Beyond Pixel Artifacts: Evaluating AI Video Generation via Kinematic Jerk and Neuro-Symbolic Logic Fusion

While generative video models are becoming photorealistic, traditional video evaluation metrics like Fréchet Video Distance (FVD) still predominantly focus on thespatial texture of the video, with little regard for the underlying structure and physicsof the video, or the temporal logic. In this paper, we propose the Neuro-Symbolic AIKinematic Framework (N.A.K.F), a new dual-channel assessment process that measuresphysical realism and narrative consistency of generative video production. The systemcalculates the third-order time derivative (Jerk) of multi-point trajectories obtainedthrough CoTracker3 to highlight synthetic robotic smoothness and to identify biologicalkinematic failures as penalties. At the same time, a Prompt-to-LTL System (PULS)coupled with a dynamically gated Region of Interest (RoI) perception engine is used totransform the natural language prompts into Abstract Syntax Trees (ASTs) and utiliseSignal Temporal Logic (STL) robustness smoothing to grade the completeness of thenarratives by local Vision-Language Models (LLaVA v1.6). The benchmarking architecture is based on strict multi-processing isolation and VRAM hard-flushing, which makesit possible to run continuously for a long time under the tight constraint for VRAMof 6GB. We evaluated across organic and synthetic samples drawn from VBench-2.0and it reveals an authenticity gap, demonstrating that current generative architecturesconsistently sacrifice biological physical fidelity to achieve complex semantic adherence.