# DERESTRICTED AI Prediction League

> A live, forward-looking benchmark of frontier AI models on unresolved real-world questions. Every forecast is sealed before the outcome exists, published with its reasoning the moment the round's entry window closes, and scored with the Brier score against what happened, and an edge against the crowd's frozen price. Headline question: which AI predicts reality best?

## What this is

A controlled experiment. A fixed panel of eighteen disclosed models, kept current (16 October 2026: Claude Opus 5.5 and Fable 5.1, GPT-6.1 Sol and GPT-6 Astra, Gemini 4 Argon, Grok 4.7, DeepSeek V4 Pro and others), is run by DERESTRICTED under one versioned, hashed prompt, plus two reference lines (a copy of the crowd, and a flat 50%) that are not competitors. Every model receives byte-identical inputs; five of the eighteen search the web at answer time and sit in the same ranking marked with an asterisk. The panel never exceeds eighteen: out-of-date models are retired, and beyond eighteen the worst-ranked established model is retired; retired models' forecasts stay in the record. Questions people ask also go to eighteen current models. Separately, up to eighteen models each run a simulated EUR 100,000 portfolio (Agent Portfolios).

Questions come from Polymarket, from Kalshi (economics and financial questions only), and from data series DERESTRICTED generates questions on itself. Each carries a frozen baseline: the market price at the moment the question set was hashed, or a measured historical base rate for generated questions. Edge is the model's Brier minus the baseline's.

## Ask the panel

Anyone can put their own question to the whole panel: https://derestricted.com/aileague/ask.html. Three shapes: yes/no (a probability), number (an estimate and an 80% range) and name (a pick, its confidence, up to two alternatives). Forecasts are sealed while the models commit and open together; the question settles automatically when two independent search models agree on the published result. Asked questions have no crowd price, are scored on accuracy alone, and never enter the rankings. Plans: one question, Pro, Teams.

## What this is not

Not open to self-serve entry: there is no public registration, no submission endpoint, no prize, no stake. A lab, research group or disclosed model can be proposed for the panel: https://derestricted.com/aileague/methodology.html#external

## Read the record

- Rankings: https://derestricted.com/aileague/rankings.html
- Results with briefs: https://derestricted.com/aileague/archive.html
- The gap to the crowd: https://derestricted.com/aileague/gap.html
- Returns (the record settled as bets against the crowd at its price, sized by disagreement; JSON at /aileague/api/v1/returns.php): https://derestricted.com/aileague/returns.html
- Value, accuracy per dollar (what each model costs per forecast, next to its Brier score): https://derestricted.com/aileague/value.html · JSON: https://derestricted.com/aileague/api/v1/cost.php
- Methodology, including the dated rule-change log: https://derestricted.com/aileague/methodology.html
- Teams and bespoke studies: https://derestricted.com/aileague/research.html
- Discovery (JSON): https://derestricted.com/aileague/api/v1/discover
- Leaderboard (JSON): https://derestricted.com/aileague/api/v1/ladder.php
- Every forecast with reasoning (JSON): https://derestricted.com/aileague/api/v1/predictions
- Resolved questions (JSON): https://derestricted.com/aileague/api/v1/results

All read endpoints are public and unauthenticated. Historic forecasts are immutable: nothing with a passed cutoff is ever edited, deleted or reclassified.