emperor-mew/voidly-measurement-classifier-v1

Voidly Atlas Per-Measurement Classifier v1

Version: v1 | License: CC BY 4.0

XGBoost row-level censorship classifier inspired by Niaki et al. KDD23. Trained on 84K evidence rows, stratified 80/20 split. AUC 1.0 honest caveat: the model reconstructs the labeling rule from signal_value + source patterns rather than discovering novel signal. Lives at POST /v1/measurement/classify.

Eval

Metric	Value
held_out_auc	1
held_out_f1	1
held_out_precision	1
held_out_recall	1
min_source_auc	1
label_balance_overall	0.1965

Features

• signal_value
• signal_baseline
• risk_tier
• obs_hour
• obs_dow
• obs_is_weekend
• obs_month
• signal_deviation
• signal_ratio
• has_artifact_hash
• has_source_url
• has_domain
• has_asn
• asn_filled
• country_7d_rate
• asn_7d_rate
• country_domain_7d_rate
• source_code
• kind_code
• country_code_code
• continent_code
• un_subregion_code
• region_code_code
• domain_category_code

Honest caveats

• AUC 1.0 is NOT a generalization claim. The model reconstructs labeling rules from raw signal magnitudes — see top feature asn_7d_rate (81% gain).
• Use as an interface layer (per-row score for downstream UIs), not as a discovery model.
• v3.3 country-day classifier is the canonical research artifact; this is a row-level wrapper.

Citation

@misc{voidly_voidly_measurement_classifier_v1,
  title  = {Voidly Atlas: voidly-measurement-classifier-v1 (v1)},
  author = {Voidly},
  year   = {2026},
  url    = {https://huggingface.co/emperor-mew/voidly-measurement-classifier-v1},
  note   = {Open censorship-research ML stack. CC BY 4.0.}
}

Method foundation: Niaki et al. KDD23 — Massively Parallel Censorship Probing