Methodology

NarrativeRadar continuously monitors over 130 sources organized into 8 geopolitical information blocks: Western media, Russian media, Chinese media, Middle Eastern media, African media, official government channels, financial press, and regional outlets.

Sources are selected for geographic and editorial diversity, not ideological alignment. New sources are added based on reach, reliability, and block coverage gaps. RSS feeds are checked every 30 minutes.

Individual articles are not shown to readers. Instead, a clustering engine groups related articles about the same development into a single event. Clustering uses five layers of matching:

• Semantic similarity (vector embeddings)
• Keyword overlap (Jaccard similarity)
• Keyword signature matching
• Named entity overlap (countries, organizations)
• Language grouping

An event is created only when multiple sources report on the same development. Single-source reports are tracked but not promoted.

Not every event receives the same depth of analysis. The tier depends on source coverage:

• Deep Analysis — 5+ articles from 3+ sources across 2+ blocks. Includes full summary, reported facts, 2-3 competing narratives, conditional scenarios, divergence analysis, and market exposure.
• Standard Analysis — 3+ articles from 2+ sources. Includes summary, facts, scenarios, and market exposure. No narrative comparison.
• Brief Summary — Remaining events. Includes summary, key facts, scenarios, and market exposure.

For events with Deep Analysis, the system identifies 2-3 competing narratives — distinct interpretive frameworks promoted by different information blocks. Each narrative includes:

• A label identifying the framing (e.g., "Western deterrence narrative")
• A summary of the interpretation
• Specific claims attributed to sources
• The information blocks promoting this framing

If only one narrative exists for an event, it is deprioritized — narrative divergence is what makes an event analytically interesting.

When narratives conflict, the system identifies the specific points of disagreement across four dimensions:

• Disputed facts — Sources report different versions of the same observable event.
• Different readings — Sources agree on facts but disagree on causes or significance.
• Blind spots — Aspects covered by some blocks but ignored by others.
• Resolution signals — Indicators that could clarify which framing is more accurate over time.

Analysis is generated using AI models applied to source articles. AI is used for summarization, narrative identification, entity extraction, and scenario generation. All AI output is constrained by editorial rules that enforce neutrality and layer separation.

AI does not decide what is true or important. It structures information according to predefined rules. Every event page discloses that analysis is AI-assisted and editorially supervised.

Take a sanctions package against Russian oil exports — the kind of multi-block event the system handles every week. Within hours, articles arrive from Reuters, AP, TASS, Xinhua, Al Jazeera and Bloomberg. The clustering engine groups them into a single event because they share entities (Russia, Brent, EU, US Treasury), keyword signatures, and embedding similarity above the 0.65 threshold.

The enrichment prompt is then run with the article cluster as input. The output is structured into the layers above:

• Facts: the package size, the designations, the date — same across all sources.
• Narratives: two emerge. Western framing reads "new enforcement teeth; shadow-fleet pressure rising". Russian framing reads "rerouted exports absorbing the shock; sanctions failing". Each is attributed to the blocks promoting it; no editorial verdict.
• Divergence points: the system flags an explicit disputed fact — Western sources give a revenue-impact estimate of 12–15 percent, Russian sources cite a single-digit impact. Both numbers are kept verbatim. The reader gets both, side by side.
• Market exposure: Brent Crude tagged with separate entries per block — Western narrative expects upward pressure, Russian narrative expects stable Asian demand.

Crucially, the same prompt is applied to every qualifying event. The rules don't change because the topic is sensitive — they apply to a Russian oil package, a Federal Reserve announcement, and an Argentine election the same way. That uniformity is the methodology.

Events progress through four phases as coverage evolves:

• Emerging — Recently detected, limited source coverage
• Ongoing — Active coverage from multiple sources
• Cooling — Coverage declining, no new articles recently
• Archived — No longer actively covered

Analysis is refreshed automatically as new articles arrive. Events are re-analyzed and may be upgraded to a higher tier as more sources report on them.