FourCast — The Caption Factory

Most caption pipelines are a prompt wrapper. FourCast is a caption factory with quality control. Frames are sampled straight off the video URL (ffmpeg remote seek — no full download), and Kimi K2.6 writes a strict Scene Report: subjects, actions, on-screen text, distinctive details. Facts only — if it isn't visiFireworksble, it doesn't exist. GLM 5.2 then writes three candidates per style against explicit style contracts; humorous_non_tech must additionally survive a regex jargon firewall — one tech word and the candidate dies. A separate model family (gpt-oss-120b) judges every candidate on the exact two axes the official leaderboard uses — caption accuracy and style match — picks the winner, and sends weak ones back for a critique-guided rewrite. The container is engineered for the evaluation harness: results are atomically rewritten after every clip (a timeout still leaves valid partial output), every style always gets a grounded caption (never a missing key), output JSON is ASCII-escaped, and one bad clip can't crash the batch. Three example clips process in ~55 seconds; a 12-clip hidden set fits comfortably inside the 10-minute budget. On the sample clips, our internal judge scores land at 0.85–1.00 on both axes — e.g. "Kitten boots from a stationary state, then begins a slow forward deploy toward the camera, tail raised high as a status indicator." Built on Fireworks AI, with serverless models hosted on AMD hardware.