Conflict Signal // Methodology

How this data is produced

Conflict Signal maps armed-conflict incidents from open-source reporting in near real time. This page states exactly how that happens, what the numbers mean, and where the method is weak — because a tracker you cannot audit is not worth trusting.

Where does the data come from?

Every incident originates from a public open-source intelligence channel on Telegram. Posts are ingested continuously, roughly every five minutes, with their attached photos and video. No incident is invented, and nothing is scraped from behind a paywall.

How is a post turned into a mapped event?

A language model classifies each post against a strict rule set: it must describe one concrete armed action that has already happened, in a named, mappable place. Diplomacy, threats, analysis, daily round-ups, crime and disasters are rejected. Roughly one post in four becomes a mapped event; the rest are discarded.

How accurate are the coordinates?

Coordinates are the model’s best estimate for the named locality, not a survey position. A point marks the town, base or district that was reported, not a precise impact site. Reports that name only a country or region are not mapped at all, because a country centroid would imply precision that does not exist.

What does the escalation score mean?

It is a relative 0–100 indicator derived from recorded incident volume and reported casualties over a rolling window. It is useful for comparing a country against itself over time. It is not a casualty estimate and not an authoritative severity rating.

How are casualty figures handled?

Fatality counts are taken from the wording of the original report, matched literally rather than inferred, because a model asked to reason about numbers gets them wrong. Where no figure is stated, none is shown. An absent figure means nothing was reported — it is not a claim that nobody was harmed.

What are the known limitations?

Single-source reporting cannot be cross-verified, so nothing here should be treated as confirmed. Coverage follows whatever the source channel covers, which is uneven by region. Military vessels and aircraft appear only when they choose to broadcast a transponder signal, so silence is never evidence of absence. Early reports of any incident are frequently revised.

Which parts are written by AI?

Classification, the per-country situation synthesis and the long-form briefings are AI-generated from the source reports and labelled as such wherever they appear. The underlying incident records, counts and scores are computed from the data, not written by a model.

Corrections

Incidents are revised as reporting develops, and records here are updated in place rather than rewritten silently. If something is wrong, it is worth telling us — an error left standing is worse than a gap.