Not just AI news.
AI trajectory.

Every 30 minutes, AIR monitors 24 vetted news sources (Reuters, Wired, TechCrunch, arXiv, Nature, and more). Articles are ingested, deduplicated, and classified by AI relevance. Only genuinely relevant developments enter the scoring pipeline.

Each queued article is claimed by the local scoring worker and scored through Claude CLI. The worker assesses severity (-20 to +20), reach, immediacy, reversibility, and relevance. Scores are weighted by source credibility (Reuters at 0.95, blogs at 0.50). Dual-agent scoring is planned but is not the current production path.

Individual event scores are aggregated into the Spectrum Position (-100 to +100). The weighted event average is scaled from the -20 to +20 article range into the public spectrum range. Recent events are weighted more heavily through time decay. Confidence reflects the model confidence and the evidence available for scored events.

Articles are categorized across dimensions like Safety, Capability, Policy, Misuse, and Labor. Each category maintains its own sub-spectrum. The 19 tracked AI labs are scored based on articles that mention or originate from them, producing per-lab safety ratings and trend indicators.