What this is
Single-country anomaly detectors (DBSCAN, STL) catch local censorship events. They miss the more interesting kind: cross-country synchronized events — e.g. every Gulf state blocking the same app on the same day, or post-Soviet states tightening together around a regional flashpoint.
The multi-country anomaly burst detector
(scripts/build-anomaly-bursts.py) scans 90 days of
DBSCAN flips, groups them into single-day buckets, and surfaces
days where K≥3 distinct countries flipped together as a
candidate coordinated event. Each burst is then characterised by
its modal shared domain, blocking method, and signal type from the
evidence underlying the flip — producing a hypothesized
common factor like shared_domain:chat.openai.com or
shared_signal_type:outage.
What the first run found
90-day scan, 3,718 (country, day) cells scored against the live DBSCAN model, 820 anomalous flips, 73 bursts with K≥3 countries. 33 are statistically significant at p_adj < 0.05 under Bonferroni correction over the 91 days scanned.
The largest burst — 2026-05-03 with 58 countries
co-flagged — was driven by OONI's globally-
coordinated probe sweep of chat.openai.com on that
date, not by coordinated censorship. This is the exact failure
mode the brief warned about: co-occurrence is not
coordination. A simultaneous burst can be a coincidence,
a shared upstream outage, a probe-network methodology change, or
real cross-state coordination — the detector can only flag
the candidate, not adjudicate the cause.
More interesting candidate bursts include the
2026-04-12–14 sequence flagging protonvpn.com
blocking spreading across 23–26 countries (which is
consistent with a real anti-circumvention pattern), and recurring
twitter.com (24 countries on 2026-04-21, 22 on 2026-04-24) and
facebook.com (31 countries on 2026-04-13, 22 on 2026-04-16) bursts
that warrant editor follow-up before any “coordinated
censorship” framing.
How it works
- For each (country, day) cell with ≥5 measurements over the
90-day lookback, score with the same rolling-window DBSCAN that
powers
/v1/anomaly/dbscan/{cc}(window 45d, eps=75th-pct kNN, min_samples=3). - Group flip events by day. Flag days where K≥3 distinct countries flipped.
- For each burst, mine the evidence table on those (cc, day) cells to find the modal blocking method, signal type, and affected domain across all members — this is the hypothesized shared factor.
- p-value: under independence per-country flip rates,
p_event = product(member_rates),p_any = 1 − (1 − p_event)^N_days,p_adj = min(1, p_any × N_days)Bonferroni. - Cap output to top 50 by significance. Significant-first sort, then by p_adj asc, then by k_countries desc.
Endpoints
GET /v1/atlas/anomaly-bursts?limit=20&significant=true&min_k=5— list recent burstsGET /v1/atlas/anomaly-bursts/{burst_id}— full details: per-country anomaly scores, per-country base rates, full shared-signal/method/domain distributions
Honest caveats
- Co-occurrence ≠ causation ≠ coordination. Any burst could be coincidence, shared infra failure, OONI methodology change, holiday traffic shift, or a fiber cut that happens to bisect a region. The detector flags candidates — a human editor must adjudicate.
- 6-hour bucketing collapsed to day. The brief
asked for 6-hour windows; upstream evidence
observed_atis day-granular, so buckets are day-level in practice. Finer bucketing would require per-test-time data we don't aggregate. - Independence assumption is wrong. Censorship- adjacent neighbors correlate. Bonferroni p_adj is a conservative bound, not a causal-inference test.
- Underlying DBSCAN is noisy. AUC vs labeled incidents is 0.65 — just above the promote floor. Bursts should be treated as exploration hooks, not alerts.
- OONI methodology bursts are a known false positive. Synchronised probe-sweep additions look identical to coordinated censorship at this aggregation level. Cross-check the burst date against OONI's probe release log before drawing causal conclusions.
Why this matters
Most published censorship-detection work focuses on single-country single-event detection — the DBSCAN baseline already covers that. The burst detector adds an explicit cross-country axis that's rarely surfaced in real-time: candidate coordinated campaigns ranked by statistical significance, with their hypothesized common factor pre-computed from raw evidence. Even if the headline number (33 significant bursts in 90 days) is half false-positives, the other half are real shared events worth investigating — and the pipeline runs daily, automatically, without an analyst needing to remember to look.