About & Methodology

Recent sync history

Known limitations

• FEC campaign finance data is reported on filing schedules. Quarterly filers may have data that is weeks or months old.
• Some members have FEC candidate IDs that point to old campaign committees (e.g., a senator's prior House campaign). This can result in missing or incomplete finance data for their current office.
• Stock disclosure coverage is incomplete and still under validation. Preview pages cannot establish that a member did or did not trade, and every row must be checked against its official filing.
• House PTR rows below 80% parse confidence are flagged for review rather than hidden. Automated confidence is a triage signal, not independent verification.
• Press release coverage depends on each member office publishing an RSS feed at a discoverable path. Offices without a feed are not represented in the timeline.
• Bill and vote coverage is limited to the 119th Congress. Earlier congresses are available upstream but have not been backfilled.
• Historical roll calls and bills are stored for current members only. A member who left office before the current Congress does not appear in vote or sponsorship records here.
• Territory delegates (DC, PR, GU, AS, MP, VI) have limited legislative data, they cannot vote on the House floor.
• The roster is sourced from @unitedstates/congress-legislators. Congress has 435 voting House seats, 100 Senate seats, and six non-voting delegate seats. The site currently tracks 537 sitting members, with 4 voting seats vacant pending special elections. These numbers are computed from the live database, not written by hand. Vacant seats have no member page until a successor is sworn in; members who die or leave office are marked out of office on the next nightly sync.
• Official and campaign biography text is the subject's own description. Publication means a reviewer confirmed the source and quote; it does not turn that statement into an independent biographical finding.

AI transparency

The Ask feature uses OpenAI GPT-5.6 Terra as its primary provider and Anthropic Claude Sonnet 5 as an independent fallback. OpenAI uses strict schemas for every tool. Anthropic uses a deterministic topic router to expose only the relevant strict retrieval schemas, keeping grammar compilation within the request deadline. Both providers send arguments through server-side type, range and scope validation before any query runs. A terminal answer tool is required, and a factual answer is rejected unless at least one record lookup was completed.

Exact SQL retrieval is used instead of embedding search for votes, bills, finance, committees, terms and races because those are structured records where identifiers and dates must match. Each Race answers dual-read state-authority candidacies where a verified adapter exists and FEC filings elsewhere, preserving the coverage label in the model context. Each answer shows the record categories checked. Stock disclosures are not exposed to the assistant while that feature is under validation.

Questions are capped in length, same-origin POST requests are enforced, provider retries are disabled, and one provider fallback is allowed only when the primary is unavailable. Per-connection and daily provider limits cap abuse and spend. Fresh questions are screened with a free moderation model before any paid provider call; that check fails open, so a moderation outage never blocks the feature. IP-derived identifiers and cache keys are HMACed or hashed; questions are not stored in plaintext in the answer cache.

Every question and answer is kept in an audit log for 90 days, keyed to a hashed connection identifier rather than an address. The log records the model's own status for each reply — answered, no matching record, out of scope, or declined — plus the records checked, the provider used, latency, and what share of the answer's sentences carry a validated citation. That makes any answer the assistant ever served reproducible and correctable, and the health page publishes live totals from it. Replies that are not grounded answers are labeled in the interface, no answer is reviewed by a person before display, and every answer carries a report link for corrections.

Two ingestion pipelines use models. House PTR PDFs are parsed with Anthropic Claude Sonnet 4.6 in vision mode. The model reads the rendered disclosure form and returns structured JSON, ticker, asset description, owner, transaction type, transaction date, amount band, plus a per-row confidence score (0–100). Every row is stored with its score; rows below 80% are flagged in the UI and the user can see exactly which rows the parser was uncertain about. Senate PTRs do not use AI: they come back as structured HTML and are parsed deterministically.

Campaign-site and official-member biography research uses OpenAI GPT-5.6 Terra with low reasoning and strict Structured Outputs, then falls back to Anthropic Claude Sonnet 5 when the primary provider is unavailable. Retrieval is deterministic and bounded: crawlers select only approved same-domain research pages, then supply their text directly. The model cannot browse or choose new URLs. Application code drops any output whose quote is not present in the cited snapshot, and a reviewer name and timestamp are stored before any fact can be published or supplied to Ask.

All other data — bills, sponsorships, votes, finance, committees and members — traces directly to an official API or community-maintained dataset, with no model in the loop.

Technical details

For journalists & site health

Reporters can bulk-download every dataset as CSV, with freshness timestamps and reporting tips, on the For Journalists page. The full pipeline status — per-source coverage, sync history, and any active issues — is public at /health.