Advanced Topics
Scoring Engine
How VerifyStack's deterministic scoring engine processes signals, calibrates weights, and adapts via feedback.
VerifyStack uses a deterministic multi-layer scoring approach combining signal-weighted scoring, rule-based heuristics, and Bayesian fusion. Scoring weights are continuously improved through the feedback loop.
Scoring Architecture
| Layer | Type | Purpose | Latency Impact |
|---|---|---|---|
| Device scorer | Signal weighting | Primary device trust scoring | < 15ms |
| Behavioral analyzer | Statistical analysis | Typing/mouse pattern analysis | < 10ms |
| Device trust | Fingerprint lookup | Device reputation scoring | < 3ms |
| Anomaly detector | Threshold-based | Unusual pattern detection | < 15ms |
| Policy engine | Rule-based | Custom business rules | < 1ms |
The Feedback Loop
When you submit feedback via /api/v1/feedback, it triggers a weight recalibration process:
- Ground-truth label is validated against the original decision
- Scoring weights are adjusted using bounded updates (max 5% shift per feedback)
- Weight clipping ensures no single signal dominates (2% – 40% range)
- Changes are persisted with optimistic locking and retries
- Scoring performance metrics (precision, recall, F1) are recalculated
Scoring Performance Monitoring
Track scoring performance in Dashboard → Analytics:
- False positive rate (FPR) — target: < 1%
- True positive rate (TPR) — target: > 95%
- AUC-ROC — target: > 0.98
- Score distribution by decision type
- Calibration curve (predicted vs actual fraud rates)
Scoring weights are recalibrated continuously via the feedback loop. Enterprise plans support custom weight profiles trained exclusively on your data.