2 Home

Claude edited this page 2026-02-23 19:41:13 +00:00

Table of Contents

• Lethean Ethical Model (LEM)
• Key Concepts
• Wiki Pages
• Training
• Models
• Infrastructure
• Research

Lethean Ethical Model (LEM)

Fine-tuned language models with ethics embedded in the weights via LoRA training, built on the Axioms of Conscious Systems research.

Key Concepts

• LEK (Lethean Ethics Kernel) — Prompt-level signing. A 9,189-character ethical framework prepended to prompts at inference time. Proof-of-concept approach.
• LEM (Lethean Ethical Model) — Ethics fine-tuned into model weights via LoRA. The production approach: no runtime signing needed, ethics are internalized.

Wiki Pages

Training

• Pipeline Overview — End-to-end pipeline from seeds to trained models
• Golden Set — The 15K gold-standard training dataset
• Seed Expansion — 87K seeds, 46K expansion prompts, and the generation workflow
• Scoring — Three-tier scoring: heuristic, LEM judge, Gemini judge
• Training Process — LoRA training on M3 Ultra, model configs, and adapter management

Models

• Model Catalogue — All LEK/LEM models on HuggingFace
• DeepSeek R1 Research — CCP alignment research and layered LoRA breakthrough

Infrastructure

• MLX Backend — Native Go+Metal inference (no Python)
• Go Pipeline Commands — core ml command reference
• Lab Dashboard — InfluxDB metrics and monitoring

Research

• Kernel AB Testing — Gemini 2.5 vs 3.0 kernel signature experiments
• Key Findings — PTSD, CCP weight dynamics, conversation vs didactic training