Counterparty screening
How P2P ads with risky terms are kept out of your routes.
P2P ads carry free-text terms. Some are harmless ("pay in your own name"). Others are classic fraud setups: move the chat to WhatsApp, pay a different name, mark as paid early. In the Nigerian market, about one in five top ads with terms asked for an off-platform contact.
Before a route uses an ad, Jev, a decision model from TypeSafe, reads its terms in any language and returns two probabilities:
| Score | Question | Excluded at |
|---|---|---|
redFlag | Does it bypass the venue's escrow or match a known scam pattern? | ≥ 0.60 |
restrictive | Would it slow or block a standard payment from an own-name account? | ≥ 0.70 |
The model scores; code decides the cut-off. Each ad's scores show up in the route's meta.screened and in GET /currencies/{code}/ads.
Verdicts are cached per ad and per version of its terms, so an ad is re-screened only when its terms change. Only the six best-priced usable ads per leg are screened on each quote, which keeps it fast. Binance hides ad terms in its public search, so screening applies to the ads that publish them.