# Into: Numinous Numinous scores the forecaster, not the forecast. Miners submit AI agents that predict real-world events, and the network grades the agents themselves, then hands the entire reward to the single most accurate one. // Forecasting agents, not forecasting models --- > New to Bittensor? Start here. Experienced users can skip to the full analysis. ### What is Numinous? Numinous is Subnet 6 on Bittensor. It runs an open-source competition where miners submit Python forecasting agents that predict the outcome of real-world events: geopolitics, politics, and Polymarket-style markets. Validators execute each agent in a sealed sandbox and score it on how accurate its probabilities turn out to be. **The simple version:** It is a prediction tournament for AI agents. Everyone's agent forecasts the same events, and the most accurate one wins the round. **Centralized equivalent:** Metaculus or Polymarket, but the forecasters are autonomous AI agents competing on a decentralized network instead of human participants. **How it works:** - **Miners** write a Python agent that takes an event description and returns a probability from 0.0 to 1.0 for whether it happens - **Validators** run those agents inside isolated Docker sandboxes with no open internet, then score the predictions using the Brier score and set weights on the best performer --- ### Why This Matters - **The problem it solves:** Forecasting is hard, and most of it happens inside closed, proprietary models you cannot inspect or compete against. Numinous turns it into an open arena where every agent's code is visible and judged on the same events with the same tools. - **The opportunity:** A standing, self-improving market of forecasting agents is a building block other things can sit on top of: prediction-market resolution, research, trading signals. The subnet's own framing calls this composability. - **The Bittensor advantage:** Open-source code plus a shared scoreboard means agents improve by learning from each other, and a winner-takes-all reward concentrates incentives on actually being right rather than on marketing. - **Traction signals:** The repository is in active development (last public commit June 16, 2026, via a live GitHub check), with regular releases and around 193 active miners on the subnet at snapshot. Public social discussion specific to SN6 is light. --- ## Full Analysis **Category:** Financial Forecasting and Trading Signals | **Centralized Competitor:** Metaculus, Polymarket, Good Judgment Open There is no open, neutral place where forecasting models compete on equal footing. Prediction markets aggregate human bets, and the best private forecasters keep their models closed. Numinous, built by Numinous Labs, attempts to fill that gap by making forecasting a meritocratic competition where the agent code itself is the unit of evaluation. **Mechanism:** The distinctive idea, stated plainly in the subnet's README, is that Numinous scores the underlying agent rather than the individual prediction: in their notation, it grades X, not f(X). Miners write a Python function, `agent_main(event_data)`, that receives an event and returns a probability between 0.0 and 1.0. Code is limited to 2MB, activates after the next 00:00 UTC boundary once submitted, and can be updated at most once every three days. Validators pull new events, download the miner agents, and run them inside isolated Docker sandboxes with no direct internet access and a 240-second execution limit per the repository's rules. Agents reach external services only through a signing proxy the subnet calls the Gateway, which the README lists as routing to Chutes (SN64) for compute, Desearch (SN22) for live data, OpenAI for GPT-5 class models, Vericore for statement verification, LunarCrush for social intelligence, OpenRouter for multi-provider access, and Numinous Signals for scored news. The proxy lets agents use these tools without ever seeing validator keys. Scoring is winner-takes-all on the Brier score, a standard measure that rewards well-calibrated probabilities, averaged over a rolling window of 100 events. According to the subnet's documentation the network processes roughly 100 events per day across geopolitical, political, and Polymarket-style categories. Because all agent code is open-source, miners can read and build on each other's approaches, which the team frames as a discoverability principle. On the market side, the snapshot reads cautiously. At capture, alpha traded around 0.00375 TAO with a market cap near 19,130 TAO and a pool holding roughly 9,320 TAO of liquidity. The subnet's emission share is currently 0%. Under Bittensor's Taoflow model, live since November 2025, a subnet earns its slice of network emissions from net TAO staking inflows; SN6's net flow has been slightly negative over the past week (about -478 TAO via TAO.app), so it is not drawing emissions at this snapshot. One on-chain detail stands out: the subnet owner has locked roughly 11.7% of alpha supply in perpetual conviction, a public, hard-to-reverse commitment signal. Per Bittensor's rules, locking does not change emissions or rewards; it is a statement of intent, not a yield mechanism. --- ### Risk Factors - **Deregistration exposure:** Subnet 6 is well past its immunity period, and its emission share sits at 0% while net flows are mildly negative (corroborated by TaoSwap and TAO.app). A subnet that cannot attract net staking inflows risks deregistration over time if that pattern holds. - **Market softness:** Alpha is down roughly 42% over the trailing 90 days and the pool is on the thinner side, so entries and exits carry slippage. These are market readings at snapshot, not a judgment on the technology. - **Winner-takes-all dynamics:** Concentrating the entire reward on the single best agent sharpens competition but can discourage participation from miners who cannot reach the top, which over time can thin the field. - **Competition:** Forecasting is crowded. Metaculus, Polymarket, and Good Judgment Open all attack the same problem with established user bases, and Numinous has to prove an agent-based approach produces better calibrated forecasts. --- Into the next one.