Pipelines — purpose, evidence and remaining work

Verification status — 6 October 2026

The pipeline audit is still open. The August inventory below records what an earlier inspection reported; its “built” labels do not establish that the complete research journey works. The program audit separates current findings, remaining coverage and proposed repairs. No repair or automation activation is approved by this page.

A pipeline is an automated path that gets something into the library and keeps it current. These pages cover content collection, brand onboarding, campaign recognition and ongoing monitoring, together with their routing and prompts.

Earlier implementation inventory — 17 August 2026

The earlier inspection described one pipeline as fully built, two as partly built, and one as fragments. It described routing as built with a dormant part. These were implementation observations, not completed product acceptance. The diagram and tables below retain that inspection’s scope and date.

The August map

Have versus need — August snapshot

Pipeline	Where it stands
The content pipeline	Built. Fifteen stages, verified against the code stage by stage
Brand onboarding	Partly built — more than expected. Entity enrichment works; a five-stage brand-guidelines pipeline exists and has run for twelve brands. Missing: a products-and-features harvest, and the "onboard this brand" routing step
The campaign pipeline	Fragments. A synthesis script exists but only runs by hand; the evidence machinery has run a handful of times across two campaigns; nothing detects a campaign on its own
Continuous monitoring	Partly built. Newsroom ingestion works; a re-scan scheduler exists but ships switched off. Missing: the daily recognition pass and any real alerting
Routing & prompts	Built, with a dormant half. Sixteen live routes pick a prompt by source; the smarter semantic layer exists but is switched off; prompts are hand-written today, and the ruled future is prompts generated from the registry

What each pipeline reads and writes (the tie to the data model)

The earlier inspection reported these reads and writes on 17 August 2026:

Pipeline	Writes to	Watched through
The content pipeline	the piece's own fields + its classification links	the execution log — every model call with its stage and cost
Brand onboarding	the entity's facts · the brand-guideline sections	the same execution log
The campaign pipeline	campaigns · membership links · the two evidence surfaces	the intelligence run tables
Continuous monitoring	observations · risk signals · relationship events	the collection-run log; the alerts surface is built but empty
All of them	nothing runs on a schedule unless the single master switch is on — and it is off; every run so far was started by hand	one flag, one query

How to read the pipeline pages

Each page introduces its purpose, labels its historical inventory and links to later evidence. Approved requirements remain distinct from implementation observations. Follow the program audit for verified findings and remaining coverage. A missing capability is work to plan; it does not invalidate the approved model.