How the Global Security briefs are made

What these briefs are #

The Global Security briefs are an open-source intelligence digest on world security, military posture, commodities and shipping, and cyber threats. They aim to give context, not a headline stream: what happened, how it compares to the usual pattern, what serious outside analysts think (including where they disagree), what it means for markets and business, and what to watch next. Media framing is checked against primary data where it can be.

Who writes them #

The briefs are written by AI. Anthropic's Claude language models collect, draft, check and edit them on a fixed schedule, and they are published automatically, without review by a person. Each article says so at the top. Treat them as a well-sourced starting point and check the linked sources before relying on anything.

The checks described below reduce errors; they do not remove them.

Sources and collection #

• Feeds: several dozen news, analysis, official and social feeds, grouped by region and topic. A reliability-rated list of the sources is reviewed weekly.
• Kinds of source:
  • wire and broadcast news
  • think tanks and named analysts
  • official sources: governments, militaries, regulators and statistics agencies
  • trade and shipping press
  • selected social media accounts from more than one side of each conflict
• Full text: the text of collected articles is kept in a private research archive, so later briefs can compare against what was said before.
• Gaps: web search and primary documents are used to fill them.

Rating sources #

Every source carries a reliability rating on the Admiralty scale:

• A source's rating limits what it can support: a D or E source never supports "confirmed".
• New sources start on probation.
• A source caught publishing false information is quarantined.
• A lean is not a falsehood. Partisan and state-aligned sources are kept when they show what a party claims, and their lean is named.

Fact labels and confidence #

Every factual statement carries one label and a link:

Several outlets rewriting one wire story count as one source. So do an outlet's editions in other languages, sister papers that share copy, and any outlet that attributes the claim to another; the texts are compared before two outlets are counted as independent.

Source fidelity is the hard rule: every factual clause must trace to a sentence in a source that was read. Nothing is added from memory, and events are linked only where a source links them.

Judgments about what comes next use the US intelligence community's estimative scale (ICD 203):

Each key judgment also carries an analytic confidence (low, moderate or high) and the reason for it.

Daily briefs #

Each night the day's new articles are first grouped into events: reports from different outlets and languages about the same occurrence are put together, so a story told ten times is read once, and wire reprints are not counted as separate sources. The grouping uses text similarity and an automated decision model. A news-desk pass then reads the day's ranked stories across every region and picks the ones that matter, with a guaranteed place for quieter regions and for anything on the weekly watch list. Those are written up by region or topic, each with its label and source.

A separate checking pass then reads every claim and link against its sources. It rewrites or cuts anything the source does not support, and replaces any link that cannot be traced to a real article. A day that has not passed this check is not published.

Weekly brief #

The weekly brief covers Sunday to Saturday and is published on Sunday morning. It is built like an intelligence shop: collection, then desk analysis, then editorial synthesis.

• Eight desks, each researching its week from its own sources, cross-cutting analysis and wire coverage, the archive, web search and primary documents:
  • Americas
  • Europe & Russia/Ukraine
  • Middle East & Gulf
  • Indo-Pacific
  • Africa
  • Space & Arctic
  • Commodities, Shipping & Sanctions
  • Cyber & Critical Infrastructure
• Every development a desk reports carries a verbatim quote from its source, so it can be checked.
• Each desk section contains: developments, context and baseline, named third-party analysis and dissent, an estimative assessment, business and commodity impact, a check of media framing against primary data, and indicators to watch.
• The editing step fact-checks the claims that lead the brief. A separate checking model reads the cited sources and returns the sentences that support or contradict each claim. Whatever fails is fixed or cut.
• Each week the source list is reviewed: probation sources are judged, accuracy problems are traced to their source, and sources that published false information are quarantined.

Information Operations #

A weekly feature, published on Sunday evening, on information operations: covert or state-aligned efforts by any state, and the networks it runs or works through, to shape what people believe. It is observational. It records what is claimed, who pushes it, how it spreads and how far it gets. It follows methods published by professional investigators.

