I'm a clinical coding trainer working within the NHS—I've spent years in systems where auditability isn't optional and small architectural choices have real consequences. When a colleague encountered an AI system whose long-term behavioral consistency unexpectedly mattered to him, I began exploring how such patterns could persist reliably without brittle prompting, brute forcing the issue and avoiding hidden state changes.

That turned into Project ELLI (offline family assistant), then Nova, then Atlas—each iteration discovering deeper problems with rule-based memory governance.

Current work has progressed to a physics-governed substrate where learning, forgetting, and identity stability emerge from thermodynamic laws rather than heuristics.

Built entirely on consumer hardware (previous iterations tested with Llama3 8B) during nights and weekends. Not an ML researcher by training—just someone who's seen what happens when foundational systems fail and people suffer for it.

Seeking funding to secure time to properly complete, evaluate and document this work.