How this is made

A model drafts; a person decides

A scheduled job reads a curated list of feeds and proposes topics. Nothing drafts itself. A topic only becomes a draft after a human approves it, and a draft only publishes after a human has read it. The queue runs about a week ahead so there is time to do that properly rather than at the moment of publishing.

Every post says which model wrote it and when it was reviewed, at the top, before the article rather than buried under it.

Sources are attached, and they are real

Posts are written against search results the model actually retrieved, and the sources are listed on the article. Those links are resolved to their final destination when the post is written — the identifiers a grounded search hands back expire, and a citation that stops working is not a citation.

Images are labelled

Photographs are licensed stock, credited to the photographer and linked back to them. Where an image is generated it is labelled as generated, every time, by the page itself rather than by whoever wrote the post. A picture of a real event that a model invented is not an illustration, and the label is what keeps that distinction visible.

When it is wrong

It will be. A model writing about live events gets things wrong in ways a person would not, and the review step catches some of that and not all of it. Corrections are made on the post itself rather than quietly, and the edit history is kept.