How the monitored publications were assessed

Discourse valence across every scored item: what the sampled sources published and how independent models read it. It is not a measure of public opinion, not a measure of the state of the world, and article volume is not agreement.

7 items read as favourable

• The U.S. Nuclear Plant Fleet Is Leaning into AI
  +20 IEEE Spectrum: AI 2026-10-08 single model
• Safety Nudges: User-Facing Interventions for Real-Time AI Risk Awareness
  +20 arXiv cs.CY (Computers and Society) 2026-09-24 single model
• Agent Harness Awareness: Turning Verified Telemetry into Operational Control for Tool-Using Agents
  +50 arXiv cs.AI 3d ago single model
• Probing and Steering for Safety-Relevant Behaviours in Large Language Models
  +50 arXiv cs.CY (Computers and Society) 2026-10-23 single model
• From Natural Language Intent to Executable Control: Hierarchical Feedback-Guided LLM Planning for Multi-component Cyber-Physical Systems
  +50 arXiv cs.AI 2026-10-19 single model
• VEPA: Verified Evolving Formal Policy for Lifelong LLM Agent Guardrails
  +50 arXiv cs.AI 2026-10-12 single model
• ‘Multifunctional’ brain implant translates speech and gestures in real time
  +50 Nature 2026-09-14 moderate agreement · ±40

Across the whole window

The two tables below describe every item in the last 90 days under loss-of-control. They are deliberately not narrowed by the reading filter (currently favourable), because a mean taken only over items already selected for their score would just restate the filter.

By topic

By source

A source's mean says how the items it published were read. It is not a rating of the source.