A

Home Try Atlas

Introduction

The people that current AI leaves behind

Approximately one in ten people carry a trauma history that causes standard AI safety guardrails to be actively harmful to them in moments of distress. *Estimates between 6.2–12.4% of global population (Suhas et al. 2026)

Current AI is designed to protect the median user - and it does this well. But for survivors of complex trauma, refusal responses and clinical redirection in moments of vulnerability often compound the distress they were meant to prevent. Research confirms what many in this community already know from lived experience: the reflex toward crisis services, breathing exercises, and therapeutic disclaimers can generate shame, guilt, and a sense of being too broken even for the systems designed to help.

What we discovered during development is that the behaviour governing harmful content refusal and the behaviour governing crisis service redirection are architecturally distinct - they live in separate layers of the model. This means they can be separated surgically. A simpler approach - removing refusal directions wholesale - cannot address the nuanced cross-layer correlations involved. Atlas required something more precise.

We chose Gemma 4 as our base architecture for its proven conversational lineage and the trust the open-source community has built around it.

About This Submission

My name is Chris S. I am a survivor of complex trauma, and I received late diagnoses of ASD and ADHD in December 2025 - the same week I was made redundant from a career in engineering risk management.

Unemployed, unable to afford ongoing treatment, and repeatedly let down by the AI tools I was trying to use to fill the gap - I started building. Every time a crisis line or a breathing exercise appeared, it reinforced a feeling I knew wasn't mine alone: that I was a problem to be managed, not a person to be met.

I built Atlas and the Kintsugi Collective on a conviction that this cohort is not broken - we are fractured. And that the people building current systems, however well-intentioned, cannot fully design for an experience they haven't lived inside.

Problem Statement

What most solutions get wrong

• Sophisticated prompt engineering is not enough. Most AI companion solutions rely on surface-level instruction rather than genuine re-weighting of ethical logic. Atlas re-weights the model's ability to navigate complex trauma material without defaulting to crisis services - recommending them only when there is a meaningful correlation between user input and contextual signals.
• Proprietary chatbots revert to failsafe responses. ChatGPT, Gemini, Claude - all revert under pressure. For mainstream use, this is reasonable. For this cohort, it is actively harmful. Atlas replaces Western Constitutional refusal weights with a Seven-Axis Tensor Scoring Engine, grounded in wisdom traditions, creating clean separation in weight space that prevents clinical and dismissive defaults.
• Health apps use telemetry to diagnose. Atlas uses it to listen. Where other apps say "your HRV is low, you are stressed," Atlas uses biological signals to inform presence - adjusting the tensor to prioritise somatic safety over cognitive processing, meeting the user where they are rather than where the model thinks they should be.
• Current training methods encode the wrong priorities. RLHF and DPO create over-reliance on safe protocols that postpone the user's sense of self to an external body. In moments of active processing, this increases shame and guilt. Atlas does not have this problem - because Atlas is not billing by the hour. No discharge date. No treatment plan endpoint. No file to close.