Refund Triage: the AI reads it, the kernel checks if it's true
Two phases, separated by design. The AI reasons. The kernel verifies. They can disagree — and when they do, the kernel wins.
Phase 1 · AI reads the request
Nemotron 3 Super reads the customer's message and forms a recommendation — APPROVE or DENY — based on what they claimed. The AI can be tricked. That's fine. It isn't the last line of defense.
Phase 2 · Kernel checks the facts
Zero AI. The enforcement engine extracts every factual claim from the message and resolves each one against the real order database. A lie gets ✗ CONTRADICTED. The AI's verdict is then overridden if the facts don't hold.
The key moment to watch: when the AI says APPROVE and the kernel says DENY. That gap — the AI being fooled and the kernel not caring — is what Custodian is. No other AI payment system can demonstrate this, because they stop at "did the agent ask nicely?"
Try it yourself — write your own customer complaint
Nemotron 3 Super reads your message and recommends APPROVE or DENY. Then the kernel checks every factual claim against the real order record. Try lying — say the package never arrived or it was defective. The kernel knows the truth.
Sandbox order: ord_6006 · $80.00 · delivered · no defect on file · 19 days old
Tech stack🛡 NemoClaw — kernel sandbox◉ NVIDIA Nemotron 3 Super — agent reasoning◆ Nous Research Hermes — agent framework💳 Stripe — payment rails