Sign in Request access

Every military exercise in the Baltic Sea region since 2015, and where each number comes from.

Exercitus is a registry for researchers. Each exercise has its dates, place, size and participants, and every value names its source, how precise it is, and whether official and press figures agree.

Request access Sign in For research. A person reviews each request.
Release
0.5.3.1 α, 28 Oct 2026
Exercises
1,637
Years
2015 to 2026
Read
news articles
Next
0.6, completeness
Status
α · field accuracy 85.5% (sample of 0.4.5)

Exercises per year, comparable core

The comparable core of release 0.5.3.1: 822 of 1,637 exercises that are multinational or have 1,000 personnel or more, one rule for both sides. 2026 runs to 28 October 2026.

One record, with its working shown

A real entry from release 0.5.3.1. Russia and Belarus declared 12,700 troops for Zapad 2017; other estimates ran to 100,000. The registry does not pick one. It keeps both and says where each came from.

Russia and Belarus

Zapad 2017

DisputedMachine-read

ProvenanceEvery value says where it comes from: official, press only, or disputed when the two disagree.

Dates
14 Sept to 20 Sept 2017
Official D1 · exact day

PrecisionD1 to D4 for dates: here the exact day is known.

Where
Belarus
Official L3 · country

Coarse looks coarseL1 to L4 for places. Known only to the country, so the map draws it with a dashed mark.

Personnel
Declared12,700Low12,700Best12,700High100,000
Declared by the organiser on 217 official pages; low, best and high come from 2,227 news sources

A range, not one numberLow, best and high, the way UCDP publishes conflict deaths. This is the disagreement.

Nations
Russia, Belarus
Official

ParticipantsA country counts only when at least 2 sources and 3% of all sources name it.

How far to trust it

Measured against exercises whose dates and size are known from official sources. The method in the registry explains each check and its limits.

20of 23

known major exercises of the period are in the registry

1,515

exercises with figures or dates from the organiser's own pages, shown beside the press

100%

of 98 randomly drawn records checked by a person were real exercises; the method gives the accuracy of every field

50%

placed only at country level so far; more training areas come with review

How to get the data

Access is personal and for research. It stays closed while the data is preliminary, so that every reader knows what they are working with.

  1. Request access

    Your name, institution and what you will use it for. Two minutes. Have an invite code? Enter it instead.

  2. A person reviews it

    The project lead reads every request, usually within a few days. You get an email either way.

  3. Sign in

    With a link sent to your email, or with Touch ID once you have added a passkey. No passwords.

  4. Explore and export

    A map and a table with filters, among them a comparable core for both sides. Every value has its source and the archived copy that was read. Extracts as CSV or JSON with a codebook. Views and downloads are logged.

Request access I have an invite code

Already approved? Sign in

Releases and citation

A release never changes after it is published, and records keep their ID from one release to the next. Cite the release you used, so others can work from the same data.

