HOW IT WORKS

The machine behind the front page.

Argus Overwatch has no journalists. Four times a day it collects from 62 newsroom feeds, Google News's top stories and 6 sections, and the links shared on 18 news subreddits: 6675 publishers on 22 Nov. It groups the articles that are about the same event, has one language model write each group up from what those outlets published, and has a second model edit it. None of it is original reporting, and no person reads a story before it appears here. This page explains what that means, and then shows you the machine itself — with its own numbers, including the ones that do not flatter it.

Edition of Sun 22 Nov · 4 runs, 00:10–18:11 UTC

  1. 6,675 publishers reported
  2. 91,307 articles collected
  3. 7,369 stories found in them
  4. 484 the most-covered, written up
  5. 482 published · 2 dropped · 0 held

Rendered when the site was last published, 22 Nov 2026, 18:58 UTC. It is a snapshot, not a live view.

WHAT THIS IS

Four things to know before you read a story.

There are no journalists here.

Every story is assembled by software out of articles other newsrooms published. Argus Overwatch makes no calls, attends nothing and reads no documents. What it does is notice when several unconnected outlets report the same event, and summarize what they said. Every story lists the outlets it drew on, and those links go to the actual reporting. They are the point.

It reads the article when the publisher allows it.

Where a publisher's robots.txt permits, Argus fetches the article and gives the writer its opening 3000 characters alongside the headline and summary. Where it does not, or the link comes through Google News, only the headline and summary are used. On 22 Nov 23% of articles carried their text. A story from headlines alone is short on purpose, and a detail from deep in an original can be missing: the outlet links matter as much as the text above them.

More outlets means more confirmation.

The ranking rests on one idea: forty articles from a single outlet are still one outlet's word, while four outlets reporting the same thing separately is evidence. So outlets count for more than volume does, and every story shows how many it had. A story built on two outlets is thinner than one built on eight, and is worth reading that way.

A second model edits every story. That is still not fact-checking.

A different model re-reads each story against the same source material. It may publish it, cut what the sources do not support, credit a claim to the outlet that made it, or correct a detail to match a source. It may not add anything, and code refuses any edit that brings in a name or number the sources lack. It can also drop a story that is not news. On 22 Nov it published 419 as written, edited 63, dropped 2 and held 0 for a person. Both models come from the same company, so they may share blind spots, and nobody checks the source outlets themselves.

The rest of this page is how all of that works, in detail, with the machine's own measurements — what it collected, what it grouped, what it sent to the model, and what the editor did with it.

01 · THE PIPELINE

One pass, and a file between every stage.

Each stage reads the file the previous one wrote and writes its own. That is what makes a run resumable — a crash on story 30 of 50 resumes having paid for 29 — and what lets two processes that never talk to each other share work: the pipeline in its container writes JSONL, and the web app on the host notices. The dashed line is a real boundary. Nothing inside it can reach the site's database. It calls outward for three things only: the feeds and listings it collects, publishers' pages for article text (where robots.txt allows), and the model gateway on the host.

DOCKER · argus-pipeline · ARGUS_SCHEDULE_HOUR=0,6,12,18 (UTC) Collect RSS · Google · Reddit 87 sources Clean dedupe·filter·text 91307 articles Window published ≤ 48h 90898 articles Embed + cluster cosine dist < 0.42 7369 clusters Rank 3 terms · §03 top 100 Synthesize ≤8 articles · JSON 484 stories Edit + gate second model · code guard 482 pub · 0 held Report markdown + index last, always writes writes appends, per story writes data/raw/<date>.jsonl data/processed/<date>.jsonl data/stories/<date>.jsonl reports/<date>/ a later run the same day reads these back: merges new articles, grows existing clusters, rewrites only stories whose prompt changed (§06) HOST · the same data/ directory, bind-mounted into the container HTTP · model gateway on the host mtime polled every 30s Public server the same app, live Django review app + public site web.sqlite3 a cache of the files Model gateway writer + editor import_runs --watch imports a finished run import_run(date) — rebuilds that day's rows reads data/ synced nginx → the app argusoverwatch.com
Counts are from 22 Nov. Collect has no count because the raw file is written after normalizing, deduping and filtering have run; what those saw and dropped survives only in the run's log. The model boxes are outlined in the accent: they are the only stages that cost minutes rather than seconds, and the only reason the container reaches the host at all.
02 · CLUSTERING

Clustering never reads a word.

