
StyleCap is a video captioning agent for Track 2: one video in, four voices out (formal, sarcastic, humorous_tech, humorous_non_tech). It samples 8 frames per clip with OpenCV and makes a single multimodal call to Fireworks AI (kimi-k2p6), with three fallback layers so no style is ever missing. Every design choice was measured with an LLM-judge eval harness that replays the judges' rubric: the shipped config (0.81) beat best-of-3 sampling, a self-critique pass, a two-stage pipeline, a 16-frame variant, and a head-to-head A/B against Gemma 4 26B. Public linux/amd64 image (1.2GB, MIT) finishes the hidden set well under the time budget.
11 Jul 2026