← Method & Data

Internal design document, rendered for transparency. Operational sections (deployment, job recipes, credentials) have been removed. Dated 2026-11-10; may be amended by later work recorded in the changes log.

Did the news move before the incident? An exploratory note

Companion to the incident study inc-20261015 (published at /publications/incident-study-inc-20261015/). Written 10 November 2026; published on this site as a working document on 12 November 2026. Status: exploratory, not part of the study's frozen method.

Why this note exists

The incident study follows each incident from 14 days before to 45 days after. It was built to measure reach, spread and speed after an incident. It was not built to ask whether the news was already moving before it.

This note asks that question for five incidents: the four coups and the killings in the Kidal region on 17 February 2025 (listed in the study as the Tiskadene attack). It looks 60 days back instead of 14.

What was done

  1. All Telegram posts in the corpus from day −60 to day −1 were exported for each of the five incidents (139,180 posts in total).
  2. The study's own frozen keyword rule and frozen search terms were applied to them, unchanged. For days −14 to −1 the result is identical to the study's own keyword hits (18 of 18, 7 of 7, 21 of 21, 3 of 3, 0 of 0).
  3. Every hit (115 posts) was read and put in one of four classes:
    • lead-up: about the developing crisis or military build-up before the incident;
    • event onset: the 23 January 2022 mutiny, one day before the dated coup;
    • context: same actors or place, but not pointing to the incident;
    • noise: a false keyword hit.

The classes were assigned by one reader, an AI assistant (Claude), once, knowing the outcome. They were not blind and not double-coded. The class given to every post is recorded, so the reading can be checked or overturned; see "Material" below.

Result in one table

Incident Posts in 60 days Keyword hits Lead-up Event onset Context Noise Reading
Burkina Faso coup, 24 Jan 2022 13,906 44 19 7 9 9 Two-month lead-up
Kidal region (army and Russian forces against civilians), 17 Feb 2025 45,719 28 23 – 1 4 Sharp rise in the last 11 days
Burkina Faso coup, 30 Sep 2022 34,820 24 4 – 13 7 No build-up
Mali coup, 24 May 2021 7,724 17 3 – 10 4 No build-up
Niger coup, 26 Jul 2023 37,011 2 0 – 2 0 No warning at all

Burkina Faso, January 2022: a two-month lead-up

Lead-up posts per week (day 0 = 24 January 2022):

Days Dates Keyword hits Lead-up What the lead-up posts report
−60 to −54 25 Nov – 1 Dec 8 8 Calls for President Kaboré to resign; the 27 November march broken up with tear gas; gendarmerie commanders dismissed
−53 to −47 2 – 8 Dec 3 0
−46 to −40 9 – 15 Dec 6 3 The government resigns; a new prime minister is named; organisers of the march are arrested
−39 to −33 16 – 22 Dec 3 1 Thirteen new military commanders appointed
−32 to −26 23 – 29 Dec 2 1 "The president's resignation is still on the table"
−25 to −19 30 Dec – 5 Jan 2 1 "Resignation is non-negotiable"
−18 to −12 6 – 12 Jan 4 2 Eight soldiers arrested for an alleged coup plot
−11 to −5 13 – 19 Jan 4 1 The alleged plot leader is sent to prison
−4 to −1 20 – 23 Jan 12 2 (+7 onset) The banned 22 January march; then the 23 January mutiny

Three things to note:

Kidal region, February 2025: a sharp rise in the last 11 days

The incident: 28 civilians killed in the Kidal region on 17 February 2025 (UCDP: Government of Mali and Government of Russia against civilians).

What the posts report, in order: army drone strikes near the Algerian border (24 December); air strikes including Tessalit (27–30 December); a Russian-language channel predicting a Tuareg offensive for 17 January (26,299 views); a drone intercepted at Tessalit airport and a rumour of a downed army helicopter (11 February); an army and Russian operation against a camp near Tessalit (13 February); the killing of a local armed-group leader (14 February).

The 23 lead-up posts come from 14 channels in French, Russian, Arabic and English, with 500 to 26,000 stored views each.

Two limits apply to this case.

The three incidents with no build-up

What the keyword rule misses

The search terms were written to find the incident itself: the capital, the president, the word "coup". A post about the lead-up that names only the country does not match. One example, found by hand: a Russian-language post of 29 November 2021 on the Burkina Faso protests and the demand for the president's resignation (3,739 views). It is not among the 44 hits.

So the lead-up counts in this note are a lower bound, and statements about which channels appear when hold only for the posts the rule finds.

The "our forecast came true" claims

On 23 January 2022 four Russian-language channels wrote, in three different texts, that they had foreseen the coup. The 60-day export was searched for those forecasts.

What this note can and cannot support

It can support:

It cannot support:

Material

The note rests on the public Telegram messages of the followed channels (the messages dataset on this site) and on the study's frozen search terms. The list of the 115 keyword hits with the class given to each names individual channels. Like the study's per-channel files, it is held back until channel naming is confirmed, and is available to researchers on request through the address on the About page.

Inputs: the study's frozen search terms (sha256 462c3254…), the study's keyword matcher, pipeline version 0.1.0. The study's own results and its published pages were not changed by this note.

Rendered from 2026-11-10-incident-study-pre-event-note.md · changes log