Methodology
How an item of public reporting becomes a mapped, classified event — and where the method stops being reliable.
This page describes the full pipeline, including the parts that do not work well. Publishing the failure modes is deliberate: a monitoring platform that only advertises its strengths cannot be assessed by its readers.
1. Collection
Each theatre has its own source list combining RSS feeds from established outlets and public Telegram channels used by OSINT communities. Sources are chosen to cover both sides of a confrontation rather than a single perspective — a theatre monitored only through one camp's media produces a distorted map.
Collection runs continuously. New material is picked up within minutes of publication.
2. Noise filtering
Most of what a conflict-adjacent feed publishes is not an event: it is commentary, diplomacy, sport or economics. Items pass a two-stage filter. The first stage requires conflict vocabulary. The second removes known false positives — multilingual expressions that contain conflict vocabulary without describing an event, such as figurative or sporting usage. Both lists are maintained per language, because a filter tuned on English alone fails badly on Arabic, Persian, Romanian or Russian material.
Items that pass only the weak stage are kept but require additional context before being mapped.
3. Side attribution
Every mapped event is attributed to one of the two sides of its theatre. Attribution is weighted rather than binary: vocabulary present in the item, the editorial alignment of the source, and the number of independent mentions all contribute. When the weighting is inconclusive, a deterministic fallback applies and the event is flagged rather than guessed.
Attribution describes who is reported to have acted. It is not a finding of responsibility.
4. Geolocation and time
Place names are extracted and resolved to coordinates, then normalised so that the same location reported under different spellings or transliterations collapses to a single point. Timestamps are normalised to UTC, which is what makes a 24-hour window comparable across theatres and languages.
5. Translation
Interface and headlines are made available in multiple languages through machine translation. Machine translation of conflict reporting is imperfect, particularly for named entities and military terminology. Translated text is an aid to comprehension; the original source is always linked and remains authoritative.
6. Measured quality, by theatre
Figures below were measured on 2026-09-13 on the most recent events available for each theatre. They move over time and are re-measured periodically.
| Theatre | Sources | Side attributed | Geolocated |
|---|---|---|---|
| Iran | 29 | 100% | 100% |
| Ukraine | 23 | 100% | 100% |
| Taiwan Strait | 14 | 85% | 100% |
| Europe | 25 | 100% | 100% |
| Cuba | 10 | 100% | 100% |
7. Known limits
- Reporting bias is inherited. If an event is not reported publicly, it does not exist for GCA. Under-reported areas look quiet; they may not be.
- Duplicate reporting inflates volume. Corroboration counts reduce this but do not eliminate it.
- Geolocation is as precise as the source. "Southern Lebanon" becomes a point, not an area. Points are indicative, never targeting-grade.
- Automated classification makes mistakes. Errors are expected, not exceptional. The corrections procedure exists for that reason.
- No ground verification. GCA never confirms that a reported event occurred.
8. Citizen reports
Users can submit an observed event. Submissions are geofenced to the relevant theatre, clustered against nearby reports, and shown as user-submitted — visibly distinct from media-sourced events. They are never merged into the verified set.