Maturity Active development 33 of 120 roadmap tasks complete (1 of 6 phases). Not a released product.
Canonical content
Sources arrive in whatever shape they were written. They leave as one canonical document that still knows where every part of it came from.
The canonical document is a content and provenance model before it is a file format.
Ingestion produces a single document that holds the episode's content and, alongside it, the record of where that content came from. Both live in the same structure, so a claim and the source that produced it cannot drift apart in the way they do when provenance is kept in a separate log.
That structure is TEI - the Text Encoding Initiative's P5 vocabulary, a long-standing standard for marking up documents with scholarly apparatus. Episodic uses its header for provenance and its body for the script. Serialization is handled by tei-rapporteur, a TEI P5 library maintained alongside the platform.
Where a prompt needs JSON, that JSON is a projection of the same TEI-backed content model. It is not a second canonical schema competing with the first.
Five source classes today, each classified and scored on arrival.
| Source type | Typically carries |
|---|---|
| Transcripts | Recorded interviews and prior episodes. |
| Briefs | Editorial direction for the episode or series. |
| RSS feeds | Syndicated items, pulled as an ongoing source. |
| Press releases | Announcements, treated as interested sources. |
| Research notes | Working material assembled by the production team. |
A source reaches the system as an upload or a remote URI attached to an ingestion job. Document classifiers assign a type and a quality score; neither is a judgement about whether the content is correct.
When two sources disagree, one of them wins. The rule that decides is written down.
Weighting heuristics compute a priority score from quality, freshness, and reliability. Where sources conflict, the weighting matrix resolves which normalizes into the canonical body. Coefficients are configurable per series profile, so a team that trusts its own transcripts over a press release can say so once rather than per episode.
Every source is retained regardless of the outcome. Rejected material stays queryable against the ingestion job, because the decision is only auditable if the discarded side of it still exists.
Recorded automatically in the canonical header. The example below is synthetic.
<sourceDesc>
<listBibl>
<bibl xml:id="src-transcript-01" n="1">
<title>Interview transcript, second pass</title>
<note type="priority">0.82</note>
<note type="classifier">transcript</note>
<date type="ingested">2026-05-14T09:31:07Z</date>
<respStmt><resp>reviewed by</resp><name>editor-14</name></respStmt>
<note type="outcome">preferred</note>
</bibl>
<bibl xml:id="src-release-04" n="2">
<title>Partner press release</title>
<note type="priority">0.31</note>
<note type="classifier">press-release</note>
<date type="ingested">2026-05-14T09:31:09Z</date>
<note type="outcome">retained, not preferred</note>
</bibl>
</listBibl>
</sourceDesc>
Identifiers generated during canonical ingestion are time-ordered UUIDv7 values, so records sort chronologically without a separate timestamp index. Large TEI payloads are compressed with Zstandard in storage; the API and domain read paths return plain text and never expose that.
Material that outlives one episode: style guides, research briefs, and the people on the show.
A reference document is versioned content bound to a series, template, or episode by a pinned revision. Because the binding names a revision rather than a document, an episode generated last month keeps the style guide that was current when it ran.
Hosts and guests are series-aligned reference documents rather than
properties of an episode. A host can be absent from an episode and a
guest can recur across several; binding rules control applicability, and
a revision can carry an effective_from_episode_id so a
change applies from a chosen episode onwards.
Transcripts, briefs, feeds, press releases, and research notes normalize into canonical TEI. The header records source priorities, ingestion timestamps, and reviewer identities.
Roadmap step 1.2, multi-source ingestion (opens the Episodic repository)Priority scores are computed from quality, freshness, and reliability, with coefficients configurable per series profile. Every source is retained for audit, preferred or rejected.
Users' guide, content creation (opens the Episodic repository)Style guides, research briefs, and host and guest profiles are versioned documents bound to a series, template, or episode by pinned revision.
Reference binding resolution (opens the Episodic repository)See the reference-document endpoints · Follow the draft into review