Content lifecycle
This is the end-to-end path of every article, as it actually runs in production: a nightly timer drives the whole cycle across the contract fleet; between runs the platform is idle.
Collect — finding what to write about
Section titled “Collect — finding what to write about”Collect harvests candidate source articles for the contract’s niche and ranks them by embedding cosine similarity against the contract’s ontology (the current major line of the Collect pipeline is vector-based). Qualified sources are seeded as pending articles; alongside, source pages are mined into research facts — the contract’s grounding corpus. A topic-deduplication check on the Collect side prevents already-covered topics from being seeded again.
Generate — producing the article
Section titled “Generate — producing the article”Generate is a staged native pipeline (a dozen-plus steps: fetch, summarize, ontology alignment, drafting, humanization, structure, images, validation, final layout). Two properties define it:
- Grounding. Numbers and claims are drawn from the research corpus and live sources retrieved for the piece — with a relevance threshold on what the retriever may cite.
- Fail-loud, versioned execution. The pipeline runs as an immutable version resolved from the catalog (per-contract pin available); any step error fails the run visibly rather than degrading silently. See Version lifecycle.
The publish path and its gates
Section titled “The publish path and its gates”Publication converges into a single choke-point used by both the nightly batch and API-triggered publishes. On that path, today:
| Gate | What it checks | On failure |
|---|---|---|
| Topic dedup | Embedding similarity of the new piece against already-published topics for the contract, plus peers inside the same batch | The article is deferred, not discarded |
| Claim verification (grounding gate) | Numerical and normative claims are checked against retrieved sources; ungrounded claims block | The article is quarantined for review |
| Quality scoring | Mechanical cleanup plus a quality stamp: verification signals, brand-safety signals, language consistency | Recorded as telemetry with the article; feeds audits and remediation |
The order matters operationally: gates run before the hub is asked to build, so a rejected article costs nothing downstream.
Publication and proof
Section titled “Publication and proof”The approved article’s rendered HTML is shipped to the publishing hub, which rebuilds the destination site. The platform then verifies the live URL responds — only that verification counts as “published”. The source↔slug pair lands in the published-URL registry.
Measurement and feedback
Section titled “Measurement and feedback”Weekly search-console collection produces per-site time series (impressions, clicks, indexation), and per-topic performance feeds a topic feedback loop: topics that demonstrably earn impressions get more Collect quota; saturated or dead topics are throttled. Cost metering records every LLM call per run and contract, so the cost of a published article is a first-class metric.