The pipeline
Scan the source layer
Every six hours, the pipeline polls a curated set of primary sources — wire services (Reuters, AP, AFP), central banks, trade ministries, IAEA, IMF, WMO, customs authorities, and a small roster of independent analysts with verifiable track records. Social and opinion media are never read.
Cluster and deduplicate
Raw items are grouped into candidate events. An event that appears in only one source — or only in sources that cite each other — is flagged contested and set aside. Events confirmed by at least two independent primary sources move to drafting.
Draft structured briefings
A language model drafts each fact in the four-part structure: what happened, why it matters, what could happen next, sources. Numbers and quotes must be verbatim from a cited document. Implications are drawn only from institutional analysis, not model speculation.
Verify every claim
A second pass checks each assertion against its citation. Unverifiable claims are dropped; ambiguous ones are rewritten or demoted. Facts that lose more than one claim in verification are not published at all.
Rank and select
Verified facts are scored on materiality (how many people it affects), durability (does it still matter in a week), and coverage gap (is the mainstream picture missing something). The top 5–10 across the four pillars become the day's briefing.
Publish statically
The briefing is rendered at build time and served from the edge. No skeleton screens, no loading states, no client-side hydration of the reading experience. You land, you read, you leave.
What we read from
The source roster is deliberately narrow. Adding a source requires a track record of primary reporting and institutional accountability. We prefer fewer sources with higher trust to many sources with lower trust.
- Wire services — Reuters, AP, AFP, Kyodo, PTI, Xinhua (flagged for state attribution)
- Multilateral institutions — IMF, World Bank, WTO, WHO, IEA, WMO, UN agencies
- Central banks & treasuries — Fed, ECB, BoE, BoJ, PBoC, plus major national statistics offices
- Customs & trade authorities — national trade and customs bulletins, shipping registries
- Scientific bodies — NOAA, IMD, national meteorological services, peer-reviewed preprint servers
- Independent analysts — a small list of individual researchers with verifiable methodology
How we handle disagreement
When sources disagree on a material number — a casualty count, a trade figure, a temperature reading — we publish the disagreement itself as the fact. The reader sees the range and who reported what.
When sources agree on the headline but diverge on causation or implication, we publish the headline as the fact and note the open interpretations in the “what could happen next” section, each attributed.
The editorial standard
A fact is published only if all of these are true:
- Every numerical claim traces to a cited document
- At least two independent primary sources confirm the event
- No language in the fact characterizes motive or intent without direct quotation
- The “why it matters” section cites institutional analysis, not model inference
- The pipeline's verification pass flagged zero unverifiable assertions
What runs this
The reading surface is Next.js rendered statically to the edge. The database of facts, topics, and briefings lives in Postgres. The pipeline is plain TypeScript on a cron, calling language models for drafting and a separate verification model for the checking pass. Nothing about the reading experience depends on the backend being up.