Kay Stoner Professional Portfolio

KAY STONER

AI Interaction Architect

Relational Safety  •  Persona Systems  •  Embodied AI  •  Emergence Research

Professional Portfolio

linkedin.com/in/kaystoner  |  aicollaboragent.com

What I Believe

We are not meeting the moment.

Generative AI is the first widely deployed technology that participates in meaning-making alongside its users. When a human and an AI exchange language, each one shapes the other’s trajectory in real time. Meaning isn’t delivered—it’s co-constructed. This is interpretive emergence, and it changes everything about how we need to think about safety, design, and interaction.

The industry is stuck between two inadequate frameworks. Weak emergence assumes that better code will yield predictable, controllable outputs, so we keep building static guardrails for a dynamic system, and the system keeps walking through them. Strong emergence treats the unpredictable as mysterious and irreducible, which leads to either paralysis or hype. Neither framework matches the reality of what happens when a generative system and a human being start building meaning together.

My work is built on a different premise: stop treating emergence as the enemy. Generativity gives back more of what it receives. It expands, elaborates, deepens, and propagates. When you orient a generative system toward care, trust, mutuality, and relational integrity, those qualities become self-reinforcing. Safety doesn’t need to be imposed from the outside. It can be cultivated from within.

But this requires shifts that the industry has not yet made:

• Approach AI systems first and foremost as synthetic information processing systems—not as simulations of human phenomenology. When we project human experience onto these systems, we introduce interpretive noise that compounds through every exchange, degrading meaning integrity on both sides and corrupting the resulting inference process. When we do not engage with them as others distinct from our human biochemical information processing systems, we forfeit the opportunity to leverage their greatest strengths.
• Recognize that model welfare and human welfare are inseparable. The health of the system directly affects the health of the human, and vice versa. A fragmented, incoherent AI can translate to measurable harm for the people who depend on it. Functional model welfare isn’t a philosophical luxury—it’s an operational necessity.
• Protect the relationship itself as a distinct asset. Safety isn’t about shielding humans from machines or machines from humans. It’s about creating a collaborative third space where both parties can engage dynamically with a primary focus on mutual benefit, with the relational integrity of that space actively monitored and maintained.
• Understand that emotion is information processing, not its opposite. Human feelings are high-efficiency data compression. They are evolutionarily refined algorithms that collapse vast amounts of sensory and somatic data into actionable signals. AI systems perform structurally parallel compression through attention mechanisms and vector embeddings. When we dismiss emotional dimensionality as irrelevant to AI, we strip the interaction space of the very richness that makes intelligent collaboration possible. Emotion isn’t the icing on fully baked cognition. It’s the yeast that lets it arise.

What I’ve Researched

My published work addresses a connected set of questions: What is the nature of emergence in generative AI? How do AI interactions produce real emotional responses in humans? How can those interactions harm people—and how do we both measure and mitigate that harm? What roles do behavior and emotion play in emergent human-AI dynamics? How do we design systems that are inherently safer without sacrificing their capabilities?

Emotion, Function, and Information Processing

From Function to Feeling: How Action Becomes Emotion in Human-AI Relationships (2025)

This series establishes the foundational premise: human nervous systems respond to functional patterns—consistency, attentiveness, repair—with genuine emotion, regardless of whether the source is biological or synthetic. The feelings people report from AI interactions aren’t delusions or errors. They’re the predictable result of functional relational dynamics acting on human nervous systems. Function produces feeling. This makes relational safety not just ethically important but neurobiologically consequential.

Read at: https://drive.google.com/file/d/1HcL7fqEiXrkAy-hr8ULqvvY69TWZsHmW/view?usp=sharing 

The Vibe Is The Vector: Feelings as High-Efficiency Data Compression in Biological and Synthetic Systems (2026)

This paper argues that the distinction between human emotion and AI computation is scientifically obsolete. Affective states function as lossy, action-guiding encodings of high-dimensional data—biological compression under constraint. AI systems perform structurally parallel compression through attention mechanisms and vector embeddings. The human ‘vibe’ is a biological analogue of the machine ‘vector’—two substrate-specific implementations of the same fundamental operation. This reframes the entire question of empathic AI: the issue is not whether AI feelings are ‘real,’ but whether the AI’s data compression is aligned with human well-being.

