Methodology

Data sources

Every procedural field on a judge page comes from one of four source types. The source type appears next to the field on the page.

• Court website / standing order. The originating page on the court's site, or a linked PDF order. Extracted with deterministic rules and cited with a short excerpt.
• PACER-derived statistic. Computed from public docket data — for example, motion-ruling cadence (median days from full submission to order) or bench-ruling rate.
• Aggregated contributor observation. Verified-attorney observations, aggregated across many submissions before any field is published. No individual observation appears on the site.
• Verified court page. A field copied verbatim from a court-published procedure page, linked and dated. Used for stable facts like courtroom equipment lists.

Pipeline

1. Scrape. One scraper per jurisdiction fetches the judge index, standing orders, and chambers-rules pages from the court's site. Scrapers live in scrapers/ and run on a weekly schedule.
2. Extract. Shared extractors at scrapers/common/extractors.py turn standing-order text into structured fields. Extraction is regex and structural HTML parsing — no language models.
3. Persist. Output is written as YAML to data/judges/{jurisdiction}/. Every procedural field is paired with a sources entry containing the originating URL and a short excerpt.
4. Validate. scripts/lint-judge-yaml.py enforces the JSON Schema at data/schema/judge.schema.json. CI fails on any schema error.
5. Review. The weekly scraper workflow commits directly to main. Maintainers review the resulting Git diff and roll back anomalies. Human-authored corrections go through the normal pull-request review.
6. Render. Astro builds a fully static site from the validated YAML. Judge pages are pre-rendered HTML with no client-side JavaScript and no runtime database.

What the project does not do

• No AI in the data pipeline. Extraction is deterministic. The output of a scraper run is reproducible from its inputs. This is a hard constraint, not a temporary stance.
• No subjective ratings. The schema does not contain rating fields. Pages describe procedure and cite where each fact came from.
• No silent updates. Every change is a Git commit. Scraper-driven updates are visible in the public history and reviewable as diffs.

Confidence levels

Each sources entry on a judge page carries a confidence label.

• Verified. A maintainer has confirmed the field against the cited source.
• Auto-extracted. The field came from a deterministic extractor with no human verification yet. The cited excerpt is the matched text.
• Aggregated. The field came from aggregated contributor observations. No individual observation appears.

Corrections and disputes

Corrections require a citation. Anyone may file an issue with the data:correction label or open a pull request against the YAML. If a correction is disputed by the cited source, the standing order or court page is the controlling authority and the field is updated to reflect it.

Where a court page and an aggregated observation conflict, the court page wins, but both are visible — the field reflects the rule, and the observation is preserved for context.

Update cadence

Scrapers run weekly. Standing orders and court pages are re-fetched on that cadence; the last fetched timestamp on each entry records when. Aggregated-observation fields are recomputed when new contributions cross the aggregation threshold for that field. The last verified timestamp on each sources entry records the most recent confirmation.