Physical-Plausibility Probe

Key Insight

A generated video can look gorgeous frame-by-frame yet quietly break the rules of the physical world — water flowing uphill, a dropped ball hovering in mid-air, an object that vanishes when something passes in front of it. These failures of physical plausibility — including object permanence and world consistency — are exactly what today's automatic metrics miss and what Sora's own report flags as still unsolved. This project builds 50 deliberately tricky prompts that should obey simple physics and uses them to probe open models, turning the fuzzy "does it understand the world?" question into concrete, inspectable pass/fail cases.