Global Conflict Intelligence Globe

How an event gets published

1. Collection

RSS and Atom feeds from news organisations, the GDELT public article index, and manual entry by analysts. Collectors respect rate limits, honour conditional requests, and do not bypass paywalls, access controls or platform protections. Connectors for public social and structured-data sources exist but stay disabled until their credentials and licence terms have been reviewed — and remain listed as unconfigured rather than disappearing, so a coverage gap is always visible.

Only the title, URL, publisher, timestamp, a short excerpt and a content hash are stored. Never full article text, and never media files.

2. Processing

Each report is normalised, classified for relevance, and mined for entities and claims: places, weapon mentions, casualty and damage figures, attribution phrases, hedging language, denials and retractions. It is then geolocated, matched against a weapon catalogue, and clustered with existing events that agree on time, place, type and wording.

All of this runs on deterministic rules and dictionaries. No language model is required, and no model output alone can mark an event confirmed.

3. Corroboration, counted honestly

Counting articles overstates corroboration: one wire story republished by twenty outlets is one source, not twenty. Reports attached to an event are grouped into independent source families by publisher and by near-duplicate text detection. Reports in the same family count once, and the resulting graph is shown on the event page so you can disagree with it.

4. Confidence, as a published formula

A 0–100 score combining independent source diversity, geographic evidence quality, temporal consistency, media and sensor corroboration, official acknowledgement and analyst review — minus explicit penalties for circular reporting, anonymous sourcing, heavy hedging, public denials, retracted or satirical sources, unattributed casualty figures and inconsistent locations or timestamps.

Every event shows the exact components, weights and penalties that produced its score. The score never sets an event's status. It ranks what a human reviewer should look at first.

5. Human review

Nothing reaches the public interface automatically. An analyst approves each event and states a reason, and that reason is recorded permanently. Corrections publish the old value, the new value and the reason. A retracted event stays visible and clearly marked, because deleting it would erase the public record of the error.

Location uncertainty

Every located event carries an accuracy radius, and the map draws it. A vague phrase is never resolved to a falsely exact point — the radius widens instead. When nothing can be matched, the event has no coordinates at all. Coordinates are never fabricated.

What this platform will not do

This is a historical and near-real-time public-source record. It is not an operational system, and the constraints below are implemented in code and in the database, not merely stated here.

• No target recommendations, target lists or strike planning.
• No ballistic, trajectory, apogee, terminal-profile or interception-window calculations.
• No prediction of where any military platform will travel next.
• No tracking of individual people.
• No precise live positions of concealed or active military aircraft, vessels or vehicles — regardless of whether those positions appear in an AIS, ADS-B or social-media report.
• No weapon fabrication information, and no guidance for evading radar, air defence, surveillance or law enforcement.
• No identification of private individuals or private homes from media.
• No face recognition, ever.

Where a report may describe an active military position, its location is coarsened to a grid and publication is held for at least an hour. Heading, speed and any notion of predicted course are never published. That is a deliberate loss of precision and timeliness, accepted because the alternative could contribute to harm.

Data and access

The platform exposes a documented, versioned read-only API with exports in JSON, GeoJSON, CSV, KML and CZML. Every export carries attribution, status, confidence and accuracy radius, so the caveats cannot be silently stripped in transit.

If you build on this data, preserve status, confidence_score and accuracy_radius_m. An event shown without them is being misrepresented, and events flagged is_synthetic are test fixtures, not real reports.

Operator

The platform is presented as a neutral public-source record. It is built, operated and reviewed by Dana Mohammad Khidhir, independent developer and researcher. Corrections and takedown requests go to danamohammadkhidhir@gmail.com — requests that name specific material can be acted on fastest.

Source material can be removed on request. The event and its audit trail are preserved, so a takedown never silently rewrites history.