emperor-mew/voidly-classifier-v3.3

Voidly Atlas Classifier v3.3

Version: v3.3 | Trained: 2026-05-21T03:01:46.793987+00:00 | License: CC BY 4.0

Country-day censorship classifier with regime-similarity-weighted geographic contagion features. Promoted 2026-05-21.

Intended use

Given a (country, day) feature vector — measurement volume, anomaly rate, probe agreement, and three contagion neighbour aggregates — return a calibrated P(censorship event). Surfaced live at GET https://api.voidly.ai/v1/classifier/score/{cc}.

Evaluation

Metric	Value
stratified_f1	0.7289
stratified_auc	0.8991
loco_median_f1	0.8696
loco_mean_f1	0.7109
loco_n_countries	127
n_features	16
n_samples	4237
n_positive	1116
n_countries	131

Stratified split = single random 80/20. LOCO = leave-one-country-out across 127 countries, the honest cross-country generalization number.

Training data

• Total samples: 4,237 country-day rows

• Positives: 1,116 (26.3%)

• Unique countries: 131

• Provenance: OONI + CensoredPlanet + IODA + Voidly probe network (84K evidence rows)

• Labels exclude IODA disruption incidents (fix 2026-05-21) — those are real network outages but not all are censorship

Features

16 inputs: 13 base + 3 regime-similarity-weighted contagion neighbours.

• anomaly_rate
• measurement_count
• spike_magnitude
• day_of_week
• month
• is_weekend
• rate_count_interaction
• probe_block_rate
• probe_node_count
• probe_avg_confidence
• probe_agreement
• rate_spike_interaction
• high_evidence
• neighbor_block_rate_7d
• neighbor_incident_count_7d
• neighbor_max_anomaly_7d

Top features by gain

• anomaly_rate (importance 0.2210)
• month (importance 0.2022)
• measurement_count (importance 0.1696)
• neighbor_max_anomaly_7d (importance 0.0887)
• neighbor_incident_count_7d (importance 0.0770)
• neighbor_block_rate_7d (importance 0.0764)
• rate_count_interaction (importance 0.0561)
• day_of_week (importance 0.0389)
• spike_magnitude (importance 0.0354)
• rate_spike_interaction (importance 0.0319)
• is_weekend (importance 0.0013)
• probe_avg_confidence (importance 0.0009)
• high_evidence (importance 0.0003)
• probe_block_rate (importance 0.0002)
• probe_node_count (importance 0.0001)
• probe_agreement (importance 0.0000)

Honest caveats

• LOCO mean F1 0.71 is the honest generalization number — the median 0.87 hides a long tail.
• 16 MENA + post-Soviet countries regress 5–29pp (OM, UZ, TN, LY, YE, JO, MA, …) due to sparse neighbor-pair overlap.
• v3.4 regime-cluster finetune was a negative result — coefficient analysis showed the stack head ignored the cluster heads. Root cause is noise-bounded F1 in countries with 5–15 positive samples; the real fix is targeted labeling, not architecture.
• EG (Egypt) recovered from a v3.2 regression (F1 0.55 → 0.73).
• Predicted probabilities are NOT isotonic-calibrated; the API surfaces per-country thresholds.

Reproducibility

# Build script (training data → fitted pickle + per-country thresholds)

python3 scripts/build-classifier-v3.3-regime-weighted.py

• Algorithm: GradientBoostingClassifier (sklearn)

• Promoted artifact: /opt/voidly-ai/models/censorship_classifier_v3_promoted.pkl

• Backup pre-promote: .pkl.bak.v3.1-2026-05-21

• Per-country thresholds: ml-deploy/classifier_v3.3_per_country_thresholds.json

Citation

@misc{voidly_voidly_classifier_v3.3,
  title  = {Voidly Atlas: voidly-classifier-v3.3 (v3.3)},
  author = {Voidly},
  year   = {2026},
  url    = {https://huggingface.co/emperor-mew/voidly-classifier-v3.3},
  note   = {Open censorship-research ML stack. CC BY 4.0.}
}