Funded by the European Union · Horizon Europe · GA 101132671FIMI Attribution Event Dataset · Deliverable D4.2 · release v2.4.3
Funded by the European Union · Horizon Europe · GA 101132671
The dataset records 3,029 incidents of foreign information manipulation that 97 European and allied institutions attributed in 534 published documents. Each record carries the source document, the page and the sentence that support the attribution. The dataset measures how European institutions attribute: which institutions publish, how quickly, about which targets and actors, and how often they corroborate one another. It does not measure how much manipulation takes place.
The four measures are recalculated for the current selection. Each describes the published record of European institutions, and none describes the behaviour of any foreign state.
The occurrence series counts incidents by the year in which they took place; the publication series counts them by the year in which an institution first published an attribution. The difference between the two is the delay before an incident enters the public record.
The attributed state is the state the source named. The actor category and the DISARM techniques record how the source characterised the operation. All three describe published attributions, and none estimates the behaviour of a state.
Incidents by the state the source named. Selecting a bar filters every panel.
Incidents by the controlled actor category assigned in the dataset.
Incidents by the type of episode the source described. An incident may carry more than one type.
Share of each state's attributed incidents that carry the technique. Columns are attributed states; rows are the 16 most frequent techniques.
Two techniques are linked when they appear in the same incident. Node size is the number of incidents carrying the technique; link width is the number of incidents carrying both. Hovering over a node lists the techniques that most often accompany it.
This section describes the institutions that publish attributions: how many incidents each contributes, how concentrated the record is among them, how often two institutions independently attribute the same incident, and which jurisdictions report on which countries.
Incidents by canonical parent institution. A jointly authored document counts for each institution behind it.
Cumulative share of incidents with an identified reporting institution, by number of institutions ranked from the largest contributor.
Share of each year's published attributions by type of attributing institution, 2015 onwards.
Three nested measures of agreement between institutions.
Two institutions are linked when they attributed the same incident independently in separate documents. Node size is the number of the institution's published attributions; link width is the number of independently corroborated incidents the two share.
Each cell counts attributed incidents. In the first view the diagonal holds jurisdictions reporting on their own information space and the other cells hold attributions about another country. Hovering over a cell shows the institution types, attributed states and publication delay for that cell.
Every country appears once. Ribbons run from the jurisdiction of the attributing institution to the country the incident concerned; a ribbon that loops outside the ring is a country reporting on itself. Hovering over a segment isolates its flows, and selecting a segment keeps them isolated.
Incident counts by target country reflect institutional capacity as well as exposure. The second measure gives the share of each country's record that its own institutions published; the third gives the number of independent institutions that attribute incidents against the country.
Incidents by target country. The blue segment is the share of the country's record published by its own institutions.
The table compares two values of one dimension over the current selection. Any filter on the chosen dimension is ignored for this table.
Every row names its source document and page and carries the sentence that supports the attribution. Selecting a row opens the full record. The documents themselves are not yet linked, so verification requires retrieving the document by its identifier.
| Year | Incident | Target | Attributed to | Published by | Confidence | Evidence |
|---|
We leave the judgement of what constitutes foreign information manipulation with the institutions that make it publicly and answer for it, and we confine our own work to reading their published documents and recording every attribution we find in a common schema, together with the sentence that supports it. A language model reads each passage under a fixed instruction set; every record must carry a verbatim quotation, fields stay empty where the source is silent, and the actor is recorded as the source names it. We normalised institution names through a curated alias table, linked duplicate descriptions conservatively, and coded actors by category.
We tested the alternative, in which a language model reads news archives and decides for itself what counts as manipulation, and rejected it. Genuine incidents are rare relative to the volume of news, so a classifier at the measured agreement rates would produce roughly 51 false flags for every incident correctly identified, and a system that labels published speech as foreign interference needs an accountable author. Deliverable D4.2 sets out the test and the design in full.
We release the data and the codebook under the Creative Commons Attribution 4.0 International licence. The codebook lists every figure we report with the expression that reproduces it from the released file, and the script reproduce_figures.py in the supporting materials runs those expressions in full. Cite as: DE-CONSPIRATOR Consortium (2026). The DE-CONSPIRATOR FIMI Attribution Event Dataset, release v2.4.3. Deliverable D4.2, Horizon Europe Grant Agreement 101132671. Until the dataset has a persistent identifier of its own, 10.3030/101132671 identifies the project alone.