What gets read (the inputs)

A scheduled job, four times a day, reads public signals:

• License filings. Virginia ABC retail-license applications and similar state/county records.
• Permits. Building permits and certificate-of-occupancy filings classified as restaurant or food-service.
• Health-department records. New-establishment inspections, which are filed immediately before a restaurant can open.
• Official social and press. Posts from the restaurants, landlords, and shopping-center developers themselves, plus their press releases.
• Targeted web search. Queries like "now open" and "coming soon" scoped to specific localities.

What gets written (the process)

1. Triage. Anthropic's Claude Haiku model reads each candidate lead and decides three things: Is this a real, novel restaurant event? Is it in our coverage area? Which of the gathered signals genuinely support the claim?
2. Corroboration. A lead only proceeds if it has either two independent supporting signals or one primary public record (a license filing, a permit, an official press release, etc.). Leads that don't meet the bar are held back or rejected. This rule is enforced in code, not just in the prompt.
3. Drafting. For corroborated leads, Anthropic's Claude Sonnet model writes an original short article — typically 120 to 180 words. The model is instructed that every factual claim must map to a listed source URL. If the available sources can't support the claims, the model returns a "needs more sourcing" verdict and no article is produced.
4. Human review. A human reads every draft, edits if needed, and approves, edits, or rejects it. Nothing publishes automatically.
5. Publish. Approved articles land here with their primary-source links and the AI-assisted disclosure visible at the bottom of every page.

What we won't do

• We won't scrape and rewrite other outlets. If The Burn, Eater DC, Washingtonian, or any other publication covered the story first, that's their work. We follow the link to the primary record and cite that — or we sit it out.
• We won't invent details. No fabricated quotes, no guessed prices, no made-up opening dates. If a fact isn't in the underlying source, it isn't in the article.
• We won't backdate. A small archive of clearly-labelled historical events is allowed, but every historical entry is tagged backfill and dated to its real event date — never dated to imply we were publishing earlier than we were.
• We won't auto-publish (with one narrow future exception: very high-confidence "now open" or "closing" items backed by a primary record may eventually clear human review automatically. Every such auto-publish is logged to an audit trail.)

Tips and corrections

If you're a business owner and we got something wrong, write to [email protected] and we'll fix it the same day. If you spotted a story we missed, send the link to the primary record — a permit, a press release, an official post — and we'll run it through the pipeline.

The stack

The pipeline is open source on GitHub. It runs on a small DigitalOcean server, costs a few dollars a month in API calls, and is built with Node.js, SQLite, the Anthropic SDK (Claude Haiku for triage, Claude Sonnet for drafting), and Astro for the static site you're reading now.