What we ship, what we won't, and why.

nwsly's briefs are AI-analyzed and human-checked. That makes the editorial floor higher than most AI-news products, not lower. Here's the workflow, the guardrails, and the things we explicitly refuse to do.

The workflow

1. Polling. The pipeline pulls RSS feeds from every source in our v1 list every few hours — national, city, and state outlets. URLs and headlines are stored, not the full articles.
2. Clustering. Stories that describe the same event get grouped using semantic similarity. A cluster typically has 3–8 sources covering the same event from different angles.
3. Selection. Five clusters get promoted each morning: three national, one city, and one state for the user's region. The three national slots start spectrum-balanced (left / center / right) for new readers, then personalize to each reader's revealed preferences — the bias positions and topics they engage with — while always reserving at least one national slot for a story outside the reader's lean, and subject to a factuality floor that excludes sources we rate unreliable regardless of lean.
4. Synthesis. Claude (Anthropic) analyzes each cluster and writes an 80–120 word brief, drawing only from the underlying articles. It's instructed to lead with the most important factual claim, close on stakes, and note when sources frame the event differently.
5. Human spot-check. A staff editor reads every brief before the 7am push. Factual errors, hallucinated quotes, or load-bearing claims absent from the source articles trigger a rewrite or a kill.
6. Ship. Briefs land on the user's phone at 7am local with source citations attached.

What we won't ship

• An unchecked brief.
• A brief whose load-bearing claims aren't traceable to the cited sources.
• A brief that quotes a source quote that doesn't actually appear in any cited article.
• A source below our factuality floor — no matter how hard a reader leans toward an outlet, we won't serve sources we rate unreliable to satisfy a preference. We personalize bias, never reliability.
• An all-one-side national deck — however strongly a reader leans, at least one of the three national stories always comes from outside their lean. We narrow the deck to taste; we never seal the bubble.
• A correction loop that's silent. When we kill or update a brief post-publication, we tell users.

How AI fits in

Claude writes the briefs. A human reads them. That's the distinction. We don't run an autonomous publishing loop and we won't until the model can clear a human-comparable error rate on multi-source news analysis — which is not today.

Specific things Claude does well in our workflow: synthesizing across politically-coded source sets without inheriting either side's framing, surfacing where coverage diverges, hitting word count, holding factual register. Specific things it does poorly: spotting load-bearing claims that aren't in the sources, recognizing when a source quote is paraphrased rather than verbatim, weighing the recency of dueling claims correctly.

The human editor exists for the second list, not the first.

Corrections

When we publish a brief that turns out to be wrong, we issue a correction in the next morning's edition with the original brief's headline and what changed. The brief permalink at /edition/{date} gets a struck-through correction notice. We don't silently rewrite history.

Source inclusion

Every source in our v1 list is profiled at /sources. The bias methodology is at /methodology. We add and remove sources based on a quarterly review; both events are announced.

Contact

Tip line / corrections / complaints: editorial@nwsly.co.