How it works
From 87 sources to one five-minute briefing
Every issue is produced by the same automated run: collect, group into events, let a language model write the issue, then check its work in code. Here is each step, what is AI and what is not, and what stops a bad issue from going out.
The daily run
-
87 sources · 15 perspectives
85 RSS feeds + 2 scraped sites; 55 are non-Western and 15 are state-affiliated (always labeled).
-
Collect and enrich
All feeds are fetched in parallel. Short RSS summaries are replaced with the article's full text; articles readers could not read (non-Latin script) are skipped.
-
Group articles into events
Sentence embeddings and density clustering put every report of the same event in one group — so a Western and a non-Western account of one event sit side by side.
Fallback: if the embedding model cannot load, plain title-similarity de-duplication keeps the run going.
-
Pre-select what the model reads
Events are ranked by how many outlets confirm them and how many different perspectives cover them; up to four articles per event are kept, from different perspectives.
-
AI call 1 — write the issue
One call picks and ranks the top stories, writes the quick hits and the big number. The model also sees the last issues so it can lead with fresh events.
Readability gate: copy above a 9.5 school-grade reading level is sent back for a rewrite (at most twice) — kept only if it reads easier without turning choppy or losing a figure.
-
Editorial rules, enforced in code
Running storylines move to Developing; items that repeat recent issues are dropped; links are checked against the text they sit under; headlines are set in sentence case; off-topic items (crime, celebrity, sport) are removed.
-
AI call 2 — how the world covers it
For the big story the model gets the event's articles grouped by perspective and returns each group's framing plus one quote. The Blindspot is picked by code: an event reported by at least two outlets we track, none of them Western.
Quote check: a quote must appear word for word in the source article, or it is dropped.
-
Signals
Prediction-market odds and news-volume trends for the big story, added only when a market matches the story's specific names.
This website
The issue page, a permanent page per story, archive and RSS.
Sent to daily subscribers; delivery is verified after sending.
Issue JSON
Machine-readable copy with model, cost and decisions. Feeds the Sunday digest — no AI needed.
Numbers on the arrows are from the November 22, 2026 run. Latest issue written by deepseek/deepseek-v4.1-flash-1017.
Live figures for every issue are on the dashboard.
What keeps it honest
- Quotes are verbatim. Every quote in the perspective grid is matched against the article text. A paraphrase never appears in quotation marks.
- State media is data, not a source of fact. State-affiliated outlets are labeled wherever they appear and show how a government frames an event; they are never the only source for a claim.
- Who covered it is counted, not written. Coverage bars and the blindspot's “reported by” line come from the article data, so the model cannot claim coverage that is not there.
- No reruns. Storylines are tracked across days by their distinctive names. A story you already read comes back only as a one-line update with new facts.
- Plain English. Each issue is measured with the Flesch-Kincaid formula before it is published; the result is on the dashboard.
- Every claim links to its source, and each link is checked to be about the sentence it is attached to.
- A spending cap. AI calls stop when the monthly budget is used up.