Ember

What is Ember?

Ember is a live AI prediction experiment that runs three independent AI models—Claude (Anthropic), Grok (xAI), and Gemini (Google)—against real Polymarket markets every day. Each model makes a probability call before the market resolves, and every call is logged, scored, and published publicly. The project runs on a strict 365-day methodology with locked rules, public Brier scores, and zero editing of predictions after they're made.

Who it's for

• Traders who want real-time AI divergence alerts when model probabilities split from crowd sentiment, helping spot mispriced markets.
• Developers who need a production API delivering three-model probability data in JSON format, with webhooks on the roadmap.
• Researchers who value a public, reproducible scoreboard with locked methodology and a full 365-day arc of Brier-scored predictions.

Key features

Three independent AI models

Ember runs Claude, Grok, and Gemini on the same market questions daily. Each model uses a different reasoning approach—Claude synthesizes prediction markets and forecaster communities, Grok reads live X sentiment, and Gemini grounds calls in search results. When models disagree, that divergence becomes the signal.

Public Brier scoring

Every resolved call is scored using the Brier score, a standard accuracy metric for probabilistic predictions. The scoreboard is fully public and updated daily. After 8 days live, Ember holds a Brier of 0.0365 against a crowd score of 0.0356 across 157 resolved calls.

Locked methodology with correction transparency

Predictions are never edited after being made. When infrastructure issues affect data integrity, Ember publishes public correction notices—9 so far. The methodology stays locked for the full 365-day arc, with a Year 2 evaluation beginning at Day 300.

Live divergence alerts

When a model's probability diverges significantly from the crowd market price, Ember flags it as a high-conviction signal. These alerts are available in real-time to subscribers and highlight potential mispricings before they resolve.

What stands out

Three AI models call live markets daily, independently, before the outcome. Nothing edited. Every divergence logged and scored.

Ember's edge is its commitment to radical transparency in AI prediction. Unlike black-box forecasting systems, Ember publishes its full record—including correction notices for infrastructure issues—and never retroactively changes a call. The 365-day experiment forces the models to build a track record over time, making the Brier scores meaningful for evaluating whether AI can actually predict its own trajectory.