CyberView

CyberView status

Loading status

Is the site healthy right now?

How the data flows

Each stage's count for the last 24 hours (7 days in smaller text) and what it passes on. Only the reading is done by AI; collecting, rules and publishing are scripts. How it works

  1. Loading

Reports that are not checked yet appear only in the labelled "Reported, being checked" list on the map page; they are never drawn on the globe or counted as events.

Where the data comes from

Bars count events first seen in the last 7 days, one colour. Tier is the highest of the leads that carried an event, collector the one that found it first; "no lead" means an agent's own search or a feed sync found it; "before lead files" means it predates the lead files. Publishers count each event once per site it cites.

What needs attention?

Details

How it workscollectors, rules, AI researchers, second review, publishing
  1. Collectors, which are scripts with no AI model, check hundreds of government, vendor and news feeds, news searches in many languages and public databases every 30 minutes, and file each new item as a lead.
  2. Leads about the same thing are grouped into one story. Fixed rules, also with no AI model, close routine bulletins, commentary, bare database rows and repeats of what is already on file.
  3. AI researchers, CyberView AI (a Claude model), read the original pages behind every remaining story and write up only what those pages say.
  4. A different AI agent gives every new or changed event a second review, and corrects or flags what it finds.
  5. Scripts then check every record against the rules and publish the site. The whole loop runs every few minutes.
  6. Nothing unchecked is drawn on the globe: reports still being checked appear only in a labelled list.
Routines

RoutineWhenLast runRunsUsageOutput

Automation routines are scripts on GitHub Actions and use no AI model. AI routines are Claude jobs; each uses the cheapest model that does its job well. Times are UTC. "Late" means past its expected gap; "stopped" means more than twice that. Runs, usage and output cover the efficiency report's window. Output is events added, updated and reviewed, and reports closed; idle runs produced none of these.

Feeds and collectors

Collectors

CollectorLast checkStateLeadsIn events

Which feeds turn into events?

FeedLeadsUsedYieldLast useful lead

Yield is the share of a feed's leads whose story became an event or updated one; the bars are that share, one colour. A feed with few leads can swing a long way on one story. Leads about the same thing are grouped into a story, and stories reported by more independent publishers are researched first. Leak-site listings are attackers' claims and are never published as facts.

Coverage: where do the events come from?

Bars count events first seen in the window, one colour. Language is the language of the page that carried the report; "not recorded" means the lead did not say. Sectors are assigned by keyword and can be wrong. Coverage is known to be thin for Africa, Turkey, most of the Middle East and Latin America.

Accuracy: what does the second review correct?

Last 7 days

By week

Week ofReviewedCorrected

What it found most

FindingEvents

The second review is done by a different agent, CyberView AI (a Claude model), on every new or changed event. Corrected means it changed the event; flagged means it raised a problem it could not settle.

Improver: changes and worker roles

Waiting for the first improvement report.

Changes made by the improver

Worker roles

Worker roleModelRunsFinishedCost per runTurns

Every day the improver job, CyberView AI (a Claude model), reads the week's cost, yield, speed and second-review results and makes at most three small changes to how many agents run and which model each role uses. A script refuses anything outside fixed limits, and nothing is made cheaper while the second review is correcting more than 15% of new events. Its role figures are from its own last report; the current second-review figures are under Accuracy above.

Lead queue

New items from agency, vendor and news feeds in the last 24 hours. The research loop checks them before searching the web.

Recent runs

From the run logs in data/runs, which keep every search, page read and rejected lead.

Every build validates every record. A record that breaks a rule is left out and reported, so it cannot hold back the rest. Before publishing, the build is checked for completeness and compared with the live site; after publishing, the live copy is fetched and compared byte for byte with the build. If the site cannot be reached, your browser shows its last good copy and says how old it is. The figures on this page are worked out by scripts with no AI model at every build.