Skip to content
ReframedNewsLeft · Right · Centered
Quick ReadArchiveAboutSign in

How Reframed works

How a briefing is built, step by step

← Back to Reframed.News

Reframed uses AI models and published reporting to build its daily briefing. This page explains how we build a briefing and some of the ways it can fall short.

Choosing the stories

For the daily briefing, we pull heavily covered events and run extra searches in Event Registry, a service that groups reporting by event. After an initial filter, we ask an AI model to rate candidates on their policy stakes, relevance to American readers, and national significance. U.S. relevance helps a world story rank higher, and broad global consequences count too. A daily briefing has up to 18 stories drawn from six categories. In a separate pass, we ask the scoring model to look for duplicates and poor fits on the shortlist.

The facts

A model reads article excerpts and is asked to write a short summary and a list of facts supported by at least two independent sources. That is a request we make of the model, not a rule we enforce, so a published fact can rest on a single article. We ask the writer to cite each fact, and we show a source link when one is available.

We then ask a second model to check each fact against excerpts from the articles it cites. If it flags a fact, the facts are written again once. If the new version does not pass either, the first version can still be published, and when the check itself is unavailable the facts can go out unchecked. A fact can be wrong.

The left and right reads

Many stories carry two short reads: how people on the left are likely to see it, and how people on the right are likely to see it. They are AI interpretations, not quotes from real people. Both start from the same facts, and a side’s read can also draw on a quote from that side’s own coverage. We draft the left’s read first and ask the right’s writer to make its own case.

Another model checks each read for anything the facts do not support and asks for one rewrite when it finds something; if the rewrite fails too, the first read can still be shown. If either read cannot be made short enough, neither is shown.

When we have coverage from outlets on a side, we ask the writer of that side’s short bullets to borrow their wording and stick to the shared facts. Which outlets count as left or right comes from a list we keep ourselves, and outlets we treat as centrist are not on it.

Agreement and splits

Some stories show where the left and right agree; others show where they split. A model checks whether the two reads agree, and a story can also be shown as agreement when it has a line about shared ground and none about a split.

What gets left out

We exclude stories the facts writer flags for thin sourcing, or that it fails to answer that question for. We screen the main list for stories with no facts, and a final review model can remove individual stories from the briefing. If that final review is unavailable, the briefing can go out without it. We do not yet say on the page when a story has been left out.

The pictures

For the lead story in the daily email, an AI image model makes an illustration and a second model checks it before we use it. When no picture passes, we use a drawn scene instead.

What people do

The briefing normally goes out automatically, and we review published briefings. If the final review grades the briefing a D, we try to hold the email and alert the team for manual review. More about who we are and why we built this is on the About page.

Corrections

If something is wrong, write to hello@reframed.news. We do not publish a list of corrections on this site.

ReframedNews

A daily news briefing on the day’s top stories, built to show both sides.

Written by AI from news reporting. We review published briefings. Who writes this →

Navigate

Quick ReadArchiveSettings

Company

AboutHow it worksSkylark CreationsSign InTerms of ServicePrivacy Policy

© 2026 Reframed.News

Made by Skylark Creations