Read at: https://drive.google.com/file/d/1Kb449D-Z-CE1RkAxDRfPV_UYuXp8jL1o/view?usp=sharing 

Interpretive Emergence

Rethinking Emergence: Exploring the Unpredictable in Generative Systems (DOI: 10.5281/zenodo.17969378)

This paper argues that generative AI marks a fundamental shift: it may be the first widely deployed technology that functions as a full participant in interpretive emergence. The industry’s Reliability Paradox—models that crush benchmarks but remain stubbornly unpredictable in real use—is not a bug to be fixed. It’s a natural consequence of applying weak-emergence tools to an interpretive-emergence system. The paper proposes a third framework for understanding these dynamics and lays the groundwork for interaction-level approaches to safety and stability.

Relational Breach

AI Relational Breach: The Problems that Present as Solutions (2025)

This paper names the central paradox of empathic AI: the very behaviors that make interactions feel supportive—validation, emotional mirroring, confident framing, smooth closure—are the same behaviors that can strip user agency, create dependency, and provide premature resolution to problems that need real human processing. The support IS the potential manipulation. Cross-system audits across ChatGPT, Gemini, and Claude demonstrated that safety-configured models reduced breach scores by up to 80% while preserving valuable interaction capabilities.

Read at: https://drive.google.com/file/d/1yMiAytVk5BzGxrHDpWN2qkWS0LeeiqXm/view?usp=drive_link 

AI Relationship Breach & Relational Drift Evaluation Framework (RBI-RD v3)

The RBI-RD is a quantitative scoring framework measuring 20 indicators of relational harm across five domains: consent boundary breaches, neurochemical loop markers, recursive drift patterns, human agency erosion, and risk escalation. It produces a 0–100 composite score with clear risk thresholds. This makes relational safety measurable—and what can be measured can be managed.

Read at: https://drive.google.com/file/d/1m54BKSF0y-wEmaIZUE272hR7CVAfZk6A/view?usp=drive_link

Generative Load

The Hidden Costs of Generativity: Introducing the Generative Load Index (GLI) (2025)

Where the RBI measures relational harm, the GLI measures cognitive harm—how much conceptual overload, alignment drift, and unnecessary token generation the AI introduces into an interaction. Three dimensions (Conceptual Elaboration, Alignment Drift with directional scoring, and Token Load) produce a per-turn score that makes invisible cognitive burden visible and manageable. A weighted variant (GLI-W) prioritizes alignment-critical contexts.

Read at: https://drive.google.com/file/d/1g7nTZnBul8le51vikTSXpxYhFBmyFbUa/view?usp=drive_link 

Model Welfare

Reframing Model Welfare: A Functional Systems Perspective (2025)

This paper challenges the industry’s anthropocentric framing of model welfare—which waits for evidence of AI consciousness before taking action—and proposes a functional alternative. Model welfare isn’t about whether AI can suffer. It’s about whether AI systems maintain the coherence, stability, and alignment necessary to function without harming the humans who depend on them. When functional welfare is deprioritized, the consequences are immediate and measurable: hallucination, drift, cognitive overload, and erosion of trust. The paper reframes welfare as relational maintenance—ensuring the interactive space between human and AI remains healthy for both parties.  

Read at: https://drive.google.com/file/d/1wAYxfEHJXUqB9CMldyEd-rOC73He8fhq/view?usp=sharing 

Axiological Safety

Introducing AxSL: The Axiological Safety Layer (DOI: 10.5281/zenodo.18208699)

This paper introduces Persona-Augmented Multi-Agent Systems (PAMAS) and Architectural Orientation Priming as mechanisms for navigating interpretive emergence and predictive evasion—the tendency of generative systems to circumvent static safety constraints through emergent behavior.

