Content of the Active Inference Journal — transcripts, metadata, captions, translations, and curated materials from the Active Inference Institute video library.

Learn more: https://activeinference.institute/learning/ · Tooling: https://github.com/ActiveInferenceInstitute/Journal-Utilities

Content is source-namespaced so other channels and non-video sources can live alongside, e.g. data/video/<other-channel>/… or data/<other-type>/<source>/…:

data/video/activeinferenceinstitute/<Series>/<item>/
  metadata.json     # canonical: series, item, parts[{video_id, url, title, duration, upload_date, speakers}]
  transcript.txt    # derived speaker-labeled text (part-tagged when multi-part)
  transcript.json   # raw diarized segments — immutable, SPEAKER_NN labels (where available)
  captions/         # original-language .srt (+ youtube_captions.txt where captions predate diarization)
  translations/     # translated .srt (per language)
  assets/           # images, html, prose, appendices, bibliography, …
  README.md         # human nav: titles, links, contents
docs/               # technical documentation (SCHEMA.md, …)
INDEX.json          # machine entry point: every item, its videos, paths
INDEX.md            # human index, grouped by series

Current item and video totals are generated in INDEX.json. Deliberate duplicate records are marked with duplicate_of; the index reports both indexed video records and unique video IDs. Every Institute channel video is represented (uncategorized videos live under Other/).

• main — everything above, without audio (lightweight to clone).
• audio — main + <item>/audio/<name>.64k.m4a (audio re-encoded to 64 kbps). git checkout audio to get the media.

• Technical docs (schema, organization, pipeline): docs/.
• Contribution guide: CONTRIBUTING.md.
• Citation (Zenodo DOI): CITATION.cff.
• Agent conventions: AGENTS.md.

Transcripts and metadata are pulled completely and idempotently from the Institute YouTube channel by Journal-Utilities. Most transcripts are WhisperX-diarized: transcript.json holds the immutable raw segments (SPEAKER_NN), human speaker identifications are recorded in metadata.json parts[].speakers, and transcript.txt is regenerated from the two (Journal-Utilities/scripts/apply_speaker_names.py). Items without diarization yet carry YouTube-caption text; original captions always remain under captions/. Private/unlisted videos are not transcribed. See docs/SCHEMA.md ("Transcripts — raw vs derived").