
Situational Awareness Terminal
Methodology & Scope
WorldWideWatchers is an automated open-source intelligence service with human oversight. Software collects and analyses public reporting on international security four times a day. Stories that pass every check are published without prior human review; one analyst sets the rules, reviews what the rules hold back, and answers for what is published. Stories an analyst has since reviewed are marked ✓. This page describes what the system actually does, including where it falls short.
What we cover
Four areas: regional conflicts, national security threats, counter-terrorism and cybersecurity. Stories outside them are filtered out before analysis.
1. Collection
- Two commercial news services (NewsAPI, NewsData.io) and about 65 RSS feeds: news agencies, regional outlets, government and specialist security publications.
- Feeds in Russian, Ukrainian, Arabic and Chinese are machine-translated before analysis.
- The pipeline runs four times a day (about 06:00, 12:00, 18:40 and 23:40 Athens time). Each day it sees about 3,000 candidate articles; most are dropped as duplicates, off-topic or low quality. About 20–25 per run are analysed in full, and 70–90 stories a day are published.
- Feeds that stop responding are quarantined automatically and retried later.
2. Source reliability
Every source domain carries a Source Credibility Index (SCI) from 1 to 5, shown on each story together with the equivalent NATO (Admiralty) grades for source reliability (A–F) and information credibility (1–6):
- 5: primary source, official advisory or top-tier newswire (e.g. CISA, Reuters, AP, BBC).
- 4: reliable specialist or quality newspaper with strong editorial standards.
- 3: generally credible, with a known slant, regional bias or occasional accuracy problems.
- 2: unreliable, highly partisan or sensationalist; used with caution.
- 1: state-controlled outlet. Its content is treated as the official narrative, not as independent reporting, and stories that rest on it carry a notice saying so.
A story resting on a source rated 1 always goes to human review. For the stricter rule on single sources, see the next section.
3. Corroboration: how many sources a story rests on
Articles about the same event are grouped together, and each story states at the top how many independent sources it rests on: “Single-source assessment” or “Multi-source assessment (N sources)”, with the sources named.
A corroboration score (0–1) combines the number of sources, how many of them are independent of each other, and whether they agree; each conflicting report lowers it.
What this means in practice. For regional conflicts, national security and counter-terrorism (cybersecurity stories, which often come from a single vendor report, are exempt), a story from a single source is published automatically only when that source is rated 3 or higher and nothing in the reporting conflicts. Otherwise it waits for human review. Most of our stories are in fact single-source: in a typical week about 85% rest on one source and about 15% on two or more. We label this rather than hide it, and the front page gives priority to stories with two or more sources when they exist. Finding a second, independent source before publication is the main improvement we are working on.
4. Analysis
- Relevance and importance. Each article is scored against our intelligence requirements, using rules (high-risk terms, actors, places) combined with a language model’s judgement.
- Writing the assessment. A language model writes the brief from the source text only: bottom line up front, key judgements, and competing hypotheses (Analysis of Competing Hypotheses), including an explicit check on whether deception can be ruled out.
- Confidence in words and numbers. Each judgement carries a confidence level (High, Moderate or Low) and a probability expressed in the estimative language used by intelligence analysts, with its percentage: Almost certain (95% or more), Highly likely (85%), Likely (70%), Probably (55%), Roughly even chance (45%), Unlikely (30%), Highly unlikely (15%), Remote (below 15%).
- Checking the text against the source. A separate model, not the one that wrote the brief, checks that each statement is supported by the original article (natural-language inference). Briefs that fail this check are not published automatically.
- Other checks. Conflicting reports on the same event are flagged rather than silently resolved; source articles that look machine-generated are marked as suspect; known threat actors are recognised under their different names; for cybersecurity stories, technical indicators (IP addresses, domains, file hashes, CVEs) are extracted and stored in a structured form that can be exported as STIX 2.1.
5. What is published, and what waits
A story is published automatically only when every check passes. It goes to a human review queue when, among other reasons, the text is not well supported by its source, the source is rated 1 or 2, the reporting conflicts, the source looks machine-generated, or a single source is below the threshold above. Sending a story further (Telegram, partners) has stricter thresholds than publishing it on the site.
Every published story ends with a Source Verification & Governance Report showing how it got there: source reliability, corroboration, faithfulness to the source and the publication decision.
6. Warnings
We watch a small set of scenarios (currently eleven, among them escalation involving Iran, Russia and NATO, the war in Ukraine, China and Taiwan, the Korean peninsula, India and Pakistan, Red Sea shipping, and an imminent terrorist attack), each with a few observable indicators. A warning appears at the top of the site when at least two of a scenario’s indicators are reported within its window (72 to 120 hours). Each indicator hit is first checked by a second model that asks whether the event actually happened, because keyword matches alone were right only about a third of the time.
7. Corrections
Every story has a “Report an error” link. Reported errors are checked by the analyst, and corrections are made on the article itself.
8. Limitations
- Most stories rest on a single source (see section 3).
- We read what is published openly, in batches four times a day; we are not a live feed and we do not use classified, leaked or paid-for private data.
- Language models make mistakes. The checks above reduce them; they do not remove them. Probabilities are analytic judgements, not statistical forecasts.
- Coverage follows the sources we collect, so some regions and languages are better covered than others.
- One analyst oversees the system; there is no editorial team.
Last reviewed: November 2026. Questions about the method: contact us.