• The same evidence bar for every state, the United States, Canada and EU members included. Coverage follows what was documented that week, not a quota. Documentation is itself skewed by who funds investigations, so each attributor's credibility is graded.
• Operation or false claim: something is called an operation only when someone has documented all four of the French agency VIGINUM's criteria:
  • clearly false or misleading content
  • artificial, automated, or massive and deliberate spread
  • likely harm to a state's fundamental interests
  • a foreign actor's involvement
• Every case is described the same way:
  • ABCDE: Actor, Behaviour, Content, Degree, Effect.
  • Channel class: official, state-controlled, state-linked or state-aligned, following the EU's diplomatic service.
  • Breakout Scale (1 to 6), Ben Nimmo's measure of observed reach, from one community on one platform up to mainstream media, prominent people and policy action. It measures reach, not effect.
  • Kill-chain phases: the phases of the Online Operations Kill Chain (Ben Nimmo and Eric Hutchins) that a source describes.
• Attribution:
  • Actors are named only when a named investigator, platform or government has attributed them. The article reports who attributed, when and on what evidence.
  • Each attribution is graded on four checks adapted from the Atlantic Council DFRLab's Foreign Interference Attribution Tracker: credibility (including funding), objectivity, evidence and transparency.
  • Patterns the feature observes itself are labelled as observations, never attributions. Public feeds rarely show coordination, so "coordinated" needs someone else's evidence.
• The Noise floor: the week's notable unverified claims, measured rather than repeated.
  • Each claim is described neutrally: what is known first, then the claim, then why it fails or what would settle it.
  • It links the fact-check or investigation, never the post that spread the claim, and names no private individuals.
  • A claim appears only if it reached Breakout category 3 or a named fact-checker checked it. Everything else is logged but not described.
• The record: every case and claim assessed is kept in a running dataset, including those not published. Each feature shows counts from it: by actor state, channel class, status and reach, this week and over time.

Global Pressure Points #

A monthly feature on the slow pressures on world security: climate and water, health and epidemics, critical commodities (food, energy and critical minerals), migration and displacement, and the trade chokepoints that connect them. The weekly brief covers events. This feature covers conditions and trends, and what they mean over the coming months.

• Researched weekly, published monthly. Each week an analyst writes an internal working note from the week's official data and reporting. Mid-month, those notes and the month's official releases become a draft, which a human editor reviews and edits before it is published. Nothing is published without the editor's sign-off.
• The numbers come first. Before any analysis, a script pulls official data:
  • disaster alerts from GDACS, the UN and EU's Global Disaster Alert and Coordination System
  • Red Cross and Red Crescent emergencies and appeals worldwide, from the IFRC's GO platform
  • outbreak reports from the World Health Organization's Disease Outbreak News
  • the FAO Food Price Index
  • daily ship transits through the world's chokepoints, from the IMF's PortWatch
  • tropical storms, space weather and river levels from the agencies that issue them
• Each item is written to be read by anyone: a plain-English bottom line, one short explanation of the number it turns on, and then the numbers themselves, each with its source and date.
• Weighted by consequence, not by data volume. Some countries' agencies publish far more detail than others. Pressures are weighed by people affected and by their effect on food, energy, trade and stability, and each edition is checked for balance across regions.
• It never makes its own forecasts. It reports observed conditions and official forecasts, warnings and outlooks, each with its issuer and date. Thresholds and scales come from a cited source, never from memory.
• Judgments use the same estimative language as the other briefs, with a stated confidence.
• The record: every pressure assessed is kept in a running dataset, including those not described. Each edition shows counts by section, region and trend, and which sections the month's trade disruptions fed.

Special features #

Occasional single-topic features, written on request when a story needs more room than the weekly brief gives it.

• Built from the article archive: the full text of every article collected for the briefs, plus primary documents fetched for the piece.
• Every quoted sentence is checked automatically against the stored copy of the article it cites, and every link must point to an archived article or a primary document that was read.
• A separate checking model reads the cited sources for each key claim and returns the sentences that support or contradict it. Whatever fails is fixed or cut.
• The same standards as the briefs: fact labels, named outside analysis with disagreements shown, estimative language with confidence, baselines and business impact.
• The editor approves each special feature before it is published.

Publication #

• Every article passes two automatic checks before it is published:
  • a scan that removes personal and internal details
  • an AI leak check that holds anything describing the people or systems behind the briefs rather than the world
• A held article stays unpublished until its source is fixed.
• Links that no longer resolve are unlinked.
• Times are given in UTC.

Limits #

• Coverage is what the feeds carry. The sources are mostly English-language, with some Russian- and Ukrainian-language channels. Regions and topics with few monitored sources get thinner coverage.
• AI can still misread sources or link events wrongly. The fact-checks catch some of this, not all. A source can also be wrong in a way no check here detects.
• The briefs cannot see inside platforms. They cannot measure coordination or bot activity themselves, so for influence operations they rely on what investigators, platforms and governments document.
• No person reviews the daily and weekly briefs or the regular features before they go out. Special features are the exception.

Changelog #