01 / DETECT
A score you can trust.
Gradient-boosted trees on causal card history and shared-device signals, calibrated so 0.10 really means about 10%.
LightGBM · isotonicCounterfactual payment risk engine
Not no.
yes, if.
Risk engines stop at a score. yes, if finds the smallest change that turns a risky payment into an approval (pay part now, pay a little later, or use a different card), then learns which friction is cheapest to apply. It plugs straight into PayPal checkout.
Loading…Live model
—
Fraud probability—
Decision
—PR-AUC, calibrated detector (logistic baseline —)
—of fraud caught at a 1% false-positive rate
—attack success before → after adversarial retraining
—lower cost than the best static policy after attackers adapt
From no to the cheapest yes.
Four layers on every payment. The detector says how risky it is; the rest decides what to do about it.
02 / EXPLAIN
Reasons, not a black box.
Every decision lists the signals that pushed it up or down: a new device, an odd hour, a device shared by many cards.
path contributions03 / YES, IF
The smallest change.
Searches 83 changes a buyer can actually make (part-payment, delay, debit or credit) and returns the cheapest ones that cross the approval line.
counterfactual search04 / LEARN
Cheapest friction wins.
A policy that learns how well step-up, hold and block really work from their outcomes, and shifts when attackers start beating one-time passcodes.
Thompson samplingMeasured, not promised.
Every number on this page is produced by one reproducible script and served from the same model the API runs.
Experiment E5 · policy under drift
When attackers learn to pass step-up
Cumulative cost (fraud loss + friction + lost customers) over the test stream. Halfway through, step-up stops catching most fraud.
Experiment E6 · adversary loop
Harder to fool each round
Share of flagged fraud a black-box attacker gets approved with 50 queries; line is clean PR-AUC.
Read this before quoting the numbers
- The model is trained on a synthetic, reproducible transaction generator in the IEEE-CIS column layout (account takeover, card testing, device rings). The IEEE-CIS pipeline is included for anyone with Kaggle access.
- Step-up, hold and block outcomes are simulated with stated, configurable assumptions: no public dataset records them. A real rollout starts in shadow mode, then an A/B test.
- Counterfactuals describe the model's decision boundary ("the model would accept this if…"), not causal truth. Buyers only see coarse guidance; exact options are for analysts.
Every decision is a PayPal call.
Orders are created with intent AUTHORIZE, so the engine decides what happens to the money after the buyer approves.
| Decision | What happens | PayPal Orders & Payments API |
|---|---|---|
| ALLOW | Approved straight away | Authorize, then capture in full |
| STEP UP | Buyer confirms a one-time code | Authorize; capture once verified |
| HOLD | An analyst reviews it | Authorization kept; capture or void after review |
| BLOCK | Declined, unless the buyer takes a "yes, if" | No order is created |
| yes, if pay part now | Part of the amount now, the rest after verification | Partial capture with final_capture: false |
| yes, if later | The payment goes through a few hours later | Authorization kept, captured on schedule |
Start free. Upgrade when it pays.
The engine and console are open to everyone today. Premium is in early access.
Free
$0forever
- Decisions API with buyer and analyst views
- Reasons and "yes, if" options per payment
- 60 analyst decisions a minute
- Console and PayPal sandbox checkout
PremiumEarly access
$10/ month
- Everything in Free
- API keys and higher rate limits
- Your own friction costs and thresholds
- Live PayPal checkout integration
- Monthly policy and drift reports
Planned launch pricing in USD. No card needed to join the early-access list.
Early access