Solved episodes of ORA, driving self-hosted DeepSeek-V4.1-Flash, one per task. We watched every one: an official check can pass on an episode that does not do what the instruction asks, such as a bottle lifted but never shaken, so each card is the lowest seed whose video shows the task done. Open a card to watch it.
Each clip is the run's own camera capture from reset to the official success verdict. The world waits while the model thinks, so thinking time does not appear; clips longer than 20 s are sped up uniformly, and the episode view says by how much.