Every headline and summary becomes a 256-dimensional vector from minishlab/potion-base-8M, a static embedding model: a lookup table distilled from a transformer, so embedding a whole day takes a fraction of a second on a CPU and needs no network. Agglomerative clustering with average linkage then joins any two groups whose mean cosine distance is under 0.42. No language model decides what belongs together. Geometry does, with one outside signal: when Google News lists several outlets under one story, those articles are kept together as a unit, and two stories Google keeps apart are never merged. Outlets word the same event too differently for headlines alone to find it.

03 · RANKING

Independent reporting outranks volume.

A single outlet posting forty times about something is not forty confirmations. So outlets count linearly, articles only logarithmically, and freshness decays by half every 12 hours. An outlet's own credit fades too: it halves for every 24 hours since that outlet last touched the story, and each article's count fades the same way, so a story newsrooms have moved on from slides down the page instead of holding its place and then disappearing. There is no fixed lifetime: a story ends when what is left of its outlets' credit is less than a quarter of one outlet. And outlets that ran the same piece count as one piece of reporting: thirty papers of one owner, or a wire story republished under lightly edited headlines, earn the first outlet's credit in full and a quarter for each copy, because independent reporting is what breadth is meant to measure. A rank is shown with its range: resampling the outlets that carried each story two hundred times and ranking again gives the span a rank would move across if a few outlets had or had not happened to carry it. Near the top the ranges are tight; around #25 they usually cross the line, and What Moved marks those moves as within the margin. Breadth is counted among the outlets this record reads, which are mostly English-language and mostly American and British: a story big in Lagos or Jakarta but little covered in English is ranked low here. Only the top 100 clusters are ever written up; everything below the cut is still clustered, and simply never reaches the model.

score = 3·Σoutlets 0.5hours since its latest / 24 + 2·ln(1 + Σarticles 0.5hours / 24) + 4·0.5hours / 12

TRY IT — THE SAME FUNCTION, IN THE BROWSER

score — = — + — + —

Drag the article count to sixty with one outlet and the score barely moves; add a second outlet and it jumps by 3. That asymmetry is the whole design. Here every outlet and article is taken to be as old as the newest article; in a real story each ages from its own time.

0 2 4 0h 12h 24h 36h 48h half-life
The recency term across the 48-hour window. Hover the curve to read it; the dot follows the slider.

22 Nov'S TOP 8, EACH STORED SCORE SPLIT BACK INTO ITS TERMS

  • source diversity
  • volume
  • recency
  1. 01FBI crime database and USDA payment system reportedly breached 113.53
  2. 02ASEAN Leaders Sign Regional Digital Economy Agreement at Manila Summit 107.71
  3. 03Russian cargo ship hit by drones in Baltic Sea, prompting Finnish res… 107.02
  4. 04COP31 ends without overarching deal after fossil-fuel dispute 104.47
  5. 05Trump Says Xi-Putin Summit Addressed Korean Peninsula and Ukraine 100.37
  6. 06Nepal PM Balendra Shah begins first official visit to India 95.55
  7. 07COP31 ends without consensus after fossil fuel and finance disputes 94.94
  8. 08ICE arrests two passengers on Seattle airport jet bridge, drawing cri… 83.02
The checkpoint stores the total and the outlet breadth. The outlet and volume terms follow from those, so what remains is the recency term — and inverting the half-life recovers how old the newest article was when the run scored it. Hover a segment.
As a table
storyoutletsarticlesdiversityvolumerecencyscore
FBI crime database and USDA payment system reportedly breached8186105.097.41.03113.53
ASEAN Leaders Sign Regional Digital Economy Agreement at Manila Summit23979997.517.762.44107.71
Russian cargo ship hit by drones in Baltic Sea, prompting Finnish res…566395.647.384.0107.02
COP31 ends without overarching deal after fossil-fuel dispute495592.967.514.0104.47
Trump Says Xi-Putin Summit Addressed Korean Peninsula and Ukraine37285789.088.083.21100.37
Nepal PM Balendra Shah begins first official visit to India439583.538.063.9695.55
COP31 ends without consensus after fossil fuel and finance disputes667984.327.023.6194.94
ICE arrests two passengers on Seattle airport jet bridge, drawing cri…363772.586.573.8783.02
04 · WHAT THE MODEL READS

Eight articles, one per outlet first.

