WARN Index

Problem

The WARN Act makes larger US employers file a notice about 60 days before a mass layoff or a plant closure. That’s a two-month head start for the affected workers and for anyone trying to hire them. But every state publishes the notices its own way: HTML tables, PDFs, CSV downloads, search forms, sometimes behind a WAF. And the sites change without warning.

The first version (2022) had about 33 hand-written scrapers that broke every time a state redesigned its site, plus a brittle Selenium enrichment scraper that produced most of the bad data. Keeping it running was the whole job.

Approach

The rebuild is designed around one assumption: scrapers will break, so make a break cheap to fix.

1. Fetch

   48 jurisdictions · daily CronJob

   ↓→

2. Parse + validate

   row-count range, required fields

   ↓→

3. Postgres

   dedup across re-filings

   ↓→

4. Enrich

   Claude + web search · every 6 h

   ↓→

5. API + SPA

   dashboard, map, reports

↺ On a parse failure the raw response is saved; /heal-scraper replays it, fixes parse(), and opens a PR for review.

• The fetch/parse split. Each state module fetches raw bytes and parses them separately. When parsing or validation fails, the runner saves the exact response as a snapshot, so the breakage can be reproduced offline, byte for byte.
• Repair is a reviewed PR, not magic. A Claude Code operation reproduces the break live, fixes parse() until it handles the new page and still passes the golden fixture, runs the tests, and opens a pull request. A human merges it, because merging is a production deploy.
• Enrichment with confidence scores. An LLM agent with web search fills in company details (industry, size, HQ) and records how confident it is. Low-confidence rows can be re-run selectively instead of being trusted blindly.
• Context, not just counts. Each layoff time series sits next to official BLS data (unemployment, payrolls, JOLTS) for the same period, so a spike can be read against the wider labor market.

Hard parts

• Hostile sites. Some state sites never reach network-idle, some reject plain HTTP clients, and one WAF resets TLS handshakes from anything that doesn’t look like a browser. Each needed its own fetch strategy: a real browser context, shared session cookies, or TLS fingerprint impersonation with retries.
• Silent failures. A malformed BLS series ID isn’t an error. It comes back as a valid series with no data. So the job checks coverage, not status codes, and tests pin every ID layout.
• Duplicates. States re-file and amend notices. Deduplication has to merge them without collapsing genuinely separate layoffs at the same employer.

Outcome

• 48 jurisdictions (47 states plus DC) scraped daily. The remaining three publish no usable public data, and each is documented.
• Enrichment every 6 hours, with re-runs for low-confidence rows.
• About 225 BLS series refreshed weekly as economic context.
• A breakage costs one reviewed PR, not a rewrite.
• Runs in production on the homelab: Kubernetes CronJobs, CloudNativePG, and Grafana and Alertmanager watching scraper health.

Stack