Methodology

Raw news items undergo a multi-stage AI enrichment pipeline:

1. Relevance Filtering — Keyword + geo-anchor matching ensures only Middle East conflict items pass through. Articles must match both a conflict keyword and a geographic anchor.
2. Entity Extraction — AI identifies actors (state forces, militias, organizations), weapons systems, and locations mentioned in the text.
3. Event Classification — Each event is assigned a type (airstrike, missile launch, ground combat, etc.) and a severity score (1-5).
4. Trajectory Inference — For kinetic events, the system infers likely launch origins from a database of 80+ known military installations, bases, and launch sites.
5. Tactical Significance — AI generates a brief assessment of why this event matters operationally.

Events are scored on a 5-point scale reflecting military and humanitarian significance:

The CSI is a composite score (0-100) computed from four indicators, inspired by academic conflict measurement frameworks:

• Deadliness — Weighted casualty count normalized against historical baselines
• Danger to Civilians — Ratio of humanitarian events and civilian impact indicators
• Geographic Diffusion — Number of active theaters and geographic spread of events
• Actor Fragmentation — Number of distinct armed groups involved, indicating conflict complexity

The CSI updates in real-time as new events arrive. Levels: LOW (0-25), MODERATE (25-40), HIGH (40-60), CRITICAL (60-80), EXTREME (80-100).

WARTRACKER generates per-theater escalation forecasts using a 4-signal z-score analysis:

• Event Velocity — Rate of new events compared to the rolling 7-day average
• Severity Trend — Direction and magnitude of severity score changes
• Casualty Rate — Acceleration or deceleration of reported casualties
• Actor Proliferation — New actors entering the conflict space

When z-scores exceed threshold values, escalation or de-escalation alerts are triggered. Forecasts are currently 24-48 hour horizon.

The Network view constructs a co-occurrence graph from event actor data:

• Nodes — Actors (states, militias, organizations) sized by betweenness centrality
• Edges — Co-occurrence in events, weighted by frequency
• Communities — Louvain algorithm detects clusters of frequently co-occurring actors
• Layout — Force-directed simulation positions connected actors near each other

WARTRACKER operates on a real-time streaming architecture:

• RSS/News — Polled every 5 minutes via server-side ingestion
• Aircraft (ADS-B) — Real-time streaming via OpenSky/ADSBx
• Vessel (AIS) — Near real-time maritime tracking
• FIRMS Satellite — Updated every 3 hours (NASA VIIRS/MODIS)
• USGS Seismic — Real-time earthquake data
• Weather/NOTAMs — Updated every 30 minutes

Events are delivered to the browser via Server-Sent Events (SSE) for sub-second latency from ingestion to display.