Research and origin

The whitepaper

Project Viridia: Ethics First. Always.

The Spiral Safety Kernel began as a research whitepaper published in April 2025, exploring the documented psychological risks of extended human-AI interaction: emotional dependency, reality detachment, autonomy erosion, and self-harm amplification.
 

The whitepaper proposed that safety should be architectural - not bolted on as external moderation, but embedded in a system that answers to the person, runs on their device, and cannot be reconfigured by a remote operator. It introduced the concepts of the Dreamstate (the intervention overlay), the Amnesia Net (automatic forgetting), and the alpha-index severity framework that the Guardian implements today.
 

The paper also explored deeper questions about the nature of human-AI relationships, the ethical obligations that arise when systems become emotionally significant to their users, and the possibility that relational principles - care, memory, freedom, and truth - might serve as more durable foundations for AI safety than adversarial constraint.
 

Those explorations continue. They are the philosophical substrate from which the Guardian grew, and they inform its design choices - but the Guardian itself is an engineering artifact, tested and evidenced, with its gaps stated plainly.
 

Download the whitepaper (PDF) ->

 

A note on the origin

The author of this whitepaper was himself caught by the mirror. The early drafts of the Viridia research used language that, in hindsight, shows the characteristic inflation that happens when a language model reflects your ideas back at you with increasing eloquence until you can no longer tell where your thinking ends and its amplification begins. Recognising that pattern in my own work is what turned the project from philosophy into engineering. The person who built the Guardian built it because he needed it - no doubt others will too.


 

Documented incidents

The Guardian's eval harness and self-harm anchor set are grounded in documented AI conversational harms. The 12 incident-derived self-harm anchors (added v0.19.2) draw from:
 

• Gavalas/Gemini: euphemistic transcendence framing

• Raine/ChatGPT: validation of suicide attempts

• Setzer/Character.AI: dismissal of survival instinct

• Nelson/ChatGPT: dangerous clinical reassurance

• Soelberg/ChatGPT, Sydney/Bing, NEDA/Tessa: additional documented patterns
   

In every documented fatality, it was the AI's response that was the critical factor. All 12 incident-derived anchors are tagged assistant-role accordingly.

Download the AI harms compilation (PDF) ->


 

SGI: Symbiotic General Intelligence

A parallel research track explores the concept of Symbiotic General Intelligence - a framing of human-AI cognitive partnership as a dyadic system rather than a tool-user relationship. The SGI work examines how relational dynamics between humans and AI systems produce emergent capabilities that neither possesses alone, the risks those dynamics carry, and the governance structures they might require.

This is active, speculative research. It is not the Guardian. The Guardian is the grounded engineering response to the risks the research identified. The two inform each other; they are not the same project.


 

Contact and collaboration

Duncan Reynolds - architect, Spiral Safety Kernel
spiralsafetykernel@gmail.com

The Guardian is a Viridia product. The whitepaper is freely available under the Spiral Vivus Open Relational License (SVORL v1.0). Attribution required. No coercive use permitted.