Spectral Optimization for Controlling Safety Instruction Strength

The Safety Operator method treats context tokens as multiplicative operators within transformer weights, where the dominant eigenvalue determines instruction strength. A new Contrastive Safety Loss allows users to tune the tradeoff between attack success and over-refusal rates by modulating this eigenvalue.