Voidly
All findingsPublished research

Per-category censorship classifiers (NEWS / ANON / GRP / PORN / COMT / MMED / SRCH)

7 specialized XGBoost classifiers exposing WHAT a country is targeting, not just whether. NEWS, ANON (Tor/VPN/Lantern), GRP, PORN, COMT, MMED, SRCH all promoted via the alt-AUC path (stratified AUC 0.913-0.977, LOCO median AUC 0.849-0.957). 5 of 7 primary categories cleared (needed 4), so family…

Read the full published summary

7 specialized XGBoost classifiers exposing WHAT a country is targeting, not just whether. NEWS, ANON (Tor/VPN/Lantern), GRP, PORN, COMT, MMED, SRCH all promoted via the alt-AUC path (stratified AUC 0.913-0.977, LOCO median AUC 0.849-0.957). 5 of 7 primary categories cleared (needed 4), so family promote passes. POLR (political opposition) and RELR (religion) honest-skipped due to zero coverage in evidence. Labels back-filled via a 37-domain hand-curated Citizen Lab category map because only 18 distinct domains in evidence have domain_category set. Endpoints return available=false with transparent reasons for skipped categories. v3.3 stays as the production default; this is additive transparency for journalists asking "is this regime suppressing journalism, circumvention tools, or messaging?"

Per-category censorship classifiers (NEWS / ANON / GRP / PORN / COMT / MMED / SRCH)

v3.3 answers is this country-day censoring something? The per-method classifier (also shipped 2026-05-21) answers HOW are they blocking it — DNS, TCP, HTTP, or TLS? This finding answers the journalist-gold question that sits between them: WHAT are they targeting?

The distinction matters editorially. A regime that blocks just ANON (Tor / VPNs / Lantern / Psiphon) is suppressing circumvention — that's a tell of internal political pressure, not free-press repression. A regime that blocks just NEWS (BBC / NYT / Guardian / Substack) is muzzling journalism. A regime that blocks SRCH (Google / DuckDuckGo) is taking a much heavier hand. The 7 promoted per-category classifiers expose those signals at the API.

Approach

Each classifier uses the same 16-feature v3.3 input (anomaly_rate, measurement_count, spike_magnitude, weekday/month/weekend, probe-block-rate, neighbor-contagion features, etc.) trained with class-balanced XGBoost. The category label is positive iff the country-day had ≥1 evidence row whose blocked domain matches that Citizen Lab category. Only ~18 distinct domains in the evidence table have domain_category populated, so we back-fill via a 37-domain hand-curated map covering NEWS (bbc, nytimes, guardian, ...), ANON (tor, protonvpn, psiphon, lantern, ...), COMT (signal, telegram, whatsapp, ...), and the rest. Negatives are category-agnostic (same negative row used for every category) so we don't leak "wrong-category" signal into the negative pool.

Results: 7 of 7 categories with data promoted

CategoryPos labelsStrat AUCLOCO median AUCPromote path
NEWS1760.9420.897alt-AUC
ANON1940.9210.866alt-AUC
GRP2870.9130.858alt-AUC
PORN850.9770.957alt-AUC
COMT2320.9330.856alt-AUC
MMED2300.9350.849alt-AUC (bonus)
SRCH1200.9710.924alt-AUC (bonus)
POLR0——SKIP
RELR0——SKIP

5 of 7 primary categories cleared (needed 4) — family promote passes. POLR (political opposition) and RELR (religion) are honestly skipped because zero rows in our evidence table currently match those Citizen Lab categories. The endpoint returns available: false with a transparent reason rather than a hallucinated probability.

Top countries per category (training-time)

  • NEWS: RU (0.378 mean prob), BD, QA, PK, VE, IN — Russia leads journalism-targeting
  • ANON: AZ (0.508) — Azerbaijan top for VPN/Tor blocking
  • GRP: EG (0.498) — Egypt top for social-organizing blocks
  • PORN: QA (0.198) — Qatar leads adult-content blocking
  • COMT: PK (0.587) — Pakistan top for messaging-tool blocks
  • MMED: EG (0.469) — Egypt top for YouTube/Instagram/Tumblr
  • SRCH: QA (0.233) — Qatar top for search-engine blocking

Promote gate

Floors are deliberately softer than v3.3's gate because per-category labels are ~5–7% positive rate (vs v3.3's 26%), so F1 at threshold 0.5 punishes any high-precision-low-recall model. We use either:

  • Classic path: stratified F1 ≥ 0.55 AND LOCO median F1 ≥ 0.60, OR
  • Alt-AUC path: stratified AUC ≥ 0.75 AND LOCO median AUC ≥ 0.70 AND threshold-optimal F1 ≥ 0.40

All 7 promoted categories cleared the alt-AUC path with margin (lowest LOCO AUC was COMT at 0.856).

Honest caveats

  • Citizen Lab categories cover ~14K domains; only 18 distinct domains in our evidence have the column populated. The 37-domain hand-curated back-fill is editorial — we cite the source.
  • Categories aren't mutually exclusive (BBC is NEWS but used heavily for COMT via live chat). Each model is independent; a country can be flagged for multiple categories on the same day.
  • Per-country coverage is uneven. LOCO AUC is the honest generalization signal.
  • The probability is NOT a counterfactual ("would they block this without X?") — it's a learned association between country-day anomaly shape and category-positive labels.
  • v3.3 generic classifier remains the production default. This is an additive transparency layer, not a replacement.
  • POLR and RELR are honest-skipped — the endpoint flags them explicitly. We will retrain when evidence coverage expands.

Endpoints

  • GET /v1/classifier/category/info — per-category metrics, feature importance, honest caveats
  • GET /v1/classifier/category/{category}/{cc} — per-category probability for that country today
  • GET /v1/classifier/category/leaderboard?category=NEWS — training-time top countries for the category

The source record.

Original source links and labels. Linked responses may have changed since publication.