Varde Research

Method andcorrections

How Varde Monitor counts, ranks and writes, what it does not know, and what it has had to correct.

How incidents are counted

Incidents come from the GDELT 2.0 event stream, which reads news reports from around the world every fifteen minutes and codes what they describe, an air strike, an armed clash, an abduction, with a place and a date. We keep only violent event types, group reports of the same event on the same day and place into one incident, and show an incident only where the UCDP register lists an active armed conflict in that country. Every incident keeps a link to each report behind it. The map is updated several times a day.

An incident is a report of an event, not a verified event. The counts on the map are counts of what was reported, not of casualties.

One source and several sources

An incident carried by a single outlet is drawn as a hollow ring. One carried by two or more independent outlets is filled. A single report can be right, but it is more often a misreading: a protest coded as a clash, an old event retold, a place guessed from a headline. So single-source incidents are shown, but never counted in the ranking. In countries the register lists mainly as parties to wars abroad, such as France, Belgium, the United States and the United Kingdom, only incidents with two or more outlets appear at all.

How crises are ranked

The list ranks crises by incidents confirmed by several sources, weighted by the severity of the event type, in the chosen window. It does not rank by raw counts: raw counts follow media attention, and English-language coverage makes some countries look far more violent than others where far more people are dying. Beside the counts, each crisis shows UCDP's own count of battle-related deaths with its year, as a measure of a different kind.

What the AI does, and does not do

Each crisis carries three to five machine-written sentences on what is happening. A language model writes them from the headlines already linked on the crisis page, preferring established agencies and newsrooms, plus UCDP and World Bank figures with their years. The English is written first; the Norwegian is translated from the approved English in a separate step.

The model does not find, count or place incidents, decide which conflicts exist, rank crises or choose sources. It does not invent figures, names or events: if it does, the text is refused (below). The text is marked as machine-written, and the sources it was written from are listed under it.

How the fact guard works

Every summary is checked in code before it is shown, and refused if it fails any test:

  • every figure, and every number written as a word ("twelve", "dozens"), must be in the sources, as the same figure or one of that size;
  • every name must be in the sources;
  • a named person must be stated as killed or wounded as the sources say, never "killed or injured";
  • no counts of incidents or percentage changes, which the page states itself, and no ranges such as "40-41";
  • a UCDP or World Bank figure must give its year;
  • no sentences about outlets or method, and no commentary ("highlighting", "marks a new phase");
  • the Norwegian must contain exactly the English text's numbers.

A refused draft is asked for once more with the reason. If it fails again, the page shows the status line alone. If only the Norwegian fails, the English is shown with a note saying so. The guard cannot check that a sentence is true, only that what it states is in the sources. When something wrong gets through, it is corrected and logged below.

What Verified by Varde means

An incident drawn in magenta has been checked by a Varde Research analyst against a source we name, with a short note on what was confirmed. It is the only colour on the map that means a person has looked.

Data sources and licences

GDELT 2.0
The event stream behind every incident. Free and open, with attribution. Terms
UCDP, Uppsala University
A data source, not a partner: the register of armed conflicts that decides where incidents are shown, and battle-related deaths. CC BY 4.0. Data and licence
World Bank
Country figures, including refugees (from UNHCR) and internally displaced people (from IDMC). CC BY 4.0. Terms
EIA
Oil flows through chokepoints and pipeline capacities in the trade and energy layer. US government work. World Oil Transit Chokepoints
News
Headlines from Google News, GDELT's article index and a list of curated feeds, always linked to the publisher. Outlets are named, never shown as logos.
Wikipedia and Wikimedia Commons
Background text (CC BY-SA 4.0) and pictures, each with its author and licence.
Natural Earth
Borders and coastlines. Public domain.
Our own data
The incidents, rankings and downloads as we publish them. CC BY 4.0.

Known weaknesses

  • Double counting. GDELT can code the same event more than once when reports differ on the date or the place. Grouping by day and place removes most, not all.
  • Country-level places. When a report names only a country, the incident sits at the country's centre. Such points say little about where the fighting is.
  • Uneven coverage. Media attention varies widely between countries. Wars covered mainly in Arabic, French or local languages are undercounted next to those covered in English.
  • Coding errors. Protests, police operations and old events are sometimes coded as armed incidents. Single-source incidents are the most affected.
  • The register lags. UCDP updates once a year. A new conflict reaches the map through our own detection and a human decision, not automatically.

Corrections