Teach the model with feedback
Feedback is free, and it is the best way to improve accuracy for your own traffic.
- Later, when you know the truth, send the address (or the
check_id) and what happened:
{ "email": "maria.santos@gmail.com", "outcome": "confirmed" }
- Outcomes:
outcome | Use when |
|---|---|
confirmed | The user verified their email |
real | The user logged in, paid or was clearly a real person |
fake | Fake account, bot or abuse of a free trial |
abuse | Spam, fraud or terms violation |
bounced | Email to this address bounced |
Rules:
- Each check can be labelled once, within 14 days, by the account that made the check. By email, it labels your latest check of that inbox.
- Your personal model starts after about 50 of your own outcomes, but only once it beats the hand-tuned rules on your own traffic.
GET /v1/modelshows"serving": "learned"when it is active for you. - Focus on checks with
"feedback_wanted": true: those teach the model the most. - Signals shared across customers need agreement from several customers, so no single account can poison them.