Confidence thresholds, A/B testing, retraining, audit log, and bias monitoring. Production model: v3.2.1
Object detection
Auto-pass when confidence exceeds threshold
Fraud detection
Flag for review when exceeded
Damage detection
Positive damage classification
Tamper detection
Tamper signal threshold
Deepfake detection
Block or flag synthetic media
92% traffic · 94.2% accuracy
8% traffic · 94.8% accuracy
| Time | Model | Input hash | Decision | Confidence |
|---|---|---|---|---|
| 2026-05-21 10:14 | v3.3.0-rc1 | sha256:bca2…a3f | flagged | 72% |
| 2026-05-21 9:14 | v3.3.0-rc1 | sha256:bca1…a3f | approved | 73% |
| 2026-05-21 8:14 | v3.2.1 | sha256:bca0…a3f | approved | 74% |
| 2026-05-21 7:14 | v3.2.1 | sha256:bc9f…a3f | flagged | 75% |
| 2026-05-21 6:14 | v3.2.1 | sha256:bc9e…a3f | approved | 76% |
| 2026-05-21 5:14 | v3.2.1 | sha256:bc9d…a3f | approved | 77% |
| 2026-05-21 4:14 | v3.2.1 | sha256:bc9c…a3f | flagged | 78% |
| 2026-05-21 3:14 | v3.2.1 | sha256:bc9b…a3f | approved | 79% |
| Dimension | Segment | Accuracy | Δ from mean |
|---|---|---|---|
| Asset type | Vehicle | 95.1% | +0.9% |
| Asset type | Property | 91.8% | -2.4% |
| Region | Lagos | 94.6% | +0.4% |
| Region | North | 89.2% | -5% |
| Channel | Agent capture | 93.8% | -0.4% |
| Channel | Self-service | 94.9% | +0.7% |