Custom llama.cpp fork with character intelligence engine: control vectors, attention bias, head rescaling, attention temperature, fast weight memory

Research notes & proposals: keeping minority-but-valuable views alive in LLM committees, and decontaminating an AI judge of its hedging bias (Proposition H-perp). NOT validated results — please falsify.

Build fast weight attention for episodic memory in Transformers with causal and chunked sequence support