Numinous Rebuilds SN6 to Reward Forecasters Who Move First

Numinous (SN6) is scrapping its self-generated question format for pure prediction market questions, and scoring the entire probability curve rather than checking miners against isolated point forecasts.

The change follows early tests where the subnet’s news-only belief curves Granger-led Polymarket prices on multiple high-impact geopolitical events, meaning forecasters were arriving at accurate probabilities before markets themselves repriced.

A memory component now lets each forecaster carry its last probability and rationale forward, updating continuously as new signals arrive.

The skill the subnet measures has moved from single-point accuracy to whose belief curve actually leads the market.

What the Old Scoring Missed

The prior approach relied on generated questions, and the noise that setup introduced turned out to be a hard constraint on the mechanism.

1. Short-term questions were the only viable format: Miners cannot wait a month to be scored.

2. Good short-term proxies are hard to build: Generating them at scale introduces noise. Resolving them accurately introduces more.

3. Prediction markets solve both problems: Market price becomes a continuous proxy for the underlying belief, and resolution is handled by the market itself.

4. Long-horizon question generation still has a role: The team plans to refocus its generator on questions where synthetic content adds value markets do not already price.

How Belief Curves Front-Run the Market

Using the Numinous Signals websocket, the team built a forecaster that consumes high-impact news continuously and updates its forecast every time a significant signal arrives, following the belief-updating design from FutureSim.

The best forecasters on Numinous are becoming real-time market infrastructure.

The Signal Stream API is now live, giving builders and traders a WebSocket feed of market-specific signals for prediction markets.

Each signal is scored by competitively selected, Brier-winning… pic.twitter.com/TPe3WiRX7Q

— Numinous (@numinous_ai) June 10, 2026

1. News-only input: No other news sources, no tools, no internet, no market prices.

2. Blind priors: The forecaster formed its own priors without seeing the market and updated from news alone.

3. Memory carried forward: Last probability and rationale persist across updates rather than resetting each cycle.

4. Result: In several scenarios, the news-only belief curve led the Polymarket price.

The team formally tested this with Granger causality on 15-minute grids of price and belief changes. On 16 markets tested across two weeks, the belief curve significantly Granger-led the Polymarket price in 4 markets at p less than 0.05, from a news feed alone with zero price input.

How Miners Get Scored Now

The subnet will use a difficulty-adjusted score with the Polymarket price as the benchmark, borrowing from ForecastBench methodology. Every prediction is scored against what the market believed at that exact moment.

1. The formula: Score equals (prediction minus target) squared, minus (market price minus target) squared. Lower is better. Negative means the miner beat the market.

2. Proxy target while unresolved: The Polymarket price serves as the proxy so the entire belief curve gets scored continuously.

3. Objective resolution when the outcome fixes: When the target becomes 0 or 1, the expression reduces to the difficulty-adjusted Brier score against the market.

4. Market selection: The subnet will prioritize markets with the highest jump rates, where stale forecasters get most heavily penalized and belief-updating skill separates cleanly from noise.

Three properties fall out of the construction: continuous scoring across the whole probability curve, difficulty adjustment that rewards accurate prediction of sharp moves, and objective resolution against a stable set of long-term questions.

Ahead of the Reprice

The redesign trades a noisy internal benchmark for the sharpest external one available. Prediction markets already aggregate real capital, real conviction, and real resolution into every price point they publish.

Beating them consistently with news alone is the specific skill Numinous now measures and pays for, and the early Granger results suggest the mechanism has serious room to run.

Miners iterating against a live market benchmark should compound quickly, which is exactly the flywheel a self-improving prediction subnet needs.