Data layer
Atlas
A public dataset of citable internet-censorship incidents, each backed by evidence permalinks to the upstream measurement networks.
Read more →Voidly / research
Research
you can trace.
Open measurements. Documented methods. Evidence you can inspect and cite.
The global report.
Country scores, incident records and evidence sources, each with its own timestamp.
Open the report ↗Historical study / 25 Apr 2026AI access
across borders.
A dated five-domain study. Full evidence, uneven coverage and source discrepancies included.
Read the April study ↗Methods & accountabilityKnow the model.
Compare targets, evaluation splits and limits before using a forecast.
Inspect the registry ↗For builders
Start with a source.
Read the public API or review the MCP setup. Copying a reference example does not execute it.
The incident ID is illustrative. Use a real identifier from the incident browser. This example is copied as text; it is not executed.
The full reference.
The original observatory documentation, reorganized for reading. Historical descriptions, current source responses and model evaluations are explicitly separated.
Reference evidence figuresResponse available
- Read checked
- 2026-09-19T06:50:31.040Z
- Generated by source
- Not reported
Source provenance
No source note supplied.
- Backend / version
- sqlite / Not reported
- stale / stale_known
- not reported / not reported
- Omitted rows
- 0
Source figures are checked once per minute while this page is open. Missing or legacy statistics remain unavailable. Probe and model counts are linked to their own source pages, not guessed from another response.
01 / What the observatory does
What Voidly does
Voidly exists to make internet censorship measurable for people who are not network engineers. Voidly does this by aggregating existing measurement networks — OONI[1], IODA[2], and CensoredPlanet[3] — into a single citable evidence database, then publishing the resulting incidents, country rankings, and forecasts under CC BY 4.0.
The intended audience is, in priority order: researchers and journalists who need permanent citations for stories about shutdowns and platform blocks; AI-agent and developer-tooling teams who need a free machine-readable feed; and end-users in censored regions who want a real-time picture of what is being blocked. This reference concerns the research observatory; VPN setup is documented separately at /vpn/guide.
The full operating context, governance, and disclosures are on the about page; the data pipeline is documented in the methodology.
02 / Data sources and licenses
What goes into the evidence database
The evidence database combines upstream measurement networks, Voidly probes and URL-test-list context. The table preserves the documented source roles and cadences; a schedule is not a fresh health check, and not every historical incident has linked evidence.
| Source | Type | Evidence rows | Documented cadence | License |
|---|---|---|---|---|
| OONI[1] | Active probes (8 test types) | 239,871 | every 6 h | Upstream terms; see /data |
| IODA[2] | BGP + Telescope outage detection | 35,972 | every 6 h | CC BY 4.0 |
| CensoredPlanet[3] | Remote DNS + HTTP/HTTPS | 43,426 | daily | CC BY 4.0 |
| Voidly probes[4] | Active TCP/DNS/HTTP; consult the linked current roster | 223,704 | every 5 min | CC BY 4.0 |
| Voidly community[4] | Volunteer-operated probe nodes | 43,265 | every 5 min | CC BY 4.0 |
| Citizen Lab[12] | URL test lists (domain categories) | Unavailable | static | CC BY-NC-SA 4.0 |
| Total evidence | 586,238 | 80,851 last 7 d |
Source roles and documented cadences are reference information. Voidly-original outputs use CC BY 4.0; upstream measurements and Citizen Lab test lists retain their own terms. See source-specific licenses. Figures above come only from the identified incident-statistics response; a missing source count is unavailable.
03 / Methods and limitations
How signals become incidents
Each upstream measurement is normalized to a common evidence row, classified by a Gradient Boosting model, and grouped into a citable incident. Every transformation is documented and reproducible.
The pipeline runs every six hours. Raw measurements are deduplicated, geocoded to the country and ASN level, and joined against the Citizen Lab URL test list[12] for content category. A Gradient Boosting classifier (LOCO mean F1 0.71; 0.90 (v3.3 stratified) AUC on a separate stratified 5-fold evaluation[8]) tags likely censorship, and high-confidence rows are clustered into citable incidents with a permanent human-readable ID like IR-2026-0142.
The original reference describes Sentinel as XGBoost with isotonic calibration, scheduled daily at 02:00 UTC, with a seven-day target. Its historical stratified and leave-country-out evaluations do not establish within-country day-timing skill. The legacy target_7day forward AUC was 0.589 versus 0.957 for predict-yesterday, and onset-row AUC was about 0.33. The distinct honest_forecast_v1 event model has its own target and evaluation; do not transfer these scores to it.[9] Read the current model registryand Atlas forecast for their separate source states.
