Don't Act Before the Action: Auditing and Repairing Future-Action Leakage in Chunk-Autoregressive Video World Models

This concise preprint introduces a suffix-intervention audit for future-action leakage in chunk-autoregressive video world models. The audit fixes the context, action prefix, and sampler noise while changing only future actions. We present Causal Action Residual Distillation (CARD), which combines an action-free content stream with a read-only causal action residual. Under the sampler and decoder assumptions stated in the paper, this structure makes earlier outputs invariant to later actions under fixed noise. Preliminary experiments on the released ABot-World-0 runtime show pathwise leakage in the original model and exact prefix preservation with active action responses in the CARD sidecar. This initial version focuses on the method and a limited empirical case study. Broader evaluation and closed-loop control performance remain future work.