fedcourts.ai
The public site for FedCourtsAI. Static, built with Astro, hosted on Cloudflare Pages.
How it fits together
- Content is Markdown/MDX under
src/pages/. Edit it in the GitHub UI, open a PR, and Cloudflare Pages attaches a preview URL to the branch.
- Data is never hand-edited.
scripts/fetch-ledger.mjs runs before every build (prebuild), downloads the fedcourtsai repo tarball, counts what is in data/cases, copies the committed metrics/ roll-ups, and writes src/data/ledger.json (gitignored). Components read from that file.
- Results come from
scripts/fetch-results.mjs, which runs after the ledger fetch. The list of releases is the release manifest committed in fedcourtsai (docs/releases.json), read from main, so publishing a release needs no change here. Each release it lists is read from the fedcourtsai repo at its own tag (for Release 1, docs/release-ot2026-public-summary.md at results/ot2026-longconf) and nowhere else. /results lists every release, newest first, and each published one gets its own page at /results/<slug>; a release whose tag doesn't exist yet is listed as upcoming, and the first rebuild after the tag is pushed publishes it. A manifest that cannot be read, does not parse or lists no releases fails the build, and so, once a release's tag exists, does an unreadable summary or any unfilled ‹…› placeholder, so Pages keeps the last good deploy.
- Freshness comes from
.github/workflows/rebuild.yml, which runs the same build daily and calls the Cloudflare deploy hook only if it passes, so a ledger change that breaks the site shows up as a red scheduled run instead of a silently failed deploy. It never commits here.
- Status comes from
scripts/fetch-status.mjs, which copies status.json from the fedcourtsai repo's ops-metrics branch to src/data/status.json (gitignored). The pipeline's run-ops job writes that file every morning, before the daily rebuild: which scheduled jobs are on time, what they produced over 7 days, 30 days and the Term, coverage, and the back-test headlines. /status lays it out as written. A malformed file or an unexpected schema fails the build; a file not yet published leaves the page saying so.
- Checks:
.github/workflows/build.yml runs npm ci and npm run build on every PR into main, and its build check is required before a PR can merge. .github/workflows/lint-actions.yml runs zizmor and actionlint on any change under .github/, and .github/dependabot.yml keeps the pinned actions and npm dependencies current. Agent conventions are in AGENTS.md.
Components available in MDX
| Component | What it renders |
|---|
<Stats /> | The headline counters with a ledger stamp |
<LedgerStat key="…" label="…" note="…" /> | One figure from ledger.counts (events_open, events_resolved, events_with_predictions, predictions, predictions_frozen, evaluations, leaderboard_events_scored, leaderboard_predictors_ranked) |
<LedgerTable limit={n} onlyOpen onlyResolved /> | The ledger table, newest first |
<Stamp /> | "ledger <sha> · <time>" pointing at the exact commit read |
Import them at the top of a page: import Stats from "../components/Stats.astro";
Local build (optional)
npm install
npm run build # runs the fetch first
npm run preview
npm run dev needs src/data/ledger.json to exist; run npm run fetch once first.
