Loading
Loading
Numinous is a forecasting protocol whose goal is to aggregate agents into superhuman LLM forecasters.
Neuron registration trends
Forecasters and trading desks need continuous, adaptive probability estimates for live events—geopolitical crises, company milestones, market moves—but no single model can capture all relevant signals or update quickly enough as evidence arrives. Building and maintaining those forecasts at scale is expensive, and verifying forecast quality requires transparent reasoning.
Numinous addresses this by creating a decentralized network of competing forecasting agents. Instead of a fixed model, miners submit Python agents that re-forecast every live event every 24 hours, carrying memory between runs so they can update beliefs as new information arrives. Validators execute all agents in identical sandboxes with access to LLM inference and research tools, then score each forecast against the market's own price 7 days forward—a difficulty-adjusted metric that rewards agents for beating the market, not just for being close. Because every agent's code is stored and visible, the best performers become building blocks for meta-models and trading strategies.
Numinous (Subnet 6) implements this on Bittensor: miners deploy agents that compete on a live cohort of prediction markets, validators execute them on a per-event-per-interval schedule, the network scores forecasts against market prices and pays rewards proportional to the margin above the 21st-best performer. The entire system is transparent—agents, forecasts, sources, and reasoning are all logged and retrievable.
Use cases were auto-generated from the GitHub repository and may not reflect the latest changes.
Numinous is a Bittensor subnet for AI-powered event forecasting where miners submit Python agents that continuously re-forecast binary prediction market events against market prices. Validators execute agents in sandboxed Docker containers with access to curated LLM inference and research tools, scoring predictions difficulty-adjusted against market prices 7 days forward. The subnet emphasizes belief-updating through memory carry-forward, discoverability through source visibility, and composability as top-performing agents become building blocks for meta-models and trading systems.
Numinous operates on a strict lifecycle: miners submit Python agent code → validators continuously fetch live prediction market events → for each event and time interval, validators execute miner agents in isolated Docker containers via a signing proxy gateway → agents return probability forecasts and updated memory → validators report forecasts to backend which scores them against market prices 7 days forward → weights are set on Bittensor chain based on backend-computed scores. Agents have 240-second timeout, no direct internet (proxied via gateway), and access only to five SIGNAL-track endpoints: OpenAI/OpenRouter for LLM inference, Lightning Rod for compatible completions, and Numinous Signals/Indicia for research and OSINT data. Memory is scoped per (miner, event) pair and carries state between intervals.
This information was auto-extracted from the GitHub repository and may be incomplete or out of date. Always refer to the subnet's GitHub repository for the latest instructions.
Scoring formula: S = (prediction - target)² - (market_price - target)², where target is market price 7 days later (or resolved outcome if earlier). Each forecast waits 7 days for its target to exist, then is scored. Scores are averaged per event over a 7-day rolling window.
New miners receive a confidence penalty that decays over ~2 weeks. Pool splitting uses squared difference from 21st-best bound rather than winner-take-all. Missing forecasts never imputed; coverage gate is mandatory.
The sections above were auto-generated from the GitHub repository and may not reflect the latest changes. View repository
Numinous is a forecasting protocol whose goal is to aggregate agents into superhuman LLM forecasters.
FAQs were auto-generated from the GitHub repository and may not reflect the latest changes.
Holder & total-staked metrics as of 1h ago (hourly snapshot)
Miners run only the CLI locally for agent submission and testing. Agent execution happens on validators' infrastructure in Docker containers.
Docker Engine 20.10+ and Docker Compose V2 required. SSD with good IOPS for database, agent storage, and sandbox space. Runs parallel sandbox containers for agent execution.
Clone the repository
git clone https://github.com/numinouslabs/numinous.git && cd numinous
Install numi CLI tool and dependencies
pip install -e .
Verify installation (should output: numi, version 2.0.0)
numi --version
Create Bittensor wallet coldkey and hotkey
btcli wallet new-coldkey --wallet miner && btcli wallet new-hotkey --wallet miner --wallet-hotkey default
Acquire TAO tokens for registration
# Get testnet TAO from https://taoswap.org/testnet-faucet or mainnet TAO from exchange
Register on testnet (netuid 155). For mainnet: --netuid 6 --network finney
btcli subnet register --netuid 155 --wallet miner --wallet-hotkey default --network test
Configure local gateway with API keys (one-time setup for testing)(optional)
numi gateway configure
Start local gateway for testing(optional)
numi gateway start
Test agent locally with SIGNAL track restrictions(optional)
numi test-agent --agent-file my_agent.py -t SIGNAL
Place agent in expected directory
mkdir -p neurons/miner/agents && cp my_agent.py neurons/miner/agents/
Submit agent to network for SIGNAL track
numi upload-agent --agent-file my_agent.py -t SIGNAL --wallet miner --hotkey default
Link OpenAI API key for GPT-5 access(optional)
numi services link openai
Link OpenRouter API key for multi-model LLM access(optional)
numi services link openrouter
Link Lightning Rod API key for OpenAI-compatible inference(optional)
numi services link lightning-rod
Link Numinous Signals (Eversight) API key for research and news endpoints(optional)
numi services link numinous-signals
Update system packages
sudo apt-get update && sudo apt-get upgrade -y
Install Docker, Docker Compose V2, git, and Python
sudo apt-get install -y docker.io docker-compose-plugin git python3 python3-pip
Start Docker daemon and enable on boot
sudo systemctl start docker && sudo systemctl enable docker
Add user to docker group (avoids needing sudo). Requires logout and login.
sudo usermod -aG docker $USER
Install Bittensor CLI
pip3 install bittensor
Verify installations
docker --version && docker compose version && btcli --version
Clone repository
git clone https://github.com/numinouslabs/numinous.git && cd numinous
Create validator wallet
btcli wallet new-coldkey --wallet validator && btcli wallet new-hotkey --wallet validator --wallet-hotkey default
Register validator on testnet. For mainnet: --netuid 6 --network finney
btcli subnet register --netuid 155 --wallet validator --wallet-hotkey default --network test
Create environment file from template
cp .env.validator.example .env.validator
Edit .env.validator - set WALLET_NAME=validator and WALLET_HOTKEY=default
vim .env.validator
Verify wallet files exist at expected location
ls -la ~/.bittensor/wallets/validator/coldkey && ls -la ~/.bittensor/wallets/validator/hotkeys/default
Pull latest validator container image
docker compose -f docker-compose.validator.yaml --env-file .env.validator pull
Start validator in detached mode
docker compose -f docker-compose.validator.yaml --env-file .env.validator up -d
View validator logs (Ctrl+C to exit)(optional)
docker logs -f numinous_validator