Voidly Atlas 7-Day Shutdown Forecast v1.1
Version: v1.1 | Trained: 2026-05-21T04:20:38Z | License: CC BY 4.0
XGBoost + isotonic-calibration forecast for 7-day censorship risk.
Retrained 2026-05-21 after IODA disruption labels were excluded from
target_7day — earlier model had been promoting raw IODA outages (94% of April incidents) as confirmed censorship.
Eval
| Metric | Value |
|---|---|
roc_auc |
0.9541 |
f1 |
0.6667 |
precision |
0.7255 |
recall |
0.6167 |
optimal_threshold |
0.2700 |
positive_rate |
0.0920 |
samples |
2148 |
Dual-holdout gate accepted (legacy holdout +18pp, temporal holdout +58pp).
Honest caveats
- Operational threshold is 0.05, not the F1-optimal 0.27 — calibrated for high recall, accepts false positives.
- Live production track record: precision 0.69 / recall 0.39 / Brier 0.56 / calibration drift +56pp over 720 predictions (model under-predicts long-tail positives).
- Adaptive Conformal Inference adjusts α online (currently 0.21, started at 0.10) — empirical coverage 91.3% over 840 observations.
Reproducibility
python3 scripts/build-forecast-features.py # excludes incident_type='disruption'
python3 scripts/test-forecast-labels-sane.py # weekly regression gate
Citation
@misc{voidly_voidly_forecast_v1_7day,
title = {Voidly Atlas: voidly-forecast-v1-7day (v1.1)},
author = {Voidly},
year = {2026},
url = {https://huggingface.co/emperor-mew/voidly-forecast-v1-7day},
note = {Open censorship-research ML stack. CC BY 4.0.}
}