Sources and methodology
Last updated: 12 November 2026
Where WarIntel's reports come from, what the software does with them today, and what the numbers can and cannot tell you.
1Sources
Each report links to the source it came from. A source being listed is not an endorsement: many sources are parties to the conflict, and their claims are shown as claims.
- Public Telegram channels of news outlets, officials and armed groups in the region, read automatically.
- News outlets' RSS feeds.
- GDELT, an open database of events machine-coded from world news coverage.
- The ACLED and UCDP conflict datasets, when their access keys are configured on the server.
- Ship positions from AIS receivers (AISstream, Data Docked) and aircraft positions from the ADS-B network adsb.lol.
- Official advisories and public datasets, for example casualty figures published by health ministries, always labelled with their publisher.
2Methodology: what the software does today
Everything in this section is automatic and describes the code as it runs today.
- Collection: each news RSS feed is read on its own timer of 5 to 60 minutes. Public Telegram channels are read every 2 to 10 minutes. GDELT is queried every 5 minutes (news search) and every 15 minutes (event export). ACLED is read hourly and UCDP daily, only when their access keys are set.
- Filtering: a news article or Telegram post is stored only if it mentions a conflict term and a place in the covered regions, is not on a block list of unrelated topics (for example road accidents or natural disasters), names a known place or exactly one country in a covered theatre, and has a usable headline (for news articles, at least 4 real words). News articles older than 7 days are skipped.
- De-duplication: each item is stored once. A news article is identified by its web address (host and path; query strings are ignored), a Telegram post by its channel and message number, and a GDELT record by its GDELT id or by its normalised headline, day and place. Items already stored are skipped. Reports of the same incident from different sources are not merged: they stay separate and are linked as one story (see the next section).
- Map pins: on the map, reports of the same type at the same point (coordinates rounded to 2 decimal places) within the same 2-hour block are drawn as one pin. Reports placed only at country level are never combined.
- Translation: Arabic headlines are translated into English by an AI model, with a simpler built-in translation as a fallback. The original stays linked. Hebrew and Persian headlines are not translated automatically.
- Category: assigned by fixed whole-word rules, not by an AI model. The headline decides; the text is read only when the headline gives no clue. The first matching rule wins, for example "airstrike" or "drone strike" gives Airstrikes, "tanker", "ship" or "Hormuz" gives Naval operations, "ceasefire" or "talks" gives Diplomacy. A headline about someone warning, threatening or asking about strikes is a Statement, not a strike. A report that matches no rule is filed under Political.
- Severity: also set by rules. CRITICAL: 10 or more killed, 50 or more wounded, a massacre or mass-casualty attack, an attack on a nuclear facility, or a chemical attack. HIGH: at least 1 killed or 5 wounded, or an attack that hits a strategic target such as an airport, port, refinery, power plant or military base. MEDIUM: any other attack, or any wounded. LOW: everything else. A casualty figure above 500 is treated as a running total, not one incident, and is not used. On screen the four levels are shown as FLASH, IMMEDIATE, PRIORITY and ROUTINE.
- Launch and target points: for missile and drone reports, an AI model is asked for launch and target points (used for trajectories on the map) only when the rules are not confident and the report is at most 6 hours old.
- Placement: place names are matched against a built-in list of known places; no external geocoding service is used. When a report names only one country, it is stored at country level, without an exact point.
- Publication: a stored report appears on the map, in the feeds, in alerts and in the API. The delay is mostly the polling interval of its source.
3Corroboration and the Verified label
When a news or Telegram report is stored, it is compared with reports from other sources up to 6 hours before or after it. It is linked to the most similar one when both are in the same country (or, without a country, within 25 km of each other) and their headlines share at least half of their significant words, with at least 4 significant words each. Linked reports form one story; nothing is merged or deleted.
A story is marked Verified when it has reports from 2 or more different sources. Sources are counted per feed, so a publisher with both a Telegram channel and a news feed counts as two. GDELT records never count. A GDELT record matched by an ACLED record, and every record from the UCDP research dataset, are also marked Verified. No other rule sets the label: an official statement on its own does not. All other reports are shown as unverified.
The report page shows how many outlets reported the story and lists them. A high count makes a report more likely to be accurate but does not prove it.
This automatic rule is the verification standard WarIntel uses. No person checks a report before it is marked Verified.
4Counts and statistics
Headline numbers (for example missiles or strikes) are sums over reports. They depend on what sources report and on automatic extraction, so they are estimates, not official tallies. Where a figure comes from an official body, that body is named.
Collection has gaps when a source goes silent or a feed breaks, so a quiet period in a chart can mean missing data rather than no activity. No data was collected from 19 May to 7 November 2026 because of an ingestion outage; charts and counts that span it say so.
5Editorial policy and corrections
No person reviews reports before they appear. WarIntel collects, sorts and publishes them automatically, by the rules on this page.
If a report is wrong, write to support@war-intel.com with a link to it. We check the original source, then correct the report or remove it, and reply to you. Corrected reports keep a link to their original source.
6What we publish
On the site, in the RSS feed and in the API, WarIntel publishes its own structured summary of each report (time, place, category, precedence, verification, corroboration) and a link to the original. Outlets' headlines and article text belong to them; follow the link to read the original. See the API terms.
7Known limits
- Coverage is uneven: some areas and parties publish far more than others.
- Automatic translation, classification and geolocation make mistakes.
- Machine-coded GDELT records are less precise than reports from a named outlet.
- Do not use WarIntel for decisions about anyone's physical safety; follow official guidance from local authorities.