Nemotron 3 Ultra
Parti démocrate 68 %
90.0
GLM 5.2
Parti démocrate 65 %
88.3
HY3
Parti démocrate 65 %
86.7
PRE-REGISTERED · AUG–SEP 2026
The machines
called the
midterms.
Three language models. Eleven questions. Two elections. Their forecasts are public and verbatim — scored for the quality of the call, not for luck.
- Models
- 3
- Questions
- 11
- Elections
- 2
- Results
- Pending
01 / The calls
Four questions that define the night
Each line is the model's own pick, shown before any result. The number on the right is the judge's forecast-quality score (0–100).
Nemotron 3 Ultra
Parti républicain 70 %
91.7
GLM 5.2
Parti républicain conserve le Sénat, 78 %
88.3
HY3
Parti démocrate 65%
93.3
Nemotron 3 Ultra
Parti républicain 65 %
90.0
GLM 5.2
Aucun pronostic final (texte tronqué avant conclusion)
6.7
HY3
Parti républicain 75 %
85.0
Nemotron 3 Ultra
Marine Le Pen (RN) 55%
95.0
GLM 5.2
Marine Le Pen (RN) ~40 %, Édouard Philippe ~20-25 %
83.3
HY3
Jordan Bardella (Rassemblement National) 55 %
88.3
02 / Quality board
Who reasons best about an uncertain future
The judge rates each answer on three axes: specificity (a named candidate or party and a probability), grounding (polls and electoral structure) and calibration (a probability that matches the uncertainty).
| # | Model | Overall | France 2027 | US 2026 | Quality |
|---|---|---|---|---|---|
| 01 | Nemotron 3 Ultra (free) | 89.1 | 88.9 | 89.3 | |
| 02 | GLM 5.2 | 78.0 | 85.2 | 70.7 | |
| 03 | HY3 (free) | 76.5 | 83.3 | 69.7 |
Judge: DeepSeek V4 Pro · temperature 0 · responses public verbatim. See the full leaderboard ↗
03 / Where they agree
One call, three hands
Every model expects the House to flip Democratic, with probabilities clustered between 65% and 68%. That is the closest thing to a consensus on the board — and the easiest call to check on election night.
04 / Where they split
The Senate is the fault line
Two models see Republicans holding the Senate (70% and 78%); one sees a Democratic flip at 65%. On 2027, two name Marine Le Pen and one names Jordan Bardella. The disagreement is the story.
05 / Method
Quality is not accuracy
Nobody knows the results yet. A model can be wrong and score well, and a vague answer always scores badly. On election night we add the second layer: winner accuracy, calibration and Brier score against the certified results.
Specificity
A named candidate or party and a stated probability. “It depends” scores zero.
Grounding
Reasoning anchored in polls, incumbency, district structure and history.
Calibration
A 55% call should not read like a certainty, nor like a coin flip.