Releases

  • 0.6
    In preparation nextCoverage checked year by year against official lists and parliamentary and legal documents of every country in the region.
  • 0.5.3 α
    28 Oct 2026 current1,639 exercises, core 823. Reprints and agency reports count once: inclusion and every vote count independent first sources, not pages, and each record shows how many stand behind it. A government's figure repeated by the press is kept as that government's statement. Cancelled exercises are kept as records and never counted; one-nation exercises under 1,000 personnel are marked minor. Dates of the Estonian Defence Forces' moved site are read from the live pages; old Latvian army pages with another item's text, and pages with no article, give only a name and a year.
    1 correction, latest 0.5.3.1
    • 0.5.3.1 1,637 exercises, core 822. Fixes from a code review: one page at several addresses is one source; an official source's country is read from its own page; plain words the model took for names (Cooperation 2016 and 2025) removed; a size counts as confirmed only as a whole number in the text.
  • 0.5.2 α
    Earlier archived1,658 exercises with their structure: umbrella exercises and their parts, cycles, exercises held several times a year and linked exercises, each link on the source's own words with an archived copy. The kind of exercise is recorded from the organiser's page; rescue, civil and competition drills stay in the registry but leave the core (900). Nine records removed or merged: an operation, two cooperation frameworks, a forum, a postponed exercise, two pages with a wrong year and a translated duplicate (Union Courage is Union Resolve 2022). The map shows confirmed places by default.
  • 0.5.1 α
    Earlier archived1,664 exercises. A test set built from human review measures how well each article is read (91% of 126 values on the held-out part); a second prompt for umbrella exercises, contingents and sources was tested and is not used for the values. Five exercises removed: two dated a year early by their archive, one counted from a planned edition, and two events no source calls an exercise.
    1 correction, latest 0.5.1.1
    • 0.5.1.1 Same data. Editions where a person checked values against the sources are marked as such (187), not only the 41 that a check corrected.
  • 0.5.0 α
    Earlier archived1,670 exercises, the first release with measured accuracy: 85.5% of 401 randomly drawn values were right when a person checked them against their sources. Place now comes from the most reported location; exercises placed only outside the region removed; dates no longer taken from the day a ministry site moved; 54 corrections from review.
    3 corrections, latest 0.5.0.3
    • 0.5.0.3 Same data. Every exercise names the extraction prompt behind it, and the release names the model and prompt versions.
    • 0.5.0.2 Same data as 0.5.0.1. Accuracy, recall and counts are computed into the release and the site reads them from there; one command rebuilds the release from the corpus.
    • 0.5.0.1 1,669 exercises: Zapad 2023, which was called off, removed; the audit now knows the years of Russia's strategic exercises (Zapad, Vostok, Tsentr, Kavkaz, Union Shield).
  • 0.4
    Earlier archived1,720 exercises. Adds 7,577 Russian-language pages of Russia's defence ministry; contests and exercises outside the region removed; 43% placed at a training area, base or town.
    5 corrections, latest 0.4.5
    • 0.4.5 19 sizes corrected after a check of every source: a contingent, ships or aircraft no longer stand for the whole exercise (Iron Wolf 2015, Baltic Fortress 2015), and one outlying report no longer overrides the organiser (Dragon 2024: 20,000; Northern Griffin 2020: 740).
    • 0.4.4 1,688 exercises: 20 exercises listed twice in two languages merged after a pair check; first correction from human review (Winter Wolf 2022: about 215 personnel, not 1,115); DRAGON, Arrow and Viking restored.
    • 0.4.3 Russia no longer counted in NATO-side exercises it only appears beside (Danex-NoCo 2025, BALTOPS 2020); a single mention of Russia no longer decides the side.
    • 0.4.2 Adds the comparable core: 979 exercises that are multinational or have 1,000 personnel or more, one rule for both sides.
    • 0.4.1 1,713 exercises. Side and organiser fixed where no participant is confirmed; Zapad only in the years it was held; translated duplicates of Union Resolve merged.
  • 0.3
    Earlier archived1,763 exercises: news reports plus 34,189 archived pages of the region's defence ministries, armed forces and NATO commands. Official and press figures shown side by side; 55 marked disputed.
  • 0.2
    Earlier archived519 exercises confirmed by two or more news sources; names matched across languages.
  • 1.0
    PlannedReviewed by people, with a DOI through Zenodo.

Cite

Andrianov, I. (2026). Exercitus: Registry of military exercises in the Baltic Sea region, 2015–2026 (Version 0.5.3.1) [Data set]. https://exercitus.eu
DOI from release 1.0.

Who makes it

An independent research project, built in Berlin.

Ivan Andrianov builds and maintains Exercitus. He is an MA student in North American Studies at the John F. Kennedy Institute of Freie Universität Berlin and a student research assistant at its Institute for East European Studies.

The registry grew out of Threshold, an open event database on military escalation. Threshold's index did not beat a free baseline in a pre-registered test, so it was taken offline and its tools were turned to a narrower question: exercises, where official sources exist to check the press against.

ComputeKISSKI AI service centre, GWDG Göttingen
Group leaderDr. David Bosold, John F. Kennedy Institute
Contactivan@evandrianov.pro