Unbiased is a score, not a claim
Any desk can say it is fair. This one writes its numbers down before the event, names the document that will settle them, and then publishes the arithmetic of how often it was right — including the times it was not.
Locked 2026-08-05 against model senate-2026.1. None can be edited.
Grading opens 3 November 2026, when the first certified canvass lands.
Lower is better. A coin scores 0.25. Blank until something is graded, because a score with no results behind it is a decoration.
Share of the coin's error removed. Zero means the desk added nothing but confidence.
Calibration
Being right is easy — call the favourite every time. Being calibrated is the hard claim: the things called at 70 per cent have to happen about seventy times in a hundred. This table stays empty until results arrive, and then it fills in whether or not it flatters the desk.
| Said | Calls | Graded | Happened | Gap |
|---|---|---|---|---|
| 0–10% | 15 | — | not yet | |
| 10–25% | 5 | — | not yet | |
| 25–40% | 3 | — | not yet | |
| 40–60% | 6 | — | not yet | |
| 60–75% | 7 | — | not yet | |
| 75–90% | 7 | — | not yet | |
| 90–100% | 10 | — | not yet |
The call sheet
53 open · 0 gradedEvery probability the desk has published, with the date it was locked and the document that will settle it. Named before the event, so nothing can be chosen after it.
| The claim | Said | Result |
|---|---|---|
Democrats hold at least 51 Senate seats in the 120th Congress senate-2026.1 · locked 2026-08-05 · settles 2026-11-03 | 18% | Open |
The Democratic nominee wins the North Carolina Senate seat senate-2026.1 · locked 2026-08-05 · settles 2026-11-03 | 66% | Open |
The Democratic nominee wins the Michigan Senate seat senate-2026.1 · locked 2026-08-05 · settles 2026-11-03 | 71% | Open |
The Democratic nominee wins the Ohio Senate seat senate-2026.1 · locked 2026-08-05 · settles 2026-11-03 | 26% | Open |
The Democratic nominee wins the Florida Senate seat senate-2026.1 · locked 2026-08-05 · settles 2026-11-03 | 19% | Open |
The Democratic nominee wins the New Hampshire Senate seat senate-2026.1 · locked 2026-08-05 · settles 2026-11-03 | 84% | Open |
The Democratic nominee wins the Georgia Senate seat senate-2026.1 · locked 2026-08-05 · settles 2026-11-03 | 85% | Open |
The Democratic nominee wins the Minnesota Senate seat senate-2026.1 · locked 2026-08-05 · settles 2026-11-03 | 88% | Open |
The Democratic nominee wins the Iowa Senate seat senate-2026.1 · locked 2026-08-05 · settles 2026-11-03 | 12% | Open |
The Democratic nominee wins the Alaska Senate seat senate-2026.1 · locked 2026-08-05 · settles 2026-11-03 | 12% | Open |
The Democratic nominee wins the Maine Senate seat senate-2026.1 · locked 2026-08-05 · settles 2026-11-03 | 89% | Open |
The Democratic nominee wins the Texas Senate seat senate-2026.1 · locked 2026-08-05 · settles 2026-11-03 | 10% | Open |
- A call is locked when it is published
The probability, the date and the model version are written together and never edited. If the desk changes its mind, that is a second call at a second date and both stay on this page. There is no revision history to hide because there are no revisions.
- The resolver is named before the event
Every call says what will settle it — here, the certified canvass from each secretary of state. A desk that picks its evidence after the result is not being graded, it is being flattered.
- Misses are printed at the same size as hits
A graded call carries one line explaining what the model got wrong, written whether it was right or not. Nothing on this page can be filtered to the wins.
- The score is computed here, in front of you
Every number on this page is derived from the call sheet in the same page load. There is no stored summary that could quietly disagree with the ledger under it.
- Beating a coin is not a result
The Brier score is printed beside two reference scores — always saying 50%, and always going all-in on the favourite. A desk that cannot beat both of those has not earned a probability.
Every probability above is read from an archive written at lock time — the number, its date and the model version that produced it, stored apart from the model itself. Retuning a coefficient tomorrow cannot move a call that was locked today. 53 calls are on the record and 0 live numbers are waiting to be locked. First grading opens 3 November 2026.
The model that produced these probabilities is printed in full on the method page, the certified inputs behind it on the 2024 reference layer, and the rules every piece is written against in the Standard.