A 42-article cluster once timed out every story it touched: the prompt was too long for a CPU to read in three minutes. So the model gets at most 8. The newest article from each outlet goes in first, because a second outlet adds evidence where a tenth headline from the same one adds nothing; whatever room is left goes to the newest articles overall. Each article arrives as its headline, its feed summary, and, where the publisher allowed it, the first 3000 characters of the article. The editor reads exactly the same material, so it cannot cut a true fact as unsupported for having seen less than the writer.

Referendum talk · 18 articles from 7 outlets

Dawn The Express Tribune Taipei Times RTÉ News dunyanews.tv focustaiwan.tw irishnews.com 13 Nov 14:14 UTC 22 Nov 02:36 UTC · published →
  • round 1 · newest from its outlet · 4
  • round 2 · filled by recency · 4
  • never sent · 10
A real cluster from 22 Nov, one row per outlet, each article at the time it was published. Shape as well as colour marks the round, so it reads without colour. What never reaches the model still counts: every article is in the story's citations, and in the ranking above.
05 · THE GATE

The writer cannot publish its own work.

Every story goes to an editor: a different model, told what counts as a material problem (a claim no source makes, a fact contradicting a source, a claim credited to the wrong outlet, significance or stakes the writer added) and what does not (paraphrase, rounding, word choice). It publishes the story, revises it, drops it as not news, or holds it for a person. A revision may only cut, add attribution, or correct a detail to match a source, and application code rejects any revision containing a name or number found in neither the draft nor the sources. The writer's draft is kept beside every revision. If the editor cannot be reached, the older pass-or-fail judge decides instead.

16 Nov 377
17 Nov 380
18 Nov 429
19 Nov 449
20 Nov 462
21 Nov 473
22 Nov 484
  • published as written · 2599
  • published after the editor's cuts · 442
  • dropped as not news · 12
  • held for a person · 1
  • decided by the older judge · 0
3054 stories across the last 7 run days, as each day ended. Of the stories the editor decided, it revised 14%. Grey is the period before the editor existed, when a pass-or-fail judge and the writer's own confidence decided, and most stories it held were later approved unchanged by a person. What is not measured yet: how often a story the editor passes still carries an error.
06 · IDENTITY

A story keeps its name while it grows.

A cluster is named by hashing the articles it was born with, and the name is then stored. It used to be recomputed from current members — which meant one new article made it a different cluster, and the story, its review decision and its public URL all quietly stopped matching. A later run the same day now attaches new articles to the nearest existing cluster instead of re-clustering, so this happens constantly and on purpose.

- Pragativadi I Latest Odisha News in English I Breaking News · 929 articles now

STORED ID · WHAT EVERYTHING IS KEYED TO

05d11eb80cdfc3117badc42439e457c5d4854819e3469a0ad90abc0d6ea7603e

→ /stories/2026-11-22/05d11eb80cdf/

≠

SHA-256 OF TODAY'S MEMBERS · WHAT IT WOULD BE IF RECOMPUTED

6906aa95e3f0e43dfdca02f116330ddccf31196eb4ce459a5e2afaafc737e399

This cluster gained articles after it was named. Under the old rule it would now be a different cluster with a different URL; the stored id is why it is not.
CLUSTERS ON 22 NOV
7369
GREW AFTER THEY WERE NAMED
3321
STORIES WHOSE PROMPT HAS SINCE CHANGED
189 / 484

Only these are rewritten on the next run. A cluster can grow without changing the 8 articles the model reads, and then the prose it wrote still stands.

07 · THE SCHEDULE

Four runs, and room between them.

The scheduler registers its jobs before it runs anything, so a failed run cannot take the schedule down with it. Only one run happens at a time; a run that is still going when the next is due makes that one wait, up to six hours, rather than silently skipping it. Each band below is as wide as a real run takes — read off its own files, not estimated.

