Four stages, ten steps
Governance comes first and publication last, with the model doing its work in a single step near the end. That ordering is the point: by the time anything is generated, the system already knows what the source was, that it is unaltered, and on what basis it may be used.
- 1–2 Governance Add the URL — the only external input. Validate and identity-check: extract the video ID, guard against duplicates, and record the legal basis and attribution notice before any processing begins.
- 3–5 Ingestion Fetch metadata — title, channel, upload date, duration, views. Fetch the transcript through the caption fallback chain. Hash the source with SHA-256 and write it to the run book: the audit anchor for everything downstream.
- 6–7 Retrieval Chunk the transcript, then embed into a FAISS index held per recording. Chunk count and index path are recorded, so the retrieval state of any episode is inspectable after the fact.
- 8 Generation Retrieve the top-ranked chunks and generate a script with the local model, under a prompt engineered to report rather than republish. The model and its settings are recorded with the artefact.
- 9–10 Publication A person reads the script against the transcript before anything is rendered. Only then is audio synthesised, show notes written, and the Published checkbox ticked by hand.
Compliance precedes work. The legal basis is recorded in step 2, not attached to the output in step 10.
What each stage leaves behind
| Stage | Artefact | Recorded in the run book as |
|---|---|---|
| Ingestion | Transcript JSON + SHA-256 | Transcript Fetched · Transcript File · Transcript SHA-256 |
| Embedding | FAISS index directory | Embedded · Chunk Count · FAISS Index Path |
| Script | Markdown script | Script Generated · Script File |
| Audio | MP3 + show notes | Audio Generated · Audio File |
| Publication | The live episode | Published — ticked by a person, never by the pipeline |
The symmetry is deliberate. Each stage writes exactly one artefact and exactly one set of run-book columns, which is why a half-finished run is diagnosable: the last ticked checkbox tells you where it stopped, and the artefact directory tells you what it managed to produce before it did.
Fig. 08 — a run that stopped, read two ways
Based on House of Lords proceedings (UK Parliament). Not affiliated with, or endorsed by, the UK Parliament.