A Patch in a Static Image Can Explain Part of the Future

A static image records the present, yet the representation of one patch may already contain geometry aligned with how that physical point will evolve. We study this possibility through PMTT, which represents a persistent patch by a Canonical State, identity-preserving observation-conditioned states, and their Patch-Level Local Feature Space. The training and evaluation chain is central to the result. DINOv3 is pretrained through image-based self-supervision rather than a temporal objective for predicting subsequent physical states, and we keep the backbone frozen. PMTT then constructs the patch-level state family and its continuous low-dimensional Manifold analytically from one current image. Future states enter only after this geometry is fixed, as independent measurements of physical-time change. On a new untouched synthetic cohort, the first four Lawful32 directions explain 35.50%, 32.54%, and 29.02% of future tangent variation for DINOv3-S, DINOv3-B, and DINOv3-L, compared with 0.39% to 1.04% for random subspaces. On natural video, the same geometry explains 23.50% to 27.58% on a common cross-scale subset and 23.03% at one-frame and 21.82% at four-frame horizons over all eligible DINOv3-B transitions. The effect is multi-directional, strongly low-dimensional, shared across distinct lawful observation families, and progressively recovered as observation coverage increases. Organizing the same acquired observation states as a continuous bounded Manifold raises future-state coverage by 29.38 to 31.55 percentage points on the untouched synthetic cohort and by 12.69 to 19.32 points on full natural video. These results establish a new representation foundation and opportunity for deep-learning physical modeling. The same static-to-future finding also exposes a systematic cross-axis phenomenon: geometry formed along the static Observation axis aligns with subsequent change along the Physical-time axis although no temporal future-prediction objective created that geometry. On New16, physically matched configurations exceed DINO-feature-closest incorrect configurations by 14.90 to 15.10 percentage points across DINOv3-S/B/L, while aggregate constraint-conditioned advantages reach 18.29 to 20.91 points. The alignment is organized by configuration, constraint axis, joint location, and motion family, survives visual-closest and object-shared alternatives, and follows persistent physical identity in natural video. These results give the phenomenon a reproducible physical structure and open a New Physics Question Space: why does an image-pretrained, frozen static representation contain local geometry that systematically aligns with future physical evolution, and why is that alignment organized by the current physical state and constraints? Code and experimental assets are available at: https://github.com/zixiongwang-ai/PMTT-physics.