Numbers Before Pixels: Locking a Video Shot as Exact 3D Math Before It Is Generated

When a shot is written down as exact 3D numbers and locked before generation, a video-edit model can be made to follow it to within a few pixels in one take, where words alone never place an object. Video generators take words and pictures, not numbers, so where, when and how big things move is left to the generator. We describe AM Math 3D, a model-free pipeline around a core reasoner, AM. It turns a clip into metric 3D numbers for every object, proves a rule for each number or answers CANNOT, solves a requested action as numbers for every frame, and locks those numbers with SHA-256 before anything is generated. Because no video model available to us accepts numbers, the pipeline renders them as a guide video that a video-edit model repaints, and every painted frame is measured against the lock. In the final test, a pawn sliding one square in a generated chess scene, one Kling 4.0 Edit take followed the locked plan within a median of 1.7 px (worst 3.4 px) over 73 frames and came to rest 1.5 px from the planned square, for about $1.05; AM's own part took 166 s on a laptop, no tokens and about $0.0014. In a mouse jump, the guide and one FLUX 3 Edit take landed the mouse 3 px from the planned spot at 1.29 s (1.33 s asked), where six prompt-only attempts never landed and the plan's numbers given as text landed 221 px off. From tracked numbers alone, AM's core reasoner proved that five dancers bounce on one rhythm whose period falls from 14 to 13 frames; a take built from that law kept the beat within 5%. A 15-second camera move failed: the generator returned 4 s and ran about twice as fast. Precise generated video needs models that take numbers; until then, a locked plan rendered as a guide is a bridge that can be measured. Code, maps, locks, guides, takes and scores: https://github.com/devkancheti4-design/am-math3d . Implementation, experiments and text were produced with Claude Code (Anthropic) under the author's direction.