Capitol Releases

A journalism and public-records project building the canonical "on the record" archive for U.S. elected officials: every original press release, statement, op-ed, floor speech, and social post, collected daily with full provenance.

Live at capitolreleases.com.

Coverage

Numbers below are from the production database, 2026-08-03.

Corpus	Size
Site releases (live records)	104,079
U.S. senators covered	100
U.S. House members covered	435
State and executive officials (beta)	257
Bluesky posts	18,678
Senate floor speeches (Congressional Record)	5,676
Published daily and weekly briefs	70

Coverage starts January 1, 2025 for Congress. Every record carries provenance — source URL, scrape run, date source, date confidence — and deletions at the source become tombstones, never removals.

Product surfaces

• Feed, search, trending — full-text search across the corpus with per-chamber and per-type filters.
• Member pages — per-official archive, release-volume analytics, and Ask the record: retrieval-augmented Q&A grounded in that member's collected releases, with server-validated citations and a disclosure footer. Vector retrieval over pgvector (OpenAI text-embedding-3-small), answers from Claude (ASK_MODEL, default Haiku 4.5); an answer that cites outside its retrieval set is discarded, not repaired.
• Daily brief (/brief) — AI-drafted summary of the day's releases, one Claude Sonnet call per edition, every claim cited back to source records. A weekly recap publishes on the same surface.
• Social (/social) — verified-handle Bluesky archive, kept separate from the press-release feed.
• Floor speeches (/speeches) — per-speaker segments from the daily Congressional Record.
• State tier (beta) — first state jurisdictions live, including Texas surfaces (/texas) and the Colorado caucus-pressroom model.
• Admin (/admin) — run history, health checks, open alerts. Google OAuth via better-auth; admin identity comes from ADMIN_EMAIL.

Repository layout

Path	Purpose
app/	Next.js 16 frontend (App Router, React 19, Tailwind 4, D3)
pipeline/	Python collection pipeline: collectors, CLI, tests, repair scripts
pipeline/seeds/	Per-source config: URL, CMS family, selectors, collection method
db/	Postgres schema and migrations
learning/rag/	RAG build log: ADRs, eval results, golden dataset
.github/workflows/	CI gate plus daily/weekly cron pipelines

Stack

• Frontend — Next.js 16, React 19, Tailwind 4, TypeScript, D3
• Pipeline — Python 3.14, httpx (async), BeautifulSoup + lxml, Playwright, feedparser
• Database — Postgres on Neon: tsvector full-text search, pgvector for retrieval
• AI — Claude Sonnet drafts the briefs; Claude Haiku answers Ask the record and runs advisory post-collection validation; OpenAI embeddings for retrieval. AI never writes to the corpus.
• Auth — better-auth with Google OAuth, Drizzle adapter
• Hosting — Vercel (frontend), Neon (database), GitHub Actions (cron + CI)

Getting started

pnpm install
pnpm dev

Open http://localhost:3003. Set DATABASE_URL in .env.local. ANTHROPIC_API_KEY and OPENAI_API_KEY enable Ask the record; without them the section stays hidden and everything else works.

Pipeline CLI

python -m pipeline update          # collect new releases (cron runs this 4x daily)
python -m pipeline health          # per-member pre-scrape canary
python -m pipeline test            # data-quality suite against the live corpus
python -m pipeline back-coverage   # flag members with truncated archives
python -m pipeline deletions       # detect source-deleted releases
python -m pipeline brief           # generate the daily brief (--weekly for recap)
python -m pipeline floor-speeches  # collect the Congressional Record
python -m pipeline daily-report    # operator digest email
python -m pipeline stats           # database overview

Additional commands: repair, review, verify-visual, tiers, source-profiles, sync-members.

Workflows

Workflow	Schedule	Job
ci.yml	push / PR	eslint, tsc, non-DB pytest
daily.yml	4x daily	collect, embed, data-quality gate
daily-digest.yml	nightly	operator digest to ALERT_EMAIL
brief.yml	Tue–Sat	daily brief
brief-weekly.yml	Fri	weekly recap
weekly.yml	weekly	deep checks

Design principles

1. Determinism first. AI assists, never drives. Every database write is traceable to a collector run.
2. Per-member accountability. A broken collector must not hide in hundreds of healthy ones.
3. Provenance everywhere. Every date carries date_source and date_confidence. Every record carries source_url, scrape_run, and scraped_at.
4. Collect wide, surface narrow. Store everything original. Show press releases by default; classify the rest.
5. No silent failures. Zero records for a member is an alert unless explicitly expected. Steps that may fail softly have tripwire tests that surface in the daily digest.
6. Archival permanence. Never hard-delete. Source-deleted releases become tombstones with deleted_at set.
7. Config over code. Adding a member, chamber, or jurisdiction is a seed-file change, not a rewrite.

Scope

• Current holders only. Where a seat changed hands during the window, only the current holder's releases are collected.
• Original content only. Curated third-party clippings and "In the News" mentions are skipped.
• Official channels only. Official .gov sites, verified social accounts, and the Congressional Record — no campaign content.

Status

U.S. Senate and House run at production quality with daily collection and a 30-plus-test data-quality gate; documented gaps remain on a small number of CMS-truncated archives. The state tier is beta. See docs/devlog.md (gitignored) for session-level history.

Roadmap

• Re-launch subscriber email with double opt-in and per-IP rate limiting on subscribe. The first iteration shipped without either and was retired 2026-08-03 before any subscriber signed up; the weekly recap continues to publish on-site. The send path is preserved in pipeline/commands/brief_send.py.

License

All rights reserved. Source code is public for transparency; reuse requires permission.