Fraud Detection
HealthyReal-time production health
Known immediately — no outcomes needed.
Predictions / hour
Latency
Inputs, context & predictions
Input drift vs reference (PSI)
Prediction distribution
Routing
- Backend
- Managed by Zero Grad
- Context
- pinned to v7
- Fallback on timeout
- Managed C · 0.02% of requests
Matured outcome quality
Performance metrics are computed only for predictions with observed outcomes. Chargebacks and confirmed fraud reviews take up to ~30 days to arrive, so recent predictions aren't scored yet.
Business impact
Aug 1 – Aug 31 · 72% coverageUses your configured value per case. Grows as the remaining 28% of August outcomes arrive.
Model quality
Aug 1 – Aug 31AUC — matured outcomes
By prediction week. Recent cohorts are provisional until their outcomes arrive.
Outcomes
Ground truth flows back by prediction_id. It improves evaluation and routing — never model weights.
Send outcomes
Every prediction returns a prediction_id. Send the real outcome whenever you learn it.
# Later, when the real outcome is known
zg.capability("fraud-detection").send_outcomes([
{"prediction_id": "prd_01J9X4K2M8", "outcome": "fraud",
"observed_at": "2026-09-30T23:41:00Z"},
])Outcomes never update model weights — Zero Grad does not train a customer-specific model. They are used to measure real production quality, compare foundation models and context-selection strategies, improve routing, and decide whether a challenger beats the current production configuration.
What Zero Grad learned from your outcomes
No training doesn't mean no learning.
- Context strategy: recent-weighted examples beat uniform sampling by +.006 AUC on matured outcomes — now the default for this capability.
- Routing: TabFM wins on transactions over $1,000. A routing rule for that slice is queued for shadow evaluation.
- Calibration: matured outcomes moved the 1% FPR operating point from 0.71 to 0.69.
- v8 (new foundation model release) beats v7 on the August cohort: +.009 AUC at 72% outcome coverage.
What adapts: which foundation model serves which slice, how in-context examples are selected, and calibration. What doesn't: model weights.