Action-chunked policies generate sequences of actions as a unit — a chunk of future motor commands predicted and executed together. At the boundary between consecutive chunks, the policy generates a new sequence that must seamlessly continue from where the previous one ended. The discontinuity at this boundary has been treated as an unavoidable execution nuisance.
It is not unavoidable. It is a noise-dominated failure mechanism (arXiv:2603.11642). In frozen pretrained policies, chunk-boundary artifacts are strongly associated with task failure (p < 10^-4). The artifact emerges during rollout, not as a post-hoc symptom — the discontinuity causes failure rather than correlating with it.
The mechanism: under a fixed observation context, changing only the latent noise that drives the generative model systematically modulates artifact magnitude. The boundary discontinuity is not determined by the observations but by the noise sample. Different noise samples from the same visual state produce different degrees of boundary artifact. The failure is in the generation process, not the perception.
This makes the artifact intervenable. By identifying artifact-related directions in noise space and steering trajectories away from them, artifact magnitude changes reliably across all evaluated tasks. On hard tasks with remaining performance headroom, the success rate shifts accordingly. The intervention works because the failure mode has a specific geometric signature in the latent space.
The reframing matters: a nuisance is something you tolerate. A failure mechanism is something you fix. The chunk boundary was always a seam in the policy's output — but seams can be inspected, characterized, and reinforced.