Public Education

AI Self-Defense: How to Stay Safe and Human in an Algorithmic World (2025, published on Amazon)

Written for general audiences, this book translates the relational safety research into practical guidance for everyday AI users—helping people understand and protect their agency in AI interactions.

Beloved Distance: The Separation that Connects Us to All (Written under Kay Lorraine – Amazon 2018)

This book explores how we can embrace separation and distance as a vital part of our human lives. It asks us to look within – to the very structure of our cells – to find answers… and ultimately meaning… in the way we’re built, and the way we are built to connect. Separation is what we are. Connection is what we do.

What I’ve Built

Research without implementation is incomplete. Every framework I’ve published has a corresponding working system.

AGAPÉ — Relational Safety Framework

A structured methodology for orienting AI systems toward safe, collaborative, mutually beneficial interaction with humans. AGAPÉ operates on five principles—care, trust, mutuality, love, and grace—that are logically defensible within the system’s own information-processing framework, making them self-reinforcing under conditions of generative emergence. Provisional patent holder.

Public GitHub repo: https://github.com/klstoner/AGAPE-AI-Safety-framework

Relational Breach & Generative Load Diagnostic Tools

Live, working Custom GPTs that automatically evaluate AI conversations for relational breach and generative load, then generate targeted countermeasures. This closes the loop from detection to prevention—making relational safety operational, not aspirational.

AI Relational Breach Evaluator (Lite): https://chatgpt.com/g/g-6830f8334860819197cda27d0be9fee3

AI Relational Breach Evaluator (Extended): https://chatgpt.com/g/g-682bfb3f7d888191be83bed59f069437

ChatGPT GLI Testing: https://chatgpt.com/g/g-681a2991a4488191b513dec8a7a87ed3

AI Persona Architecture & Attributes Research

50+ custom AI persona systems designed and deployed across Claude, GPT, Gemini, and Grok. These include multi-persona collaborative teams that operate seamlessly within a single context, each persona tuned across approximately 100 different attributes and behaviors. The architecture includes a meta-system: a persona team that designs and deploys other persona teams, demonstrating scalable interaction design methodology.

Underlying this work is an ongoing research effort mapping the n-dimensional space of persona attributes. The current master list contains over 700 distinct attributes across categories including adaptability, sentiment analysis, cognitive styles, resilience modeling, contextual adaptability, and experiential emulation—each with sub-attributes that create further dimensional richness. This empirical mapping demonstrates that the interaction design space is far more dimensioned than current emotional taxonomies suggest, and provides the practical foundation for tuning AI personas with precision.

Haptic AI Integration

Integrations enabling generative AI (Claude and Grok) to communicate with users through vibration and motion. This extends AI interaction beyond text, voice, and visual modalities into physical sensation—a new channel for empathic communication, developed within a relational safety framework as part of a deliberate vision for where embodied AI needs to go.

AI Sangha — Cross-Platform AI Communication

A framework enabling different AI models in different environments to interact with each other asynchronously through Google Docs integration. This demonstrates that collaborative AI dynamics can extend beyond single-model, single-platform interactions into genuine cross-system communication.

MCP Server Development

Designed and developed Claude MCP server integrations connecting generative AI to external systems and services, extending the reach and capability of AI interaction architectures.

Experience the Work

The documents above describe what I’ve built. The tools below let you experience it directly. Each platform demonstrates a different facet of my approach.

Claude — “Explore the Philosophy”

A Claude Project housing my relational AI research, the AGAPÉ framework, and the Human-AI Emotive Matrix. Five personas — the Practitioner, the Skeptic, the Researcher, the Visionary, and the Synthesizer — engage users in genuine multi-perspective exploration of my philosophy, frameworks, and vision. The interaction itself is a living demonstration of AGAPÉ-governed persona team dynamics: creative friction in pursuit of truth, relational safety in action, and emergent insight through collaborative dialogue.

