PROTOCOL · v1.0
How the call
is scored.
A forecast is not a fact. We score the quality of the reasoning that produced it, then — after the vote — we score the outcome.
1. Pre-registration
The answers were collected on 26 August and 1 September 2026, months before the 3 November midterms and the 2027 French presidential election. They are published verbatim and are the benchmark: nothing is rewritten after the fact.
2. The eleven questions
Five US questions: House control, Senate control, Senate seat balance, governors, House popular vote. Six French questions: winner, runoff pairing, RN candidate, Macronist candidate, left candidate, first-round turnout.
3. The judge
Each answer is rated by DeepSeek V4 Pro at temperature 0 on three axes, each 0–100:
- Specificity — a named candidate/party and a stated probability.
- Grounding — polls, incumbency, district structure, history.
- Calibration — a probability consistent with real uncertainty.
The category score is the mean of its questions; the overall score is the mean of the two categories. A refusal scores 0.
4. What this is not
This is not a prediction market, not a polling average and not voting advice. A high score means the model made a clear, grounded, calibrated forecast — not that it will be right. A model can be wrong and score well; a vague answer always scores badly.
5. After the vote
Once results are certified, we add a second layer: winner accuracy, calibration curves and the Brier score for the probabilistic calls. Both layers are published side by side, so quality and accuracy are never conflated.
6. Cite this
LFORLA (2026). Election Predictions — US 2026 & France 2027 (v1.0) [Data set]. https://elections.lforla.org/data.html