Production path
From a bundle of sources to a delivered episode
Seven stages, each with a declared input, output, and owner. Find the stage that concerns you; its maturity is stated on the stage, not inherited from the pipeline.
The seven stages
The order is the content, so it is an ordered list. No diagram is required to follow it.
-
Source intake
AvailableBriefs, transcripts, feeds, press releases, and research notes arrive as uploads or remote URIs and are recorded against an ingestion job.
- In
- Heterogeneous source documents
- Out
- Ingestion job with attached, classified sources
- Owned by
- Integration client
-
Canonical content
AvailableSources are weighted, conflicts resolved by documented rule, and the result normalized into canonical TEI with provenance captured in the header.
- In
- Classified source documents
- Out
- Canonical TEI episode with provenance
- Owned by
- Ingestion service
-
Draft generation
In progressStructured briefs and guardrails compose the prompt; a LangGraph state graph drives generation with metered, interruptible model calls.
- In
- Structured brief, pinned reference revisions, guardrails
- Out
- Draft TEI script, show notes, chapter markers
- Owned by
- Content generation orchestrator
-
Automated evaluation
In progressThe in-progress evaluator stack is designed to run five model evaluators and the Chrono local estimator in parallel over the draft, returning typed findings. Malformed structured output and invalid timing stop the run rather than publishing quietly.
- In
- Draft TEI script and cited source packets
- Out
- Typed findings, runtime estimate, usage metrics
- Owned by
- Quality assurance stack
-
Editorial checkpoint
PlannedFindings are surfaced to an editor, who approves, requests changes, or redirects the graph. The workflow suspends without losing state while that decision is pending.
- In
- Draft with aggregated findings
- Out
- Approval decision and audit record
- Owned by
- Editor
-
Speech and mixing
PlannedApproved segments render to speech with pinned pronunciation, then mix with beds and transitions and normalize to a broadcast loudness target.
- In
- Approved script segments
- Out
- Mastered episode audio
- Owned by
- Audio pipeline
-
Preview and export
PlannedEditors preview the rendered episode, embed chapter markers and metadata, and export delivery formats.
- In
- Mastered audio and canonical metadata
- Out
- Published episode package
- Owned by
- Editor and export service
Reading the maturity labels
Two stages are complete. That does not make the pipeline complete, and the labels are per stage for exactly that reason.
Source intake and canonical content are available: an integration client can push material through them today and get a canonical TEI document with its provenance attached. Draft generation and automated evaluation are partly built. Everything after the editorial checkpoint is designed and not implemented.
A finished early stage is the usual way a pipeline misleads. It is tempting to read "ingestion works" as "the pipeline works", so each stage above carries its own state and its own evidence link, and no summary on this site rounds them up into a single verdict.
What stops a run
Fail-fast is a design commitment. These are the conditions that halt work rather than publishing quietly.
- Malformed structured output from a model Available
- The run stops. Nothing partially parsed reaches the canonical episode.
- Chapter-marker timing that does not match segment transitions Available
- Validation rejects the payload. Starts must be ordered, integer-only, and in PT#H#M#S form.
- Optimistic-locking conflict on a profile or template Available
- The write is refused so a concurrent edit is never silently overwritten.
- Repeated request with a matching idempotency key Available
- The stored response replays. The same key with a different canonical body returns 409 Conflict.
- Per-episode budget exhausted In progress
- The reservation is refused before the billable call is made.
Stages in detail
Three pages cover the stages that carry the most weight.