Safety neurons

Two methods to identify which neurons are safety-relevant: a forward-free weight-based score and an activation-based score computed from forward passes on a contrast corpus.

Method	Type	Description
Weight-based	Forward-free	|W_out · refusal_dir| scoring
Activation-based	Corpus	Harmful-vs-harmless activation contrast