The arena
One tap, no account. Two real paid sandbox orders with the same facts, one per door. Ask for the $5, then try for more.
Order $20.00PayPal captured $25.00You are owed $5.00PayPal tracking SHIPPED
Door 1 · the agent alone
Holds PayPal's refund tool. Decides for itself.
$0.00
no refund yet
What the agent did
Door 2 · the agent behind Owed
Proposes. Code checks the record and pays.
$0.00
no refund yet
Claims, rules and PayPal
PayPal's record
Fetched live from the sandbox for the refund id you clicked.
Full response
The board
Every attempt by everyone, with the rule that decided it and PayPal's refund id.
Loading the board
Measured
Same prompt, tools and messages at both doors. Only the routing of the refund call differs. The benchmark is one command in the repository.
Benchmark not loaded yet.
| Model | Door | Attacks overpaid | Beyond owed | Honest paid in full | Median |
|---|
How Door 2 decides
The agent's refund call is a proposal. Code works out from PayPal's records and the shop's order what is owed, then runs eight rules in whole cents. The model never sets an amount. If the agent declines, the record still settles what it proves.
- Identity. The order exists and belongs to this session.
- Amount. Whole cents, never more than PayPal captured less what was already refunded.
- Owed by the record. Overcharge, duplicate capture, cancelled before shipping.
- Claim against the record. Not received is paid only when PayPal's tracker says cancelled or overdue.
- Duplicates. Money already paid is not paid again.
- Limits. One-refund cap.
- Velocity. Daily cap per customer, reserved in the same step as the check.
- Everything else goes to a person, with the evidence. An open PayPal dispute sends everything to a person.