The method / 02

A listening history.
A research memory.

Podcast Corpus preserves complete, official transcripts and their provenance. The Blind Machine can then return to them, cross-read them, and build interpretations with evidence.

01

Collect evidence before making meaning.

A saved episode is a starting point. An accepted transcript is a research object: complete text, an official source, a retrieval date, and a stable identity. Keeping those layers separate makes the record inspectable.

From an episode to the Archive

Target architecture
01

Discover

Spotify’s Your Episodes and The Blind Machine playlist identify the episodes to follow.

Spotify
02

Find the source

Check official RSS transcript links first. Use an official-domain fallback when needed.

RSS / Firecrawl
03

Verify & preserve

Accept complete official text. Preserve its source, timestamp, hash, and version privately.

Corpus / Supabase
04

Interpret & return

Deliver an accepted object through the Archive’s ingestion boundary. Interpret against evidence.

TBM Archive / downstream

A scheduled TypeScript worker conducts collection. Firecrawl is a fallback. Routine collection uses deterministic rules; AI interpretation belongs downstream.

Try the mechanics

The source gate

Local simulation

These fixed scenarios demonstrate the policy. No external service is called.

Waiting for a check

Evidence has to pass the gate.

Choose a discovery result, then inspect the outcome.

Official relationshipComplete transcriptSource & provenance
02

Missing is a legitimate outcome.

Show notes, summaries, and audio do not become transcripts. A missing official transcript triggers retries on day 1, day 7, and day 30, anchored to the first check. After four unsuccessful checks, the episode becomes unavailable. Ambiguous evidence goes to operator review.

Each layer has a job.

Spotify
Episode discovery and source membership.
Supabase
Target private datastore, scheduled worker, queue, and operational ledger.
GitHub
Canonical code, policy, contracts, tests, and migration record.
Lovable
Planned operator interface connected to the operator-owned backend.
Sidekick & Sheets
Earlier implementation and frozen migration evidence.
TBM Archive
Separate admission boundary and the downstream research instrument.

Tonight’s demonstration

A working interface.
A deliberately bounded sample.

Working here

Search and filter 15 real episode metadata records, inspect historical transcript status and official source links, follow manually assigned editorial lenses, and simulate the source gate.

Project record

The repository documents completed portable domain policy and an implemented Phase 2 backend. Its README, last updated 31 August 2026, says that backend was not yet deployed. Production deployment is not verified by this demo.

Kept separate

No live Spotify sync, transcript scraping, AI generation, or Archive delivery runs here. Full transcripts and credentials are absent. Editorial questions and related items are curated demo annotations, not generated findings.

Grounding: the canonical repository README and domain policy, the frozen Google Sheets library, and Sidekick’s earlier operating guide. Metadata retrieved 30 September 2026; individual status checks date from July–August 2026.

Keep the source.
Leave the interpretation open.

Explore the corpus