Why do video generators forget physics after two steps?

Video generation models that start from images often violate physics as denoising progresses, even though early steps produce more realistic motion. The culprit: the model's internal phase information degrades by ~18% during the 50-step denoising process while visual details sharpen. PhaseLock locks in the correct motion from just 2 steps and enforces it throughout generation via a training-free guidance method, improving physical consistency by 6.2 points across models with minimal overhead and without expensive external constraints.