Cleaned-up captions (v2)

Approved cleaned-up version: sparse tag/UI details, font attribution, motion notes and labeled translations are consolidated into existing descriptive fields. Original model annotations and earlier exports are preserved. Transcripts and timing values are unchanged; this is data cleanup, not human acceptance.

This export contains 1,609 captions covering 1,610 collected source records, including one byte-identical alias. Captioning and upload are distinct from human acceptance.

Segment, classification, tag and utterance confidence fields, plus utterance uncertain_text, are omitted from the current caption set. Text, timing and free-text uncertainty notes are preserved.

Method and limits

Gemini 3.8 Flash analyzed the source videos and available audio at requested 2 fps and high spatial resolution. Captions describe content beats, layout, motion, UI behavior, identity requirements and speech. All records passed structural, reference and timing validation. Content accuracy and exact boundaries remain unreviewed model annotations.

Category durations describe coarse segment presence. Primary formats partition the unique footage; content tags overlap. Do not sum overlapping tag hours into unique footage.

Source media stays on SeedHost; this text-only review does not stream or republish the videos. This public subset excludes credentials, internal source paths, provider request paths and operational manifests. Existing team files and review deployments are preserved.

Download complete caption JSONL (gzip) · Source index · Coverage summary

Published using the static Cloudflare workflow documented in PR 2247. This standalone export includes only caption-review navigation, not the full project tracker.