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/ · Site: https://activeinferenceinstitute.github.io/ActiveInferenceJournal/ · Tooling: https://github.com/ActiveInferenceInstitute/Journal_Utilities
The coverage target is that every non-duplicate video in the supplied channel
manifest belongs to exactly one canonical item. Validation uses the pinned
Journal-Utilities manifest; it does not enumerate the live channel or establish
current hosted CI status. Billing-lock failures were recorded on 2026-11-12
(see TO-DO.md); those dated receipts do not establish today's hosted
availability or results. The repository is generated: the per-item metadata.json core
keys, transcript.*, captions/, translations/, INDEX.json/INDEX.md and the
item/series README.md files are produced by
Journal-Utilities and
hand edits to them are overwritten on the next run. There is no build step: the
gate is ./run.sh (the arms CI runs), and the static-site build and the
channel ingest run from a Journal-Utilities checkout (see
Operations).
- Works offline: cloning, browsing, grepping and editing the hand-editable files
(see
CONTRIBUTING.md). - Uses pinned Journal-Utilities: index/transcript validation executes a fresh exact-commit snapshot. A sibling working tree is not required as executable source; the static-site workflow uses its separately pinned producer checkout.
- Needs network: the weekly ingest workflow (YouTube enumeration via
yt-dlp). - Not on
main: audio — check out theaudiobranch for the 64 kbps media.
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. Coverage is assessed against the
supplied manifest, not guaranteed against the live channel. Indexed videos without
a more specific series 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 audioto get the media.
Clone it (add --depth 1 for a lightweight copy):
git clone https://github.com/ActiveInferenceInstitute/ActiveInferenceJournal.git
git clone --branch audio https://github.com/ActiveInferenceInstitute/ActiveInferenceJournal.git # with mediaBrowse or consume the content offline: start from INDEX.json (machine)
or INDEX.md (human), then read each item's metadata.json and
transcript.txt. The static reader site is published at
https://activeinferenceinstitute.github.io/ActiveInferenceJournal/.
Browsing requires nothing to install or build. Validation runs from this checkout;
cold Journal-Utilities acquisition and GNN provisioning can require network and
an appropriate runtime. Run Journal-Utilities directly when changing the generator
(enrichment is dry-run unless --apply is passed):
./run.sh all # lint, docs, identity, full, gnn; stop at first failurecd ../Journal-Utilities
uv run python scripts/enrich_metadata.py --journal ../ActiveInferenceJournal --apply
uv run python scripts/generate_journal_indexes.py --journal ../ActiveInferenceJournal
uv run python run.py journal-check # read-only integrity gateSee Operations for every arm, workflow and pin.
- Technical docs index (schema, organization, pipeline, operations):
docs/. - Flat index of every long-lived doc:
docs/documentation-index.md. - Gate arms and measured timings:
docs/OPERATIONS.mdanddocs/_generated/gate-timings.md. - Contribution guide (what to edit vs. generated):
CONTRIBUTING.md. - Security rules (credentials, audio):
SECURITY.md. - Scoped improvement list and open items:
TO-DO.md. - Citation (Zenodo DOI):
CITATION.cff. - Agent conventions:
AGENTS.md.
Transcripts and metadata are produced from Institute YouTube content 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").
This repository carries a committed .aii/ sidecar with curated
InstituteOS metadata (aii-sidecar/v1): category, maturity, ecosystem relations,
key artifacts, provenance, and portable tasks. It is read by
InstituteOS tooling and
does not affect the project's build or runtime.