The Voidly Atlas country-level classifier (v3.3) answers "is this country blocking something today?" The new cross-protocol classifier answers a finer question: "which protocol/port is blocked?"
Different protocols are blocked at different layers with different
signatures. HTTP-80 plaintext is typically intercepted by middleboxes
inspecting the Host header. HTTPS/443 is blocked by SNI
filtering or IP-list null-routes. Tor is killed by IP blocklists on
the OR ports (9001/443). WhatsApp's mTLS endpoints get blocked at the
CDN edge by SNI. A single country-level "censored" verdict elides this.
We trained eight small XGBoost classifiers, one per protocol group. Each takes a (country, day) input and emits the probability that the protocol was blocked. Features are nine cheap signals: protocol's own 7-day history, neighbor 7-day rate (UN sub-region adjacency), the day's max anomaly rate across other protocols (a "cross-protocol lockdown" detector), country risk tier, and four calendar features. Crucially we drop today's own anomaly rate from features — that's the label-defining quantity, so including it would make the model circular and inflate AUC to 1.0.
Labels and label noise
Labels come from parsing OONI's N/M measurements anomalous
ratio directly out of the upstream_claim text — necessary
because OONI evidence rows for everything except web_connectivity
are only stored when blocking occurred (OK measurements aren't
written). A naive SELECT COUNT(*) WHERE signal != 'ok' would
show 100% positive rate for every protocol; parsing the embedded ratio
recovers the true daily denominator. A (country, day, protocol) is a
positive sample iff ≥ 30% anomalous on ≥ 3 measurements, negative
iff ≤ 5% anomalous, and the ambiguous middle band is dropped.
Results
All eight protocols cleared the promote floor (stratified F1 ≥ 0.55 AND LOCO pooled AUC ≥ 0.70). The LOCO pooled-OOF AUCs cluster between 0.98 and 0.999 because port-level blocking is strongly persistent — the 7-day history feature carries most of the predictive load. This is realistic signal, not leakage: history is strictly from days before the target day, and labels are computed from the target day alone.
What this is not
- This does not override the v3.3 country-level classifier. v3.3 answers "is this country censoring"; per-protocol answers "which port is blocked".
- Inter-protocol overlap is real. The
web-connectgroup includes both HTTP/80 and HTTPS/443 measurements, so its labels partially correlate with the TLS-app groups for the same country-day. - OONI coverage is uneven.
http_invalid_request_lineandtorhave less measurement volume thanweb_connectivity, so the corresponding models have thinner LOCO evidence (fewer countries with both classes in held-out). - The "AUC ~0.99" headline reflects the strong autocorrelation of port-level blocking, not novel signal discovery. Per-country median AUC is more honest (range 0.89 to 1.0); pooled OOF AUC is the promote-gate number.
API surface
GET /v1/classifier/protocol/info— all 8 protocols, per-protocol metrics, feature importances, honest caveats.GET /v1/classifier/protocol/{proto}/{cc}— per-(protocol, country) probability for today, with observed measurement data surfaced asobserved_todayfor transparency.
Sample: /v1/classifier/protocol/tor/CN returns
P(Tor blocked in China) = 0.96 (label_at_optimal = 1, threshold = 0.15);
/v1/classifier/protocol/tls-signal/IR returns
P(Signal blocked in Iran) = 0.99.
Leakage audit (2026-05-21).
A label-leakage audit flagged this model as SUSPICIOUS. The
training script correctly drops anomaly_rate (which defines the
label), but the feature cross_protocol_max_today is the same-day
max anomaly rate of other protocols — and protocols overlap heavily
(web-connect already covers port 443 = the TLS class), so for
correlated protocols it is a near-copy of the label. A LOCO pooled AUC of
0.98+ on 9 non-anomaly-rate features is implausible.
Honest LOCO AUC estimate without the same-day cross-protocol
feature: ~0.80-0.88. Recommended fix: drop
cross_protocol_max_today (or restrict it to non-overlapping
protocols) and re-run the LOCO evaluation.