The Midterms 2026 board tracks U.S. Senate, House, and Governor races on the standard 7-point competitive scale, along with live congressional chamber balance and per-state news sentiment. It is an independent project — not affiliated with any campaign, party, or forecaster — built and operated by one person with substantial AI assistance, updated on a fixed weekly cycle with a public audit trail. This page explains exactly how every number is produced, verified, and corrected.
Race ratings
Ratings follow the 7-point scale (Solid D · Likely D · Lean D · Toss-up · Lean R · Likely R · Solid R) and are aggregated from professional forecasters — primarily the Cook Political Report, Sabato's Crystal Ball, and Inside Elections. This site does not run its own ratings model; when forecasters move a race, the dashboard follows, with the move and its source recorded in the weekly briefing.
Chamber balance
Current House and Senate composition comes from primary sources (House Press Gallery, Senate.gov, official clerk records), including vacancies and pending special elections.
News & sentiment
Headlines are collected weekly per competitive race via news search. Outlet political lean uses a fixed AllSides-style lookup (−2 left … +2 right) maintained in the open in news_config.json — it is an external classification, not this site's judgment of any outlet.
Scores are produced by an AI model (Anthropic's Claude), which reads each week's collected headlines and scores two things: coverage tone per race (0–100, where higher reads more favorable for Democrats; 45–55 is mixed) and candidate favorability (0–100 tone of coverage toward that person). A human reviews the output weekly before it publishes.
States without scored coverage display a rating baseline — a number derived mechanically from the forecaster rating (Solid D=75 … Solid R=25), shown dimmed and labeled. Baselines contain no news signal.
Known limitations
- Small weekly samples; a single big story can swing a race's tone score.
- One scorer (the model) — no inter-rater reliability measurement yet.
- Coverage tone is not a prediction. Favorable coverage ≠ winning.
- The outlet-lean table is static and coarse; local outlets are mostly unrated and excluded from lean-based charts (they still count toward mentions).
Every weekly update passes through four independent layers before it reaches the page:
| Layer | What it does | Runs |
|---|---|---|
| 1 · Independent AI fact-check | Up to three verification agents receive the week's claims — stripped of the sources and reasoning that produced them — and re-research each from scratch (forecaster sites, official records). Contradictions block publication until resolved by a primary source. | Weekly, before human review |
| 2 · Human review | Every change ships in a briefing with inline source links and a verification section; a human approves before anything is applied. | Weekly |
| 3 · Deterministic validation | Machine checks: the 435-seat House partition, Senate=100, the 7-point rating enum, schema and key integrity. Invalid changes cannot be applied. | Every change, automated |
| 4 · Post-publish assertion | After data is written, an automated job re-reads the live source and asserts the published values match the approved change set, row by row. A mismatch fails the run, and any error that reached this page is logged in the corrections log below. | Every apply, automated |
No AI system writes directly to the published data. All writes go through validated, deterministic code. The code itself is private; the record is public: every week's ratings in the history, and every correction and open question below.
The board closes with a retrospective. Every weekly snapshot since May is replayed and set against the final results once states certify them: what the pipeline collected, which forecasters moved which races and when, how coverage tone tracked the ratings, and what the verification layer caught along the way.
Weekly snapshots through November 2 are forecaster ratings only. After Election Day, a called race shows its result as Solid D or Solid R with a note marked RESULT; uncalled races keep their final forecast until they are called.
Substantive data errors are fixed promptly and logged here permanently — no silent corrections. Each entry records what was wrong, what's right, how it was caught, and what it affected. An empty log would be a red flag, not a badge: this project publishes hundreds of data points weekly, and honest operations surface mistakes.
Loading corrections…
Not every judgment call has a clean answer. When the data is genuinely ambiguous — a source that contradicts our own convention, a row whose note disagrees with its value — the weekly update logs the question here and takes the conservative option rather than guessing or quietly picking a side. This is the companion to the corrections log above: that one is what we got wrong, this one is what we are not yet sure about.
Loading open questions…
Updates land Mondays by ~1:15 PM PT. Mid-week corrections happen when news warrants (see the log above).