Three layers, one direction
10
Videos in bronze
raw, immutable
10
Silver rows
typed and validated
184
Total chunks
avg 18.4 per recording
11
Topics covered
levy · defence · NHS · HS2 · …
- Bronze Raw — immutable One row per recording with every raw field: title, channel, upload date, duration, view count, transcript source, transcript SHA-256 and path. Schema-light and never overwritten. This is the audit anchor.
- Silver Cleaned — typed, normalised, validated Typed columns, ISO dates, whitespace-normalised transcript text, and a boolean state registry for index, script and audio. Schema-validated, so a bad row is rejected rather than carried silently downstream.
- Gold Curated — analytics-ready Episode metrics: duration, transcript word count, chunk count, script word count, audio duration, detected topics. The layer analysts and data scientists actually query.
Why a medallion, and not just a folder of artefacts
A folder of artefacts is fine until a second person needs them. Then every consumer re-implements the same parsing, makes slightly different decisions about a malformed date, and produces numbers that disagree. The layers exist so those decisions are made once, in the open, in a place with a schema.
The critical property is that bronze never changes. Silver can be rebuilt when the cleaning rules improve; gold can be recomputed when the metrics change. Neither can quietly rewrite the record of what arrived — which is the same argument the SHA-256 hash makes at the row level, applied to the whole store.
Fig. 06 — two of these three can be rebuilt
Silver and gold are opinions about the data. Bronze is what actually arrived. Keep them in different places.
Based on House of Lords proceedings (UK Parliament). Not affiliated with, or endorsed by, the UK Parliament.