Risk & Prediction
Honest validation for the 90% prediction target. Scores are measured against known outcomes on held-out cases — a model's confidence is never reported as accuracy.
All metrics below are computed on synthetic sample data and are not a performance claim. The prediction target must be agreed with the organisers before pilot training; if labels are insufficient, this remains a research track.
Target definition
Probability that a case exceeds 60 days from registration to chargesheet. Horizon: 60 days from registration.
Validation set
Time split: trained on 50 cases before 2025-01-01, tested on later cases. Prevalence 78%.
Classification quality
Accuracy 78% vs baseline 78%
Probability quality
Lower is better; see calibration chart.
Confusion matrix
True positives
7
False positives
2
False negatives (missed)
0
True negatives
0
Baseline (historical rate 64%) predicts the majority class for every case.
Calibration (predicted vs observed)
Performance by district
| District | n | Accuracy |
|---|---|---|
| Krishna | 7 | 71% |
| Guntur | 2 | n too small |
Fairness check: performance monitored across districts where sample size permits.
Open cases — predicted delay risk (explainable)
| Case | Probability | Uncertainty | Top factors | Recommended review action |
|---|---|---|---|---|
| C-2025-H144 | 100% | 98%–100% | Offence: Ganja (NDPS lab & procedure); District workload: Krishna (Vijayawada); Offence: transport (simpler evidence) | Supervisor review of pending lab reports & milestones |
| C-2025-H115 | 100% | 98%–100% | Offence: Ganja (NDPS lab & procedure); District workload: Krishna (Vijayawada); Offence: transport (simpler evidence) | Supervisor review of pending lab reports & milestones |
| C-2025-H126 | 88% | 80%–97% | District workload: Krishna (Vijayawada); Offence: ID liquor (lab report dependency) | Supervisor review of pending lab reports & milestones |
| C-2025-H129 | 88% | 80%–97% | District workload: Krishna (Vijayawada); Offence: ID liquor (lab report dependency) | Supervisor review of pending lab reports & milestones |
| C-2025-0042 | 88% | 80%–97% | District workload: Krishna (Vijayawada); Offence: ID liquor (lab report dependency) | Supervisor review of pending lab reports & milestones |
| C-2026-H157 | 81% | 70%–92% | District workload: Krishna (Vijayawada) | Supervisor review of pending lab reports & milestones |
| C-2025-0310 | 59% | 45%–72% | District workload: Krishna (Vijayawada); Offence: transport (simpler evidence) | Routine monitoring |
| C-2026-H160 | 43% | 30%–57% | base rate | Routine monitoring |
| C-2026-0021 | 21% | 10%–32% | Offence: transport (simpler evidence) | Routine monitoring |
A score is not evidence of guilt and must not be the sole basis for any enforcement decision. Model logit-baseline-0.3 (synthetic).