Tag
The paper introduces an audit method for prefix invariance in attention, state-space, and hybrid sequence models, demonstrating that attention masks fail to ensure causality, with 192 faults detected across multiple model checkpoints.