Auditing Hidden Future-Token Leakage in Adaptive Sequence Models

Investigation into an adaptive residual-routing decoder reveals that sequence-level routing summaries can bypass causal masks, causing significant future-token leakage. In one 24-layer instance, leakage artificially dropped validation loss from 3.5958 to 1.4259, though the model failed prefix-invariance audits and showed poor generation perplexity.