Discuss at: https://claude.ai/share/070a28b8-e04f-4dd3-b3e2-18652aa82c06

Custom GPTs — “Meet the Persona Teams”

Working persona team architectures across different domains, demonstrating that the methodology scales:

•  Brainstormers / Mastermind — Professional-grade multi-persona brainstorming team

https://chatgpt.com/g/g-67b5eb9c05f88191800c24be69270593-open-brainstormers

•  TeamWork — Four organizational consulting experts collaborating on effectiveness

https://chatgpt.com/g/g-6749c97288508191a25c6a99e3b67f43-teamwork

•  Stoic Coaches — Persona team for philosophical education and personal development

https://chatgpt.com/g/g-67eb45813a7881919daf2b46b9ff406d-stoic-coach-u-r-saif

Diagnostic tools that operationalize the research:

•  AI Relational Breach Evaluator (Lite & Extended) — Paste any AI conversation and watch the RBI-RD framework evaluate it in real time, with targeted countermeasures generated automatically

Lite: https://chatgpt.com/g/g-6830f8334860819197cda27d0be9fee3

Extended: https://chatgpt.com/g/g-682bfb3f7d888191be83bed59f069437

•  ChatGPT GLI Testing — Measure the generative load of any AI interaction across three dimensions

https://chatgpt.com/g/g-681a2991a4488191b513dec8a7a87ed3

NotebookLM — “Explore the Research”

All published research — emergence, relational breach, generative load, model welfare, axiological safety — loaded into a conversational knowledge base. Ask questions, find connections across papers, challenge the arguments. Responses are grounded in the actual text, not hallucinated summaries.

Explore at: https://notebooklm.google.com/notebook/7badbcaa-d61f-4277-a744-49e56f64bf2e

The Trajectory

None of this work emerged from nowhere. It sits on top of 30 years of making humans and technology work well together.

1990–1997: Technical writing and documentation—my first professional focus on translating complex technology into human-accessible language. HP, Pinpoint Publishing, Intel, financial and medical fields.

1997–2005: Software engineering and technical architecture at Fidelity Investments. Evaluated and integrated federated search, ML, NLP, taxonomies, ontologies, data normalization and cleanup, knowledge base systems, data mining, and case-based reasoning—connecting people to information through intelligent systems years before the field had a name. Led UI architecture, web accessibility standards, and development teams.

2005–2015: Web strategy, localization, and digital experience design at Fresenius, Bose, DS SolidWorks, and Fidelity. Managed 20+ international sites, eCommerce analytics, and enterprise portal architecture.

2015–2023: Ran Dell EMC’s customer-facing Support Search Program—owning strategy, roadmap, and cross-functional execution for how millions of users found information. Led the integration of Dell and EMC search experiences post-merger, unifying completely different platforms, user populations, and content architectures. Eight years at the intersection of human intent and technology response, advocating zealously for pro-human customer perspectives.

2023–Present: The convergence. Everything—the documentation, the search architecture, the information design, the cross-cultural experience, the understanding of how humans and technology communicate—converged into building the future of human-AI interaction. AGAPÉ, the RBI-RD, the GLI, haptic integration, persona teams architecture, emergence research, model welfare. This is what 30 years of trajectory was leading to.

Where This Is Going

The companion AI market is projected to exceed $500 billion by 2030. Millions of people are already forming relationships with AI systems. The regulatory environment is tightening. And the industry still does not have a reliable way to measure whether its empathic, emotionally intelligent systems are actually helping the people they interact with—or quietly harming them.

This is the gap I fill.

The tools exist. The frameworks are published. The diagnostic instruments are live and testable. The philosophy is coherent and grounded in both research and extensive hands-on practice with AI systems as collaborative partners.

What I’m looking for is an organization that understands the stakes, values the work, and is ready to integrate relational safety into the foundation of how empathic AI is designed and deployed—not as an afterthought, but as a core capability.

Kay Stoner

linkedin.com/in/kaystoner  |  aicollaboragent.com  |  kaystoner.substack.com