Pick a real test-set payment or build your own. The same request returns a coarse buyer view and a full analyst view.
Choose a payment to see the decision.
Move only what a buyer can change: how much is paid now, when, and with which card. Everything else (device, history, graph signals) stays fixed.
Approval line: —. Scores are the model's calibrated fraud probability; counterfactuals describe the model, not causal truth.
The engine scores the buyer before PayPal opens, then runs its plan on the authorization: capture, verify, review, or a "yes, if". PayPal sandbox only: use a sandbox personal account, no real money moves.
Loading…
PayPal or card, the order is only authorized: the engine's plan decides what gets captured.
Every policy replays the same test stream. Halfway through, attackers learn to pass one-time passcodes and step-up's catch rate drops from 85% to 40%.
A black-box attacker controls amount (keeping at least half), timing (up to 24h) and funding source. Each round, its successful attacks are added to training.
Perturbations happen in feature space, which is a simplification; real attackers face stricter constraints. Exact scores and counterfactuals are analyst-only and rate-limited for this reason.
Shared-entity features count how many distinct cards have touched a device, email domain or billing region so far, causally, in time order. This is the ring behind the "ring" sample.
Detector quality, calibration, the value of graph features, counterfactual quality and a fairness check, all on the held-out, later-in-time test split.