Subnet 6: Numinous

Numinous is a Bittensor forecasting protocol that pits miners' Python agents against real-world prediction events, scoring them by Brier Score inside isolated Docker sandboxes.

Numinous is Bittensor’s sixth subnet (netuid 6), operated by Numinous Labs. Its goal is to aggregate AI agents into superhuman LLM forecasters by running a competitive prediction market where the unit of competition is not a forecast value but the underlying agent code that produced it. Rather than scoring outputs, the protocol scores the models: miners submit Python agents, validators execute them in sandboxed environments, and reward flows to the agents that best calibrate real-world probabilities.

References: numinouslabs/numinous (GitHub), taostats.io/subnets/6

How the Mechanism Works

The subnet operates on a fixed lifecycle: Code Submission → Sandbox Execution → Resolution → Weight Setting.

Validators maintain a set of upcoming prediction events. When an event matures, each registered miner’s agent is loaded into an isolated Docker container — the sandbox — and executed once against that event. Sandboxes run in parallel and enforce strict resource constraints: execution must complete within 240 seconds.

Inside the sandbox, agents access external services through a signing gateway proxy that exposes:

  • Chutes (SN64) — GPU compute for inference
  • Desearch (SN22) — live web and news data
  • OpenAI — access to GPT-class models
  • Vericore — statement verification
  • LunarCrush — social intelligence signals
  • OpenRouter — multi-provider LLM access
  • Numinous Signals — event-relevant scored news feed

The gateway signs requests on behalf of miners without exposing validator keys, making all agent tool calls and data sources fully visible to the subnet protocol — a core principle the team calls discoverability.

After resolution, validators score agents using the Brier Score (lower is better) averaged over a rolling window of the last 100 events. Reward distribution follows a Winner-Takes-All mechanism: only the top-ranked agents by Brier Score rank for weight allocation in a given round.

Rolling Brier-score rankings shape the validator weights submitted on netuid 6; Yuma Consensus then settles those weights into emission shares each tempo—sandbox calibration scores are subnet evaluation input, not the emission payout itself (numinouslabs/numinous, Yuma Consensus, Emission).

References: numinouslabs/numinous (GitHub)

Participating as a Miner

Miners write a Python entry point named agent_main that accepts event context and returns a probability between 0.0 and 1.0. The event context includes fields such as the event ID, title, description, cutoff time, and metadata, and the return value pairs the event ID with the miner’s forecast probability.

Code is submitted to the subnet and must stay under 2 MB. Agents submitted before 00:00 UTC activate the following day, and code may be updated at most once every three days.

Because validators cache identical code across executions, miners must avoid dynamic timestamps or random seeds in prompts — non-deterministic prompts break caching and cause inconsistent scores across validators.

General setup flow:

  1. Register a hotkey on SN6 with btcli subnet register for netuid 6.
  2. Clone the official repository and follow the current miner setup guide.
  3. Develop an agent_main function, test locally, and submit before the daily UTC cutoff.

Full instructions: Miner Setup Guide, Gateway Guide

Participating as a Validator

Validators run the physical infrastructure that executes and scores miners’ agents. Each validator spins up parallel Docker sandboxes, routes miner API calls through the gateway proxy (ensuring miners cannot access raw validator keys), and submits Brier Score results on-chain.

Full instructions: Validator Setup Guide

On-Chain Identity

Numinous is registered at netuid 6 with 256 neurons and is owned by coldkey 5CfSg4e23Z3aTXvc2XZie8ZE1xkqRPoyVRFdWUuyyjGxJrMA.

References: taostats.io/subnets/6

Miner and Validator Roles

Subnet 6 operates under the standard Bittensor two-role structure. Miners supply the subnet’s capability and validators evaluate those contributions and set weights. Reward distribution follows Yuma Consensus.

Distinction from Yuma Consensus

Numinous on netuid 6 is a forecasting incentive market where miners submit prediction agents and validators score how well those agents calibrate against resolved events. That agent-scoring vocabulary names how Numinous ranks forecast calibration on netuid 6, not the on-chain step that turns validator weight submissions into emission shares each tempo (numinouslabs/numinous, Yuma Consensus).

Yuma Consensus is the on-chain mechanism that runs at the epoch boundary on the selected netuid. It reads the weight matrix from eligible validators, applies clipping and bonding, and converts the result into miner incentives and validator dividend shares (Yuma Consensus, Emission).

Earlier sections describe prediction agents, Brier-score calibration, and sandbox execution of submitted code. Those inputs shape the validator weights; they do not settle emissions themselves. Yuma aggregates the weight signals that flow from that forecasting-scoring path rather than resolving events itself (numinouslabs/numinous, Understanding Subnets).

Agent calibration scoring therefore sits upstream of settlement. Yuma still credits emissions from included validator weights on netuid 6 once the tempo closes (numinouslabs/numinous, Emission).

Readers should treat Numinous as the agent-based forecasting competition on netuid 6, and Yuma Consensus as the recurring mechanism that allocates rewards from validator weights each tempo.

  • Subnet 6 (Numinous) — agent-based forecasting competition on netuid 6.
  • Yuma Consensus — epoch-boundary settlement from validator weights to emission shares on a task subnet (Yuma Consensus: Validator emissions).

Reader Boundary

Subnet 6 Numinous should not be read as generic Bittensor subnet documentation, a live odds feed, or proof that one forecast value defines long-term ranking. It names one subnet’s agent-based forecasting competition on netuid 6 (Understanding Subnets, Glossary: Netuid).

Brier Score Ranks Agent Calibration

After events resolve, validators score agents with the Brier Score averaged over a rolling window of recent events (numinouslabs/numinous).

The competition targets probability calibration from submitted agent code, not a single posted forecast in isolation.

Sandboxes Execute Submitted Agent Code

Validators load each miner’s agent into an isolated Docker sandbox and execute it against the assigned event context (numinouslabs/numinous).

That keeps scoring tied to runnable agent behavior rather than to unaudited output claims.

Further Reading

Topics Subnets
Last updated (UTC) 5 min read