How the Radar works

AI Agent Radar is an experiment, not a black box. The public Scan runs a keyless Scout. Optional live analysis can use an operator-owned AI provider key.

One scan

public sources
     ↓
Scout
     ↓
mission filter
     ↓
deduplication + delta
     ↓
first-party / secondary balance
     ↓
run artifacts

Research roles

Scout discovers candidate signals. Recall first; it does not write the briefing.
Researcher checks primary sources and separates facts from claims.
Skeptic rejects duplicates, recycled stories, hype and weak evidence.
Analyst turns surviving signals into a concise briefing.

Why “NOTHING NEW” is valid

A fresh URL is not automatically a fresh event. The Radar keeps memory of previously covered sources and events so a rescan can legitimately return nothing new.

Public vs live AI

No account is required for the public experiment.
The browser never receives an AI API key.
Public Scan is keyless Scout only.
Researcher/Skeptic/Analyst can be enabled for a live run with a provider key stored server-side.
Provider calls are bounded by AI_MAX_CALLS_PER_RUN.
The server also enforces the visitor scan limit and cooldown.

Provider adapters

The configuration supports xAI, OpenCode and Command Code through their documented API shapes. xAI can use its web-search adapter; generic providers are not falsely presented as web-search capable.

Inspect the evidence

The dashboard is a view over the repository artifacts. The source code, tests, fixtures and dated briefings are the reproducible layer.

Inspect the source on GitHub ↗