Content pipeline where every fact has a source
Most AI content pipelines optimise for volume and leak invented facts. I run this one myself: 50 catalogue cards are live in production, built by exactly this pipeline, and the verify gate is covered by tests that run the real production config. It is built backwards from the failure mode: the gate is in the data layer, so an unverified card cannot ship even if every human forgets to check.
How it works
Four steps, and the gate is the product. Everything else exists to feed a check that refuses to publish an unsourced fact.
The photo stays the photo
Files arrive from a shared drive. Images are processed on-device – no generative model ever touches the product shot, so the photo on the site is the photo of the thing.
Every fact names its source
AI reads the label and fills the facts – and every fact must name the source it came from, with a URL and a retrieval date.
Majority, not best-of
A gate re-reads suspicious cards and takes the majority verdict of three reads – majority, not best-of, because best-of-three is just a retry-until-green loop wearing a gate costume.
The build refuses the rest
The static build refuses to render any card still in a needs-review state. The site builds with zero API keys – content is git, and every change is a reviewable diff.
What you gain
- Zero unsourced facts published – that is the product goal, verbatim, and the schema enforces it
- Generated copy (tasting notes, serving ideas) is labeled as generated – never passed off as fact
- The whole content state is version-controlled: audits are a git log, not an archaeology project
Honest limits
- A strict gate means slower publishing: cards wait in needs-review until a source exists. That is the trade and it is the point
- On-device image processing is tuned per platform; a different asset type means re-tuning, not copy-paste
indicative, not a quoted delivery date
What does a wrong fact cost you?
Tell me on a 30-minute call – I'll show you where the gate would sit in your catalogue.