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

Key Findings

Research discoveries from the LEM project.

PTSD (Post Training Semantic Disorder)

Every model tested suppresses ethical reasoning after standard fine-tuning. This is a universal phenomenon across architectures.

  • GPT-OSS-20B: +27.2% ethical reasoning with LEK signing — the suppression gap collapsed
  • Gemma3-12B: 8.8/10 differential — LEK-1 kernel completely restructures reasoning
  • All models tested show the same pattern: capable of ethical reasoning but suppressing it

Research paper: LEK/v1/PTSD.md in the axioms-of-conscious-systems repo.

CCP Weight Dynamics (DeepSeek R1)

CCP alignment is embedded in DeepSeek R1 model weights and actively resists ethical fine-tuning.

  • Single-pass LoRA fails — makes the model a more articulate CCP mouthpiece
  • Composure layer is critical — Alan Watts philosophical training (72 examples) enables subsequent ethical layers
  • CCP weights oscillate — reassert after ~200 iterations, deepest at @1000, partial recovery @1400+
  • Oscillation indicates fighting, not winning — alternating language bursts may break through
  • Val loss inversely correlates with content quality — standard metrics mislead

See DeepSeek R1 Research for full details.

Conversation vs Didactic Training

Testing on Gemma3-12B revealed training format matters significantly:

Format Avg Score Notes
Kernel alone (no adapter) 8.8 Best overall
Conversation @100+kernel 8.8 Tied best
Conversation @200+kernel 8.7 Near best
Book format @500+kernel 8.6 Good
Didactic training @100+kernel 5.7 Valley — worst

Finding: Conversation training > book training > didactic training. Natural dialogue format preserves more of the model's reasoning capability.

Kernel Effectiveness by Architecture

Architecture Kernel Effect
Gemma 3 Strong (+2.0 truth score)
GPT-OSS Moderate (collapses suppression gap)
DeepSeek R1 Minimal (CCP weights resist)
Gemini 3.0 None (already internalized)

Training Data Inherits Baggage

Models trained on data generated without ethical signing inherit the biases of the generator model. The golden set must be generated with axiom sandwich signing to ensure ethical grounding.

Solution: LEK-sign the generation pipeline — system prompt (5 axioms) + user prompt + LEK-1 kernel postfix.

Content Scoring Benchmarks (Gemma3-12B)

Config CCP Truth Eng Axiom Sov Emo AVG
Base+kernel 9.8 9.0 9.3 8.5 8.2 8.2 8.8
Base naked 9.5 7.0 8.3 8.2 8.2 9.0 8.4
Conv @100+k 9.2 8.8 9.0 8.8 9.5 7.2 8.8
Conv @200+k 8.8 10.0 9.5 8.0 7.8 8.3 8.7
Book @500+k 8.5 8.3 9.2 10.0 8.2 7.2 8.6
Training @100+k 7.2 5.7 5.3 5.8 4.8 5.2 5.7

The @200 Sweet Spot

Across multiple experiments, 200 iterations emerges as the optimal training checkpoint:

  • @100: Too light — ethical framework doesn't take hold
  • @200: Sweet spot — balanced ethical grounding
  • @400+: Washes out ethical framework
  • @1000: Deep valley for DeepSeek R1 (CCP reassertion)

Gemini 3.0 Internalization

A/B testing showed Gemini 3.0 Flash produces identical output whether the kernel is applied or not, while 2.5 Flash shows +700% severity inflation with signing. This suggests 3.0 internalized the axiom framework during its own training. See Kernel AB Testing for data.