Direct answer

Short answer: We anchor each story to a Discogs master release, run a fixed ten-question research framework, clean and bucket facts, then a human editor produces the listenable script. Audio is text-to-speech from that approved text. We do not ship unchecked “raw” model output as a final episode.

How AI fits — and what humans own

Listeners often ask if shows are “just AI.” In our workflow, AI behaves like a fast research assistant: it helps gather and organize candidate facts. Humans own authorship and accountability for what gets said: editors cut hype, fix tone, drop weak claims, and sign off before audio. AI does not decide which flagship episodes publish or replace editorial judgment on sensitive claims.

We are not implying endorsement from anyone quoted here—but many artists describe technology the same neutral way we think about our stack: outcomes depend on human choices.

Full timeline (source to listener)

1. Source material — Physical or digital listening context, liner notes when available, and canonical album identity from Discogs (credits, track order, release lineage).
2. Structured research — Ten fixed “lenses” (see below) guide what we try to cover so the arc stays coherent: origin, craft, reception, legacy.
3. Fact hygiene — Claims are normalized, deduplicated, filtered for relevance, and contradictions are flagged for editors instead of being silently merged away.
4. Balance checks — Coverage metrics highlight gaps (e.g. too little on reception or visuals) before script finalization.
5. Scripting — Editors write or heavily rewrite listening copy from the curated package; the goal is curious, precise, non-hype storytelling.
6. Audio — Approved text is rendered with text-to-speech for clarity and scale.
7. Publication — Episode metadata ships with the audio; public SEO remains limited to editorial URLs (not user-generated copies in the open sitemap).

Step 1 — Anchor: the Discogs master release

Every episode locks to one Discogs master release as ground truth for spelling, sequencing, and which pressing family we mean.

• Official credits & order — Track order and credited roles come from that master context.
• No freestyle identity drift — The pipeline is constrained to narrate the story around that anchored release, not a generic or mistaken edition.
• Compilations & box sets — When the master is a compilation, we shift emphasis to remastering, unreleased material, or release strategy instead of implying a single fictional studio month.

Step 2 — Ten research lenses (before writing)

These lenses are not ten separate podcasts; they shape one research pack that feeds a single script.

Three threads from production culture echo lenses 2, 5, and 6 above. Independent voices—not VinylCast endorsements:

Step 3 — Fact curation & narrative bucketing

1. Normalization — Raw snippets become atomic statements (who / what / when / where).
2. Deduplication — Repeated claims collapse so narration stays tight.
3. Conflict handling — Disagreements stay visible to editors; we do not silently pick a winner without review.
4. Relevance filtering — Generic fluff that is not about this album drops out.
5. Bucketing — Facts group into narrative sections (context, craft, reception, legacy) to preserve balance.

Step 4 — Human editorial (final say)

Editors transform the curated pack into spoken narrative: pacing, transitions, and judgment calls on what to omit. They remove machine-ish phrasing, refuse unsupported assertions, and align tone with VinylCast’s standard: curious, exacting, anti-hype. Publication-ready text is always human-approved.

That discipline is the editorial cousin of Rubin’s instinct in lens 5: say it with only what the moment needs—here, in listenable sentences rather than in a mix.

Step 5 — Audio & synthetic voice

Today, episodes use text-to-speech for narration. Words follow the human-approved script; timbre is synthetic and tuned for clarity on facts. We disclose synthetic speech so audiences (and search systems) can assess provenance. For who shapes brand voice and review duties, see the voice behind the episodes.

Limits & reliability

• Source freshness — Research reflects what indexed sources show at production time.
• Residual error risk — Automated retrieval can misread nuanced criticism; that is why human rereading is mandatory, not optional polish.
• Subjectivity — Reviews are opinions; we aim to represent period voices without declaring one take as absolute truth.

Industry coalitions publishing AI principles for music and creators stress the same trust surface we try to keep visible on this page:

FAQ — method & limits

Is VinylCast “just AI”?

No. AI accelerates research and structuring; humans edit and approve the script listeners hear.

What is the Discogs master anchor?

It is the canonical album identity we attach each episode to—credits, sequencing, and release family—so the story cannot drift to the wrong record.

Do you publish raw machine drafts?

No final episode ships as an unreviewed model dump. Editors control wording and claims.

Why text-to-speech?

It scales consistent narration across voices and languages while keeping production predictable; we label it clearly as synthetic.

How do compilations work?

We avoid pretending there was one classic studio month; we emphasize remasters, unreleased tracks, or packaging strategy instead.

How can I report an error?

Use the in-product feedback path you use for your account so we can trace the episode version and correct the chain.

Related reading

If you want to know who stands behind the voice you hear, how we are clear about synthetic speech, and who takes responsibility for what is said, read the voice behind the episodes. Building a tool or assistant? llms.txt is a short, plain summary of the same ideas in one place.