SEV-Cap: Grounded Multi-Style Captions

Single-pass video captioners fail the LLM judge on two axes: hallucinated details (accuracy) and style labels that sound right but read wrong (tone). SEV-Cap is built around those axes. The pipeline samples ffmpeg keyframes, writes one shared scene description, then self-verifies it against the frames before any styled writing. For each required style it drafts multiple candidates (text writers plus vision-grounded drafts), scores them with a vision prejudge on accuracy and tone—the same metrics Track 2 uses—and keeps the best. Weak styles get a polish/reselect pass while time remains. Jokes are locked to concrete nouns from the verified description so humor does not invent cities, brands, camera moves, or plot. The container matches the Track 2 contract: linux/amd64 image ghcr.io/skx56/sevcap-grounded:latest, argument-free entrypoint, /input/tasks.json in and /output/results.json out. An anytime algorithm writes placeholders early and upgrades captions with atomic writes under a global budget and per-clip timeout, so timeouts degrade quality but do not produce MISSING_TASKS or OUTPUT_MISSING. Default model is Kimi K2.6 on Fireworks (VLM + text). Gemma remains available via env override for the bonus track. On the eight official AMD sample-style clips, an internal judge mirroring Track 2 axes scored combined 0.966 ((mean accuracy + mean tone) / 10).