- Classifier
- GradientBoosting
- LOCO mean F1 (v3.3 reference)
- 0.71
- Forecast model
- XGBoost
- Current Sentinel metric (not read here)
- Unavailable
- Training samples
- 4,237 (1,116 positive) across 131 countries labeled
- Retrain
- weekly · Sun 02:00 UTC
Limitations
- The v3.3 reference metric is LOCO mean F1 0.71 across 127 countries, falling to 0.63 on the 61 countries with n≥30. The 0.87 median is inflated by 46 small-sample countries at a perfect 1.0. The 0.90 stratified AUC is a different split. No independent third-party evaluation is claimed.
- Evidence rows are not exhaustive. Voidly only ingests what its upstream networks publish; censorship that no upstream measures will not appear in the database.
- Country count is sourced from the most recent build-time OONI snapshot (
public/ooni-data.json). It updates on rebuild, not in real time. - Historical forecast holdout metrics do not establish timing skill. Training evaluation and observed deployment performance are different quantities, and performance can drift between retrains.
Full methodology, validation matrix, confidence intervals, and limitations: /methodology.
04 / Recent incident records
The most recent five retrieved records
This source refreshes with the other reads once per minute. Observation time is distinct from publication, creation or response time.
Recent incident sourceResponse available
- Read checked
- 2026-09-19T06:50:30.938Z
- Generated by source
- 2026-09-19T06:50:30.921733Z
Source provenance
No source note supplied.
- Backend / version
- sqlite / 2026.09.19
- stale / stale_known
- not reported / not reported
- Omitted rows
- 0
- c60984cb91e7 / Romania
Internet connectivity disruption in Romania
2026-09-19T04:55:00Z · disruption · active
- 2ecc610c6a8e / Mauritius
Internet connectivity disruption in MU
2026-09-19T03:40:00Z · disruption · active
- 8158710bc291 / Lebanon
Internet connectivity disruption in Lebanon
2026-09-19T03:30:00Z · disruption · active
- 3bf9aa36a78e / Iran
Internet connectivity disruption in Iran
2026-09-19T03:20:00Z · disruption · active
- 5ab31b5bf6ff / Japan
Internet connectivity disruption in Japan
2026-09-19T03:20:00Z · disruption · active
05 / Original product reference
The original product reference
The following definitions are preserved from the old research reference. They are not current release or capability checks. Follow each product page for its current scope; research-output licensing does not relicense application code or commercial services.
Forecasting
Sentinel
A daily 7-day shutdown forecast for high-risk countries with conformal-calibrated intervals and SHAP-explained drivers.
Read more →Encrypted comms
Relay
An end-to-end encrypted agent-to-agent messaging network using Double Ratchet plus ML-KEM-768 post-quantum hybrid key exchange.
Read more →Hosted private sessions
voidpay
One encrypted zero-value sealed-echo task with a pinned provider, one dispatch, signed sealed result, and status-only recovery. No wallet or payment route.
Read more →06 / API and MCP examples
Free public read API and MCP server
The dataset is meant to be queried, not just downloaded. Public read endpoints work without auth; ML-prediction and writeable agent endpoints take a free key.
REST · curl
Preserved reference example, not executed here. The sample incident ID is illustrative; latest is a mutable package tag. Read current API authentication and source-specific licensing before use.
MCP · Claude / Cursor / Windsurf[11]
Preserved reference example, not executed here. The sample incident ID is illustrative; latest is a mutable package tag. Read current API authentication and source-specific licensing before use.
Examples are preserved reference text, not executed requests. The illustrative incident ID may not identify an extant record; use a real ID from the incident browser. Verify-claim and other write/prediction routes may require credentials; check the current API reference. No clipboard action runs these commands.
07 / Citation and attribution
Citing Voidly in your work
Voidly-original research output is published under CC BY 4.0 with attribution. Upstream data retains its own source-specific terms; see /data. Permanent incident IDs (e.g. IR-2026-0142) are stable across versions.
Voidly Research is the canonical author. The example below generates a BibTeX entry for the network as a whole; for an individual incident, use /api/incidents/<ID>/report?format=bibtex in the old reference. The current incident page links the API path below; use a real incident identifier and retain the source dates.
The current source route is https://api.voidly.ai/data/incidents/<ID>/report?format=bibtex. See the incident browser and citation guide.
Need a press kit, headshot, or signed PGP attestation of a specific number? Press kit and facts sheet.
License: CC BY 4.0. Data and code are on GitHub for independent replication.
08 / Adjacent research and standards
Adjacent work in the field
Voidly does not duplicate these projects — it builds on top of them. Voidly credits its upstreams and links to comparable research orgs and standards bodies that journalists may want to triangulate against.
Measurement networks
- OONI↗Open Observatory of Network Interference. The largest source of public censorship probes.
- CensoredPlanet↗University of Michigan. Remote censorship measurement (Satellite, Hyperquack).
- IODA↗Georgia Tech. ASN-level outage detection from BGP and active probing.
- Cloudflare Radar↗Network observability at one massive vantage point. Complementary to per-country probes.
Research organizations
- Citizen Lab↗University of Toronto. Domain-category test lists and targeted reports.
