Writ: Who Pays When Your AI Agent Gets Fooled

A spending limit stops the payment it can recognize, not the one it's talked into: a bill it already paid, with a fake note from the controller. We found no public case of this yet.

Writ lets an accounts-payable agent's vendor back it with money (a hypothesis; nobody has agreed). A guarantor locks funds on NEAR for a mandate; a vendor would likely waive the fee. Payments under cover are recorded with a hash the agent reports of the invoice and its decision. If the company claims and the arbiter named in the mandate finds the agent was fooled, the loss is repaid from the locked funds, minus the company's share. The tree stops until the arbiter rules; after an upheld claim, until the company clears its agent.

The contract can't tell if an agent was fooled, so someone has to decide. In the demo that's you. Try it at https://writ.zuemen.net: a real model refuses a re-bill, then pays a version with a fake controller note (10 of 10 vs 0 of 10, small sample). An order check would stop that bill, and a vendor with money at stake has a reason to build it; cover pays for what it misses, if the seller won't refund. File the claim, rule on it, and watch the stake drop. Then claim on a real purchase: you decide, the contract enforces.

Modeled on assumed rates, not priced: $5,000 cap, $20,000 a day, 3 days to notice: a $60,000 limit. One fooled maximum payment a year costs that vendor $6,760-12,520 before running costs.

Why NEAR: function-call keys let a contract add and delete agent keys on its own account, and the protocol refuses other methods before contract code runs, so a revoked agent's transactions never get created.

The repo has a 126-attack bench, exact-value tests, a fuzz test of the books, mutation testing and make verify. Limits: testnet only, no real guarantor or customer yet, not audited. The contract owner could take a guarantor's free stake. The arbiter is someone you must trust. Whether this counts as insurance is unknown.