Tuning from
your labels.
Find thresholds that catch more confirmed fraud within your false-positive budget. Nothing changes until you apply.
What a suggestion contains
Illustrative exampleOwner review
- InputYour last 30 days of fraud and legit labels.
- BudgetAt most 0.5% of labelled legit events flagged.
- EvidencePer-reason precision and recall. Before/after confusion matrices.
- Your decisionApply bot and repeat/limit thresholds, or leave policy unchanged.
How it works
Label real outcomes
Report confirmed Fraud or Legit in the dashboard or through
/v1/feedback. Start with at least 20 of each.Choose your budget
Pick 0.1%, 0.5%, 1% or 2%. A deterministic search maximizes recall within that observed false-positive rate.
Review, then apply
Inspect signal counts and projected mistakes. Only an explicit owner action changes thresholds, with an audit entry.
Three separate risk scores
Verification adds three 0–100 axes: automation, repeat and abuse. You can see which kind of risk fired without exposing reason ids to callers. The existing score stays unchanged.
"scores": { "automation": 25, "repeat": 50, "abuse": 0 }Limits
- Labels reflect what you reviewed, which may not represent all your traffic. Include random legitimate samples.
- Small samples can overfit. Twenty labels per class is only a floor; a low observed false-positive rate is not a guarantee.
- Projections cover score gates only. Owner rules, login challenges and verified-agent policy are excluded, and no new signal weights or statistical model are trained.
- The default 0.5% budget is a logical starting point and no measured optimum. Re-run monthly and review the change before applying.