LaVGuard: Selective Decoding for Contextual Value Drift

LaVGuard is a selective decoding method for contextual value drift that separates an authorized ten-dimensional value profile from a frozen estimate of the visible prefix's values, and implements intervention via logit tilting constrained by modified positions and a cumulative KL budget. Profile conditioning raises the lower-bound TPR at 5% FPR from .58 to .72; at the same average intervention proportion of .12, LaVGuard scores .88 on risk-safety, versus .72 for position-matched editing. The degree of intervention is explicit and bounded.