What this is
WW3Watch is a real-time aggregator of conflict and geopolitical news from a catalog of 200+ sources across every major region and perspective — US and European wires next to Iranian state media, Israeli papers next to Arab ones, Russian outlets next to OSINT researchers. It does not tell you what is true. It shows you who is saying what, side by side, as they say it.
In conflict coverage, the wording is the data: whether an outlet writes "martyred", "killed", or "neutralized" tells you something no summary can preserve. So WW3Watch never summarizes, never paraphrases, never blends voices. Every headline appears exactly as its newsroom wrote it.
The rule the system is built on
Machine intelligence routes stories; it never rewrites them. Classifier and embedding models decide where things go — whether an article is conflict-relevant, which story it belongs to, what is trending, which tags it carries. Every one of those judgments comes from a model that cannot generate text at all: it scores, or it answers narrow typed questions with a probability. None of them touch what a journalist wrote. The single exception, translation — into whatever language you read in, set once — is the only place a text-generating model is used: opt-in, clearly labeled, and one click away from the original.
- Relevance — two tiers, cheapest first: a small local classifier settles the obvious cases, and a decision model (TypeSafe's Jev) gives the rest a calibrated probability — its verdict is final. A random slice of the local classifier's confident verdicts is re-checked by the decision model, every run.
- Story grouping — a multilingual embedding model maps every headline into a shared semantic space; articles within a tight similarity threshold and time window join the same story, which is how a Persian headline and a Norwegian one about the same strike end up grouped. Similarity means "same subject", not "same event" — so when a match is close but not certain, the decision model is asked one question: are these two headlines the same news story?
- Trending — for each of the biggest stories of the last few hours the decision model judges three things: how consequential the event is, whether it is a new development, and whether it is only talk. Code weighs those against how many independent sources, regions and languages carry the story. The weights are in the source, not in a prompt.
- Tags and filters — "major", "statement", "analysis", "unconfirmed", the topic and the parties involved are the same kind of judgment, made once per article from its headline and summary. They are a classifier's reading, not an editor's — they let you filter; they change nothing you read.
- Wire detection — articles whose text is near-identical to an earlier article in the same story are marked "wire", so "12 sources covered this" doesn't overstate independent confirmation when most are reprinting one agency's copy.
The current window, by region
Distribution of the 500 most recent articles (the feed's serving window). Volume varies with the news cycle and with which feeds are reachable — shown here so the skew is visible rather than implicit.
Recently highlighted
The stories the curator surfaced to Trending over the past few days — distinct picks, most recent first. Trending itself only ever shows the current top three; this is the trail it leaves.
Loading recent highlights…
Every source, with its health
The full roster, live from the database the pipeline maintains. Green: fetched successfully on recent runs. Amber: failing recently (many news sites block datacenter IPs; a proxy rescues most). Red: failing for half a day or more. Gray: disabled.
Loading the live roster…
Follow it elsewhere
Two RSS feeds, one item per story, opening in the reader here: everything, and major events only — the low-volume one to point a phone's notifications at. How each party's activity has moved over the last month is on Trends.
Built in the open
The entire system — pipeline, clustering, this page — is open source under AGPL-3.0: anyone running a modified version as a service must publish their changes, so every derivative of this site stays as auditable as this one. Reader content is cached to survive link rot. No visitor accounts or advertising analytics. Abuse-prevention records protect the service; see our privacy details.