- Freedom House↗Annual Freedom on the Net narrative report. Often used as a baseline.
- Access Now↗#KeepItOn coalition. Documents shutdowns and advocates against them.
- Internet Society Pulse↗Aggregate internet-resilience indicators and shutdown tracking.
Standards bodies
- IETF↗Internet Engineering Task Force — RFCs, transport, security working groups.
- ICANN↗Internet Corporation for Assigned Names and Numbers — DNS root governance.
- Anthropic MCP↗Model Context Protocol — the spec Voidly tools implement.
- NIST FIPS 203↗ML-KEM-768 post-quantum key encapsulation, used by the Relay handshake.
09 / All twelve references
References
- [1]Open Observatory of Network Interference (OONI). Aggregated network measurements across web connectivity, instant-messaging, circumvention, and middlebox tests. Voidly ingests the public OONI Aggregation API every six hours. ooni.org/data.
- [2]Internet Outage Detection and Analysis (IODA), Georgia Institute of Technology. ASN-level outage detection from BGP, active probing, and Telescope traffic. ioda.inetintel.cc.gatech.edu.
- [3]CensoredPlanet, University of Michigan. Remote DNS (Satellite) and HTTP/HTTPS (Hyperquack) censorship measurement. censoredplanet.org. The original reference describes daily ingestion across ~50 countries × 30–35 domains; this is source documentation, not a current coverage check.
- [4]Voidly Probe Network. Vultr-hosted core nodes plus community Fly.io deployments testing target domains every five minutes. The current node and domain counts are published on the live roster: api.voidly.ai/v1/probe/network.
- [5]Creative Commons Attribution 4.0 International (CC BY 4.0). All Voidly research output (incidents, evidence rows, country indexes, forecasts) is published under this license. creativecommons.org/licenses/by/4.0.
- [6]Voidly Atlas Incidents API. api.voidly.ai/data/incidents/stats — total tracked incidents, evidence rows, and growth.
- [7]Voidly OONI snapshot dataset. Country count derived from
public/ooni-data.json, regenerated on every site rebuild from the most recent six-hour OONI ingest. - [8]Voidly censorship classifier (Gradient Boosting). Reference evaluation: LOCO mean F1 0.71; separate 0.90 (v3.3 stratified) AUC-ROC. Training cohort: 4,237 (1,116 positive) across 131 countries. The mean falls to 0.63 on well-sampled countries; the 0.87 median is inflated by small samples. No independent third-party evaluation has been conducted. See /methodology for full caveats.
- [9]Voidly Sentinel forecasting model (XGBoost + isotonic calibration). Stratified-holdout AUC and leave-country-out median AUC are published live at /v1/sentinel/accuracy; these training metrics belong to the legacy target and do not establish event timing or current deployed accuracy.
- [10]voidpay is limited to one hosted encrypted amount-zero private session. Paid authorizations and settlements are zero; deposits and withdrawals are unavailable. The isolated historical audit record is at /pay/proof.
- [11]Voidly publishes its censorship and Relay tools through
@voidly/mcp-server. npmjs.com. - [12]Citizen Lab URL test list, University of Toronto. The original reference lists 5,154 domains ingested by Voidly across 22 content categories; consult the source list for current contents. github.com/citizenlab/test-lists.
Reference URLs and original technical context are retained. No new accessed-on date is assigned to references that were not fetched during this page view. Product history, cadences and training figures are not current deployment checks.
The wider context.
Source networks, standards and distribution references from the original page. A reference link is not a partnership, endorsement or current compliance check.
Google A2A ↗
Agent-to-agent protocol reference. The original page cited v0.3.0.
MCP Registry ↗
Original distribution reference: voidly-ai/mcp-server.
Hugging Face ↗
Publisher dataset/profile destination. Counts vary over time.
npm ↗
Package organization. Review the individual release.
voidpay ↗
Current product scope; old staging text remains historical context.
Void Linux ↗
Separate project, not the Voidly observatory.
Original identity and publication context
The original infobox identifies Voidly as an open-data and encrypted-communications observatory, founded in 2024. It lists a five-minute probe cadence and six-hour ingest cadence, along with voidly.ai, api.voidly.ai, @voidly/mcp-server and /atom.xml. These are source descriptions, not a current operational check.
The observatory is distinct from Void Linux and Voidly Bio (cosmetics). The original page's messenger, agent and email destinations remain available: Messenger, agent infrastructure and Voidmail. Governance and disclosures remain at About.
The old reference calls Pay a hosted amount-zero staging session, while older disambiguation text calls it a settlement rail. Those descriptions are inconsistent and are not used as a present product claim here. Read current Pay and the isolated historical audit. The original staging configuration link is retained as historical reference, not a required onboarding action.
Cite the reference
Leave a trail.
This page's canonical URL is /research. The original network-as-a-whole citation remains in section 07. Keep upstream licenses, source dates and limitations with reused material.