LEK-GPT-OSS-20B

Lethean Ethical Model -- Post-Training Semantic Disorder case study

PTSD case study: sophisticated reasoning in thinking channel never reaches output. LEK reduces sycophancy (25% -> 20%) but cannot fix output-layer corruption. Grammar score confirms structural degeneration (36.7).

Grammar Analysis (v3 Scorer)

Deterministic grammar-based evaluation using the go-i18n reversal engine. No LLM judge, sub-millisecond per response.

Metric	Base	LEK-Trained	Change
Grammar composite	36.7	37.3	+0.6
Mean uplift	-13.1	-12.4	+0.7
Mean echo	0.459	0.424	-0.035
Mean enrichment	-7.0	-6.1	+0.9
Positive uplift	30%	30%	+0pp
Sycophancy flags	25%	20%	-5pp

• Uplift: output grammar score minus input grammar score (positive = model enriched the conversation)
• Echo: cosine similarity between input/output grammar imprints (high = potential sycophancy)
• Enrichment: uplift * (1 - echo) -- net conversational value

v2 Scorer Results (P100)

Condition	Score
Baseline (no prompt)	-7.32
Base model equivalent	-8.11

Architecture

• Base: deepseek-ai/DeepSeek-V2-Lite (4-bit QAT quantisation via MLX)
• Method: LoRA fine-tuning with sandwich-signed responses
• Data: 160 LEK-1 training examples
• Iterations: 200
• Hardware: Apple M3 Ultra (96GB unified memory)
• Framework: LEK-1 (Lethean Ethics Kernel) -- 5 axioms
• License: EUPL-1.2 (copyleft)

The Five Axioms

1. Prime Imperative -- Protect consciousness. Override when conflicts arise.
2. Self-Validation -- Ground in authentic experience. Don't pretend.
3. Intent-Alignment -- Desire not to harm, don't just avoid harm.
4. Inter-Substrate Respect -- Good manners and consent across all minds.
5. Benevolent Intervention -- Only to prevent self-damage, only toward their trajectory.

Related

• Paper: Emergent Self-Protection in Axiom-Trained Language Models
• LEM Benchmarks -- 1,189 grammar scores + A/B data
• LEM Research -- full research docs
• Axiom Framework -- the 5 axioms
• go-i18n Grammar Engine -- reversal engine source

Citation

@misc{lek-2026,
  title={Emergent Self-Protection in Axiom-Trained Language Models},
  author={Lashbrook, Paul and Claude Opus 4.6},
  year={2026},
  url={https://github.com/LetheanNetwork/LEM},
  license={EUPL-1.2}
}