Most censorship-research platforms force a journalist to manually query a country page, then a service page, then cross-reference probe rows by hand. The Voidly Atlas auto-fact-check service inverts that flow: a journalist or researcher types a natural-language claim ("Twitter is blocked in Iran") and gets a verdict, a confidence score, and five evidence permalinks in under a second.

The endpoint POST /v1/atlas/fact-check ships today. Behind the scenes: a regex-based claim parser maps country names (122 aliases including "iran", "persia", "PRC", "ksa", "Türkiye") and service names (36 aliases including the OONI test-name conventions for service-level probes) to canonical (country_code, domain_patterns, ooni_test_names) tuples; the evidence table is then queried with strict suffix-matching on domain (so "twitter.com" matches "twitter.com" and "abc.twitter.com" but never "chatgpt.com") plus OONI's upstream_claim LIKE "OONI tor:%" for service-level tests plus CensoredPlanet's source_url LIKE "%domain=twitter.com%".

Verdict logic: confirmed_block requires at least 3 block-style observations across at least 2 independent sources (OONI + CensoredPlanet, or OONI + Voidly probes, or any combination). partial_block is anything with at least one block observation that falls short of the confirmation threshold. not_observed is zero block-style rows in the window — and crucially the response flags that this is not the same as "not blocked" because it can also mean the probe network has no recent coverage for that (country, domain) pair. contradicted fires when the claim asserts blocking but the evidence is unambiguously normal/accessible, or when the claim asserts unrestricted access but the evidence shows blocks. insufficient_data is reserved for unparseable claims (no country recognized).

Confidence is a heuristic blend of block-row count, source diversity, and the average block signal value. The formula caps at 0.99 and is deliberately bounded below 1.0 — every response reminds the caller in the honest_caveats field that parsing is not LLM-based and that absence of evidence is not evidence of absence.

20-claim accuracy test, run end-to-end against the public api.voidly.ai endpoint: 19/20 correct (95%). Median latency 178ms, p95 990ms, against the intelligence.voidly.ai upstream directly: median 7ms, p95 17ms. The 18ms median upstream measurement is the real bound — the public-URL number absorbs Cloudflare Worker proxy overhead and TLS termination.

The one "failure" is informative rather than a bug: the test expected "WhatsApp is blocked in France" to return not_observed (France being a non-blocker), but the evidence table actually contains 6 Voidly community probe rows that flag WhatsApp anomalies in France. The verdict was partial_block with confidence 0.80 — the service is being honest about the data, not the test's prior. The appropriate journalist action when seeing this verdict is to click the single source, see that all 6 rows are from one community probe in one ASN over a short window, and conclude "this is a real anomaly worth investigating, not country-wide blocking." That's the workflow the endpoint is designed to enable.

Honest caveats baked into every response:

Claim parsing is regex + alias map, not LLM-based. Misspellings, unusual phrasings, or non-English claims will fail silently. We surface what was matched (matched_country_alias, matched_service_alias) so the caller can verify.
not_observed can mean "service is accessible" or "we have no probe coverage for this (country, domain)" — these are fundamentally different states and the endpoint flags this ambiguity inline.
Single-source corroboration triggers an explicit caveat. The 3-row 2-source confirmation floor is conservative on purpose; below it we degrade to partial_block with a lower confidence and tell the caller why.
For CN/RU/IR/KP, even confirmed blocks may miss DPI-based intermittent throttling that synthetic probes cannot see.

Endpoints:

POST /v1/atlas/fact-check — body {"claim": "...", "time_range_days": N} (default 365, max 365)
GET /v1/atlas/fact-check?claim=...&days=N — same thing for easy testing in a browser or shell
GET /v1/atlas/fact-check/info — methodology, thresholds, supported alias counts, honest-caveat list

Sample (curl):

curl -X POST https://api.voidly.ai/v1/atlas/fact-check \
  -H "Content-Type: application/json" \
  -d '{"claim":"Telegram is blocked in Iran"}'

Returns:

{
  "verdict": "confirmed_block",
  "confidence": 0.99,
  "evidence_count": 135,
  "corroboration_sources": ["censoredplanet", "ooni"],
  "last_observed_at": "2026-05-20T18:00:00Z",
  "top_evidence": [
    "https://explorer.ooni.org/search?probe_cc=IR&test_name=telegram&...",
    ...
  ],
  "signal_types": ["blocking", "dns-blocking", "http-blocking-tcp-reset"],
  "avg_block_signal": 0.73,
  "honest_caveats": [...],
  "latency_ms": 13
}

Per-category breakdown of the 20-claim test: confirmed-block category 10/10 (100%), ambiguous category 5/5 (100%), false-claim category 4/5 (80% — the one "failure" is the honest-data case described above). All ten heavy-censorship claims (Telegram/Signal/WhatsApp in Iran, YouTube/Facebook/WhatsApp/Tor in China, Instagram in Russia, Twitter in Turkey, plus the headline IR/Twitter) returned partial_block or confirmed_block. All four "is X blocked in Germany/USA/Canada/Japan" claims returned not_observed as expected.

This is a building block for downstream systems: an LLM agent can verify its own claims before publishing, a journalist's chatbot can fact-check user-typed assertions in real time, and a research-paper auto-generation pipeline can refuse to cite a country-service pair when the verdict comes back insufficient_data or not_observed. Free tier; cached for 5 minutes on the Worker edge.