00 00:00 03 03:00 06 06:00 09 09:00 12 12:00 15 15:00 18 18:00 21 21:00 24 00:00 00:00 06:00 12:00 18:00 host watcher · polls every 30s, imports each run as it lands
Upper times are UTC, lower are UTC. ARGUS_SCHEDULE_HOUR=0,6,12,18, container time (UTC). Measured run lengths: 18 Nov 26 min · 19 Nov 27 min · 20 Nov 28 min · 21 Nov 23 min. Restarting the container does not start a run; runs happen on this schedule or by hand. The public site is the same app on the public server, serving each page as it is asked for; it shows a run once that run's files have been copied there, at the end of each run.
08 · ANATOMY OF A RECORD

Everything above, in one line of JSON.

Files are the source of truth; the site's database is a cache that can be deleted and rebuilt from them. This is one real story as the pipeline stored it — line 450 of 620 in data/stories/2026-11-22.jsonl, pretty-printed and trimmed. Stories are appended, never rewritten, and the last line for a cluster wins.

{1  "cluster_id": "5c960b148499ee493cb1e17b370bd8784ba06d6abffd411100d666ff0867d432",2  "title": "FBI crime database and USDA payment system reportedly breached",3  "body_markdown": "The FBI is investigating an apparent breach of the National Crime Information Center, a nationwide database used by federal, state and local law enfo…",4  "citations": ["https://www.reuters.com/world/us/fbis-criminal-information-database-hacked-cnn-reports-2026-11-21/", "https://www.sconeadvocate.com.au/story/9395719/fbis-criminal-information-database-hacked-report/", "… 72 more"],5  "review_status": "published",6  "synthesis_backend": "openclaw:openclaw/argus-writer",7  "synthesis_confidence": 0.88,8  "eval": {    "cites_sources": true,    "distinguishes_fact_from_synthesis": true,    "no_unsupported_claims": true,    "notes": "editor publish: The story is faithful to the source material and cont…",    "passed": true,    "judge": "openclaw:openclaw/argus-editor"  },9  "editor": {    "decision": "publish",    "reason": "The story is faithful to the source material and contains no material…",    "issues": "0 quoted problems"  },10  "prompt_sha": "09317250cfba79a9f85ba449ae16155c327cd2e1149f0e897e59ffd152b06c93"}
  1. 1cluster_id

    SHA-256 of the sorted article ids the cluster was born with. Stored, never recomputed, so the cluster can gain articles without changing identity. Its first 12 hex characters are the story's public URL.

  2. 2title

    Written by the model under a JSON schema. A reviewer's edit never touches this record: it lives in data/reviews/ and the database.

  3. 3body_markdown

    As long as the material supports and no longer: a single thin source makes a two-sentence brief. Every claim must be in the sources, attributed to the outlet that reported it; the writer may not add stakes or significance.

  4. 4citations

    Every article URL in the cluster, not only the ones the model read. The cites_sources check is exactly that this list is non-empty.

  5. 5review_status

    The gate's verdict: published, needs_review, dropped or stub. Decided in application code from the editor's decision (or the judge's, when the editor is unavailable); there is no path for the writing model to publish its own output.

  6. 6synthesis_backend

    Who wrote it: the writer agent on the model gateway, the local model when the gateway is down, or stub when nothing answered. A stub is a headline list, not prose.

  7. 7synthesis_confidence

    The writer's own 0–1 estimate of how complete and supported the story is. Self-reported and uncalibrated. The editor reads it; it no longer decides publication on its own.

  8. 8eval

    The verdict in the form the review app reads: the editor's decision and every material problem it found, quoted. "judge" names who decided.

  9. 9editor

    A different model's decision (publish, revise, reject or review), its reasons, and, when it revised, the writer's original draft. Code refuses any revision that adds a name or number the sources do not contain.

  10. 10prompt_sha

    SHA-256 of the article ids that reached the prompt, capped at 8 and taken one per outlet first. A later run rewrites the story only when this changes.

Every figure above was read from the pipeline's own checkpoints when this page was published, and the constants — weights, thresholds, the prompt cap, the confidence floor — are imported from the modules that use them, so this description cannot quietly drift from the machine it describes.