Alibaba Cloud's Project Hub described Dailies on 3 November, an automated quality-control system for AI-generated video built in its Global AI Hackathon series. A Qwen agent turns a plain-language brief into validated shot parameters, the server expands them into a pipeline, and the user reviews the shot plan and stills before any video is generated. Wan then produces the clips. Deterministic checks run on OpenCV against the properties the team could specify: duration, brightness, flicker, scene cuts, camera motion and palette. Qwen-VL adds advisory checks on visual meaning and continuity, the parts pixel arithmetic cannot judge. When a blocking check fails, the system does not only flag it: Qwen revises the prompt and Wan attempts a bounded retake, which must itself pass verification before it is allowed to replace the original shot. The post notes that the deterministic checks use no model tokens and work on output from any generator, and that identical requests are cached so a shot is not paid for twice.
Our takeThe architectural choice is the split: properties that can be written down are checked cheaply and travel across generators, while semantic ones stay advisory instead of blocking. That keeps cost per take flat and makes failures reproducible. So what: the gate that matters is the retake rule — a repaired shot has to clear the same check before it replaces the original, otherwise automated repair quietly lowers the floor while reporting success.