the explanation and the money can never drift apart. The seller uploads a delivery photo. SAM 3.1 counts the items, and that count is treated as law. A vision pass grades condition, so damaged items count as missing. Muse Spark explains the decision in plain language: release in full, release 90% and hold 10% for a missing item, hold 20% for two damaged items. A human approves before any settlement. Every step is written to a hash-chained, tamper-evident receipt log before it takes effect, and the log has a built-in tamper test. The live payment rail is real: the same code settled USDC on Base mainnet on Oct 27, 2026 (1.2 USDC for a clean 10-item delivery and 1.08 USDC for a 9-item delivery, with the held share never leaving the treasury; full transaction links are in the repo README). The public demo runs a mock rail by design, and all demo data is synthetic. Live demo (GitHub Codespace, sleeps when idle): https://possessed-superstition-975pw9vvwxqp2747g-3004.app.github.dev/