Help the model learn
The Voidly Sentinel classifier's lowest-confidence forecasts. The model is most uncertain about these 13 predictions — human labels here have the highest information value for the next retrain. Submit observed outcomes via the auth-gated /v1/sentinel/report_miss endpoint.
Method: Calibrated uncertainty sampling — arXiv 2510.03162 (2025) · Updated every 15 min · Raw JSON
How to submit a label
- Get an API key — admin via Voidly, or subscriber key via /alerts.
- POST to
/v1/sentinel/report_misswith:
curl -X POST https://api.voidly.ai/v1/sentinel/report_miss \
-H "X-Voidly-Subscriber-Key: $YOUR_KEY" \
-H "Content-Type: application/json" \
-d '{
"country_code": "IR",
"what_happened": "Observed shutdown 2026-05-19 14:00 UTC, BGP withdrawal of AS31549. Sourced from NetBlocks twitter + my SSH probe in Tehran.",
"source_url": "https://twitter.com/netblocks/status/..."
}'Rate limit: 10 / hour / key. The label enters error_queue on the Vultr ML server and is incorporated by the next nightly retrain. See /methodology for the gate logic that decides whether new models replace the serving one.
13 uncertain forecasts (lowest distance from 0.5 first)
- VenezuelaVEMay 20, 2026not yet resolvedProbability33.5%Distance from 0.516.5 ppPred above 0.5?↑ yesnear_threshold · not_yet_resolved
- EgyptEGMay 20, 2026not yet resolvedProbability33.0%Distance from 0.517.0 ppPred above 0.5?↑ yesnear_threshold · not_yet_resolved
- VenezuelaVEMay 18, 2026not yet resolvedProbability33.0%Distance from 0.517.0 ppPred above 0.5?↑ yesnear_threshold · not_yet_resolved
- VenezuelaVEMay 19, 2026not yet resolvedProbability32.7%Distance from 0.517.3 ppPred above 0.5?↑ yesnear_threshold · not_yet_resolved
- EgyptEGMay 18, 2026not yet resolvedProbability32.7%Distance from 0.517.3 ppPred above 0.5?↑ yesnear_threshold · not_yet_resolved
- EgyptEGMay 19, 2026not yet resolvedProbability32.1%Distance from 0.517.9 ppPred above 0.5?↑ yesnear_threshold · not_yet_resolved
- VietnamVNMay 21, 2026not yet resolvedProbability68.2%Distance from 0.518.2 ppPred above 0.5?↑ yesnear_threshold · not_yet_resolved
- BangladeshBDMay 21, 2026not yet resolvedProbability68.2%Distance from 0.518.2 ppPred above 0.5?↑ yesnear_threshold · not_yet_resolved
- MYMYMay 21, 2026not yet resolvedProbability68.9%Distance from 0.518.9 ppPred above 0.5?↑ yesnear_threshold · not_yet_resolved
- ThailandTHMay 21, 2026not yet resolvedProbability69.4%Distance from 0.519.4 ppPred above 0.5?↑ yesnear_threshold · not_yet_resolved
- Saudi ArabiaSAMay 21, 2026not yet resolvedProbability69.8%Distance from 0.519.8 ppPred above 0.5?↑ yesnear_threshold · not_yet_resolved
- BrazilBRMay 21, 2026not yet resolvedProbability69.8%Distance from 0.519.8 ppPred above 0.5?↑ yesnear_threshold · not_yet_resolved
- PhilippinesPHMay 21, 2026not yet resolvedProbability70.3%Distance from 0.520.3 ppPred above 0.5?↑ yesnot_yet_resolved
Why active learning matters
Most ML systems retrain on random labeled samples. Active learning literature shows that labeling the lowest-confidence predictions instead gives the biggest F1 gain per labeling-budget unit.
Voidly's weekly classifier retrain has been rejecting new model candidates since 2026-05-03 because they F1-regress on a dedup'd evaluation set. The gate is working as designed, but without fresh labels we can't distinguish “real regression” from “data drift the model needs to learn.”
This page closes that loop. Journalists + censorship experts see the model's most uncertain forecasts, submit observed outcomes, and the next retrain gate uses their labels for evaluation. Slow but durable.
References
- Calibrated Uncertainty Sampling for Active Learning (arXiv 2510.03162, 2025)
- RiskID: Active learning for network traffic
- /sentinel/backtest — honest model reliability
- /atlas/models — full model registry
- /methodology — retrain gate logic