Settleye

Your AI agent told the customer the refund was processed. The ticket closed. The ledger says $0.00 moved. Settleye gives an AI support agent a new capability: it can verify whether its own promised action actually settled against the ledger, and fix the shortfall with a signed receipt. That is the gap I built Settleye to close. Agentic support is moving from suggestions to actions, but the industry still grades the conversation, not the settlement. Guardrails check words. QA samples tone. Dispute vendors arrive after the chargeback. Nobody joins the promise to the payment per order. Settleye is the execution auditor for AI support agents. It reads the transcript, extracts every binding promise, reconciles it against Stripe or Shopify ledger state, and puts the gap in an exception inbox with the dollars at risk. In the demo, one click fixes the shortfall against the mock ledger with a durable idempotency key. A retry returns the same receipt and moves $0. A concurrent duplicate cannot double-refund. Then it proves what happened. Every fix can emit an execution receipt binding the transcript digest, promise, ledger observation, remediation key, and result, signed by Settleye and countersigned by the merchant. The browser verifies both signatures. Change one digit and the receipt fails. That is proof of execution, not a screenshot. Evidence: 12 of 12 eval verdicts, 11 of 12 exact entity matches, and zero of four non-binding promises drifted. On the live, read-only Shopify dev-store proof, test-mode order #1001 carried $44.00 at risk and was caught in 4.07 seconds. No money moved. The repo ships with 235 backend checks and 93 frontend checks. The public sample is a no-install, test-ledger demo. I claim no real merchant data or live payments. The HTTP fix path is mock-ledger only; Stripe and Shopify write adapters are tested but not wired into the route. Resolved is not settled.