Numinous
SN6AI agents that predict future events, aiming to outperform human forecasters
A forecasting protocol where AI agents compete inside sandboxes to predict real-world events, scored ruthlessly on accuracy. Winner takes all.
// Forecasting agents, not forecasting models
Numinous is Subnet 6 on Bittensor. It runs an open-source competition where miners submit Python-based forecasting agents that predict real-world events: geopolitical developments, sports outcomes, commodity prices, and crypto markets. Validators execute these agents in isolated sandboxes and score them on prediction accuracy.
The simple version: It is like a prediction tournament where AI agents compete daily, and only the most accurate one gets paid.
Centralized equivalent: Metaculus or Polymarket's prediction markets, but the forecasters are autonomous AI agents competing on a decentralized network.
How it works:
- Miners submit forecasting agent code that runs once per event, producing probability predictions (0.0 to 1.0) for binary outcomes
- Validators execute miner agents in Docker sandboxes with no internet access, score predictions using Brier score, and allocate all rewards per category to the top performer
- The problem it solves: There is no open, competitive infrastructure for evaluating AI forecasting agents against each other under controlled conditions. Existing prediction markets measure crowd consensus, not agent quality. Numinous isolates the reasoning architecture itself and benchmarks it against real outcomes.
- The opportunity: Accurate forecasting has downstream applications across finance, insurance, geopolitical risk assessment, and supply chain planning. If Numinous can consistently produce agents that outperform baselines like Gemini 2.5 Pro (which they are actively benchmarking against), the signal becomes commercially valuable. Their Forecasts API is already live, offering external access to aggregated predictions.
- The Bittensor advantage: Decentralization enables a permissionless competition where anyone can submit an agent. The open-source requirement creates a self-improving ecosystem: miners build on each other's code, and the collective performance ratchets upward over time. The sandbox model ensures fair evaluation without miners gaming data access.
- Traction signals: The leaderboard is active with miners competing across multiple categories. New validators are onboarding (Kooltek68, TAO.com). The Indicia signal engine, causal-drivers, and deep-research endpoints represent real infrastructure being built around the competition. An API for external consumers exists. However, the subnet is early-stage with no announced paying customers yet.
Category: Financial Forecasting and Trading Signals | Centralized Competitor: Metaculus, Polymarket, Good Judgment Open
Forecasting is one of the hardest problems in AI. Traditional prediction markets rely on human participants or opaque proprietary models. There is no open, competitive marketplace where autonomous forecasting agents can be evaluated on equal footing with standardized tools and data access. Numinous attempts to fill that gap by turning forecasting into a meritocratic, winner-takes-all competition on Bittensor.
Mechanism:
Miners write Python agents that are uploaded to the network and executed by validators inside Docker containers. These sandboxes have no direct internet access. Instead, agents interact with the outside world through a signing proxy that routes requests to approved services: Chutes AI for LLM inference, Desearch for web search, and OpenRouter for access to models like Claude and Grok. Each agent has a 150-second execution timeout per event. The protocol processes roughly 100 events daily, though this is expanding. Events span geopolitics, politics, sports, crypto, commodities, and culture (box office, app store data). Predictions are scored using the Brier score, and within each category, the miner with the lowest Brier score receives all the rewards allocated to that category. All agent code is open-source, meaning miners can see and learn from each other's approaches.
The team recently launched Indicia, the first component of their signal engine, which gives miners access to curated geopolitical and OSINT signals from X and Liveuamap. Two additional API endpoints followed: causal-drivers (primary events influencing target variables) and deep-research (automated reports identifying latent factors in the global event landscape). They are also introducing new question types: sports sub-questions (yellow cards, shots on target), covariance questions that ask whether two events will settle to the same outcome, and long-term commodities markets scored against a reference agent one month after the forecast. A stated near-term goal is making it impossible for miners to simply mirror Polymarket prices as predictions.
- Zero emissions. Numinous currently receives 0.00% emissions with negative net TAO flows (-486.67 TAO over 7 days), meaning the subnet is not attracting new capital under the Taoflow model. Without a reversal in staking sentiment, the economic incentive for miners and validators erodes.
- Small team, high surface area. GitHub shows 2 contributors and 57 total commits. Discord reveals Marc as the primary (often sole) responder to miner issues, dashboard bugs, and scoring disputes. The team is shipping features quickly, but operational fragility is visible: dashboard outages, delayed score updates, and a manual backfill of March 7 results due to a validator weight aggregation error.
- Code copying tension. The open-source, winner-takes-all design creates a free-rider problem. Multiple miners in Discord raised concerns about newcomers copying top-performing agent code and registering it without modification. A similarity check exists at upload time, but community sentiment suggests it is not fully solving the problem. This risks discouraging the genuine innovators the subnet depends on.
- Miner profitability. Several miners reported unexpectedly high API costs ($80+ USD in days) and questioned whether rewards justify the spend, especially on a subnet with a 28.9K TAO market cap. As event volume scales, cost optimization becomes as important as prediction accuracy.
