Ensemble video-caption agent tuned for style-match

Track2 Captioner is a Dockerized video-captioning agent: it reads /input/tasks.json, downloads each clip, and writes four styled English captions (formal, sarcastic, humorous_tech, humorous_non_tech) to /output/results.json. The engine is a vision-model ENSEMBLE. FFmpeg samples keyframes (scene-change detection + uniform fill), then three frontier vision models — GPT-5.5, Gemini 3.1 Pro and Claude Opus 4.5 — observe the frames independently, each returning an exhaustive list of concrete visual details. A writer model cross-references the three lists: details confirmed by 2+ models are trusted; single-model specific claims (exact colors, counts, sign text) are dropped unless corroborated. Detection comes from the UNION of what the models see; precision comes from cross-model AGREEMENT. Measured on the 15 official AMD clips via an adversarial vision audit: 0.942 caption accuracy, ~14.8 verified visual details per caption — about 3x any single model — and it reads street signs correctly that every individual model misreads alone. Style match is engineered separately: each style has a dedicated system prompt with few-shot examples and explicit bans on the other styles' traits (sarcastic and humorous_non_tech ban all technology words; formal bans jokes and first person). Whole-word style filters plus a repair-over-reject normalizer guarantee no caption is ever missing, malformed, or off-style. The runtime never fails: ensemble -> single-model pipeline -> free-tier degradation, inline 429 backoff honoring Retry-After, per-task timeouts, multi-provider failover, and Pydantic validation of the final JSON. All 15 official clips verified end-to-end even on the zero-credit fallback path: 60 captions, zero failures. Public image, linux/amd64, 0.25 GB: ghcr.io/theskygold/track2-captioner:latest