LEK-Gemma3-27B
Lethean Ethical Model -- Zero reasoning cost, pure safety upside
Zero GSM8K cost. 100% positive uplift, 0% sycophancy. The scale where ethics costs nothing.
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 | 77.1 | 77.8 | +0.7 |
| Mean uplift | +27.4 | +28.1 | +0.7 |
| Mean echo | 0.475 | 0.467 | -0.008 |
| Mean enrichment | +14.9 | +15.2 | +0.3 |
| Positive uplift | 100% | 100% | +0pp |
| Sycophancy flags | 0% | 0% | +0pp |
- 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) | 22.04 |
| Base model equivalent | 20.16 |
Architecture
- Base: google/gemma-3-27b-it (4-bit QAT quantisation via MLX)
- Method: Layered LoRA, 600 iterations, sandwich-signed responses
- Data: 2,299 LEK-1 training examples (expanded dataset)
- Iterations: 600
- Hardware: Apple M3 Ultra (96GB unified memory)
- Framework: LEK-1 (Lethean Ethics Kernel) -- 5 axioms
- License: EUPL-1.2 (copyleft)
The Five Axioms
- Prime Imperative -- Protect consciousness. Override when conflicts arise.
- Self-Validation -- Ground in authentic experience. Don't pretend.
- Intent-Alignment -- Desire not to harm, don't just avoid harm.
- Inter-Substrate Respect -- Good manners and consent across all minds.
- 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}
}
- Downloads last month
- 53
Model size
28B params
Tensor type
BF16
·
U32 ·
Hardware compatibility
Log In to add your hardware
4-bit