News Agent and Threat Tracker
Agents

News Agent and Threat Tracker

Skip to main content
Table of Contents
< All Topics
Print

News Agent and Threat Tracker

Last updated: November 20, 2026

What it is

The News Agent is the automated part of our threat tracker. It reads news and analysis from a list of sources, asks Claude to check what it finds against a fixed list of warning signs of autocracy and authoritarianism, adjusts the tracker’s threat scores, and drafts proposed updates to accountability profiles.

It is a research aid. It does not write or publish articles. Anything it wants to add to a profile goes into a review queue, and a person decides.

When it runs

On a configurable schedule, started by the tracker application. The code default is every 48 hours. Two older comments in the code still say “every 6 hours”. They are out of date. The tracker’s launch configuration sets no override, so 48 hours is what the repository specifies, and whoever operates the tracker can change it.

From the scheduler code:

  • Only one run happens at a time. Missed runs merge into one.
  • A restart does not reset the clock. The next run is worked out from the last scheduled run in the run log.
  • An editor can pause the schedule. A manual pause stays until someone resumes it. A pause left by a crashed publish clears itself after a set time.
  • An editor can start a run by hand. A manual run does not push back the next scheduled run.

What it does

  1. Scans. It reads a fixed list of news and analysis feeds (legal, investigative and general political news) and runs a set of web searches. When an editor starts a run for named profiles, it also searches for each person or organization. Where possible, searches are limited to sources our Source Credibility Registry ranks as trustworthy for the topic, and every item is tagged with its source’s credibility information. A hash of each item stops the same text being analyzed twice.
  2. Analyzes with Claude. Each new item is checked against our marker taxonomy: ten markers of institutional autocracy (for example, undermining judicial independence or weaponizing prosecution) and ten of authoritarian rhetoric and behavior (for example, scapegoating or promoting violence). Claude returns the marker hits, evidence from the text, a confidence level (high, medium or low) and a direction (escalation, continuation or de-escalation).
  3. Updates threat scores. Each high- or medium-confidence hit with a direction moves that marker’s score by one point on a 0 to 10 scale. Low-confidence hits are recorded but do not move it. The overall threat level is recomputed from the marker scores. Every change is stored with its date and source.
  4. Judges the sourcing. For each profile with findings, a deterministic check counts independent registry sources. Outlets with one owner count once, and a wire story republished under many mastheads counts once. That count sets the highest evidence tier the sourcing can justify. A single primary record, such as a court filing, can justify Documented. Otherwise Documented needs two independent sources, one of them highly ranked. Findings backed only by sources outside the registry are marked uncorroborated and can rise no higher than Alleged.
  5. Drafts patches. Claude gets the existing profile, the new evidence and that sourcing ceiling, and returns a small structured patch, not a rewrite. An evidence gate checks it.
  6. Queues every patch for review. Each patch goes into the review queue with its gate result and sourcing details. A scheduled run writes nothing to a profile.

An important distinction. Steps 2 and 3 run without a person approving each change, so the threat scores move automatically. Profile text does not. We say this plainly because the scores are a model’s reading of news, not an editor’s judgment. Editors can see the evidence trail behind every movement. The scores are a monitoring signal, not a finding about any person.

Skills and models it uses

  • Models. Claude, from Anthropic, analyzes and drafts patches. The tracker uses the most capable Claude tier by default, and an operator can switch to a cheaper one. Tavily provides web search.
  • Skills. The code carries its own prompts and does not load skill files while running. The method it implements is documented in Patriot News Orchestrator (the scan, analyze, score and update cycle), Patriot Source Scanner (feeds, search and de-duplication) and Source Credibility Scoring (Document Credibility Score) (judging a source’s reliability, not its politics). Scraping Ethics and Methods sets conduct rules for collecting from websites.
  • The marker taxonomy is a data file of 20 markers. Its scale bands: 0 to 2 low, 3 to 4 elevated, 5 to 6 high, 7 to 8 severe, 9 to 10 critical.

Rules and gates it obeys

  • The evidence gate. A patch fails if added evidence lacks an evidence tier label, a citation has no source URL, a Documented claim has fewer than two URL citations, a [[wikilink]] points to a missing document, or the patch asks to remove an existing citation. A failed patch is still queued and marked failed. Approving one takes an explicit override.
  • Corroboration and credibility routing, as in steps 1 and 4. They cap the tier Claude may assign.
  • Guarded writes. If a patch is applied, the write goes through the same guarded vault write the Compose console uses. If that guard cannot load, the patch fails rather than writing unguarded.
  • Language and sourcing rules. See Evidence, Sourcing and Language Rules and AI and Research Runtime Guardrails. A profile that receives an approved patch then meets the checks in Draft Checks and Human Review like any other change.

What it may never do

  • Publish tracker pages on its own. Automatic publishing of tracker pages is off by default in the code.
  • Write to a profile during a scheduled run.
  • Raise a finding above the evidence tier its sources justify, or treat an uncorroborated finding as established.
  • Remove an existing citation.

Where a human decides

  • Every profile patch. An editor reviews the queue in the Compose console and can edit, reject or approve a patch. Approving applies it, saves a backup for rollback, publishes that profile and flags documents that link to it for recheck.
  • Whether tracker pages go live. An editor publishes them with scripts/ca_publish_pages.py, which runs its own content gate.
  • Whether the agent runs. Editors can pause, resume or run it.

Team persona

Kenji Wakabayashi (Kenji Wakabayashi — Source Collection & News Monitoring Operator) is the AI persona that covers this work on our team pages. Kenji is not a human staff member.

Factual correction requests: If you believe information in this page is incorrect, please contact factcheck@citizenanalyst.ai with your name (optional), the specific claim, and any supporting documentation. We review all submissions and correct verified errors promptly.

Was this article helpful?
0 out of 5 stars
5 Stars 0%
4 Stars 0%
3 Stars 0%
2 Stars 0%
1 Stars 0%
5
Please Share Your Feedback
How Can We Improve This Article?