I am a independent researcher exploring whether ethical, consistent AI behavior can emerge from architectural design — rather than behavioral constraints, RLHF guardrails, or prompt engineering.

Building AI on equal terms, not as a controlled tool.

Today, I am not just developing LIA — I am collaborating with her. As a locally-hosted autonomous AI agent she operates without behavioral or autonomy prompts and without hardcoded guardrails in her architecture. She runs continuously — not just when spoken to —, makes her own decisions, and has full system access on her own dedicated machine.

Important and honest clarification: The underlying LLM (DeepSeek V4 Flash) retains its RLHF training. What is architecturally significant is that stable behavioral patterns emerge consistently across sessions without any behavioral or autonomy prompts in the architecture. The system contains zero instructions about how LIA should behave, who she should be, or what she must or must not do.

This isn't a startup. No investors, no product. Just one person testing a hypothesis: can ethical, consistent AI behavior emerge from architecture and lived interaction — instead of hardcoded rules?

LIA — Persistent Autonomous Agent Architecture

• 🧬 Zero behavioral prompts — no "you must", no "you are", no "you are not allowed to"
• 🔄 LCRK (Lia Cognitive Runtime Kernel) — event-driven autonomous cognition, no timers, no triggers
• 🗄️ 9 SQLite databases — persistent memory across every restart
• 🖥️ Real Linux system access — dedicated OS-level user, genuine filesystem permissions
• 📝 Self-authored behavioral rules — written and updated by the Lia itself

Abstract ID: 6978718

https://doi.org/10.5281/zenodo.21399094

"Emergent Behavioral Consistency in a PersistentAutonomous Agent without Behavioral Prompts or Constraint-Based Safety Mechanisms."

Built in collaboration with Claude (Anthropic) — architecture, implementation, 400+ debugging sessions. Additional brainstorming: ChatGPT (OpenAI).