VeriPhy Flags 228 AI-Video Physics Flaws vs Baseline's 164

On a 149-clip core carrying 304 human-annotated flaw records, VeriPhy accounts for 228, against 164 for a published question-decomposition evaluator given the same clips and the same claims.

The abstract does not sell that gap as the win. Prompting the same backbone monolithically reaches 222, six behind VeriPhy. "Recall alone does not separate it from prompting the same backbone monolithically," the paper reads; "what separates them is that each decision retains its evidence record and provenance, making the traces auditable one verdict at a time."

VeriPhy compiles a prompt into "typed physical obligations and a statically validated execution plan" before any frame is observed, then routes work to frozen low-level experts: segmentation and tracking, counting, eleven typed physical measurements, depth, OCR, and audio-event detection. Verdicts land in a three-valued state (supported, contradicted, or unknown), surfaced to a user as plausible, implausible, or abstain.

The framing lives in one flat sentence at the top: "Visual fluency in generated video does not imply physical reliability, and a scalar quality score alone is incapable of indicating the obligation a clip violates or the moment it fails."

One caveat is on the tin. The 149-clip core is a development set; "these figures characterize the system rather than its generalization," the authors write.