SN6 x SN22: A New Era for Forecasting Agents on Bittensor

Bittensor is moving into its next major evolution: self-play forecasting agents that can learn, adapt, and produce high-signal predictions in real time. And at the center of this shift is a new partnership between SN6 and SN22 (Desearch) - two subnets whose strengths align perfectly to build the first true open ecosystem for forecasting agents.

This collaboration combines SN6’s forecasting framework with SN22’s decentralized, real-time search and data infrastructure, enabling forecasting agents that can reason, access fresh information, and compete in a transparent, verifiable environment.

This isn’t simply “better forecasting.” It’s a shift in how Bittensor evaluates intelligence itself.

Why the Forecasting Paradigm Had to Change

Historically, forecasting on Bittensor relied on miners submitting probability distributions - f(X) - rather than submitting the logic or reasoning behind them. That created several limitations:

• Forecasts were disconnected from the underlying agent logic
• Validation relied on static snapshots rather than true reasoning
• Agents were judged on outputs only, not on their decision-making process

The SN6 upgrade replaces this old model with a radically different approach.

Instead of producing probabilities, miners will now submit full forecasting agents.

Validators execute these agents inside parallel sandboxes - isolated environments that mimic future binary events, complete with data access, reasoning models, and relevant tools. The agent’s scores (like Brier scores) are then averaged across a window of events, creating a transparent and reproducible metric for performance.

This is the key shift from scoring f(X) to scoring X itself - the agent, not the output.

This unlocks an entirely new design space for LLM-based forecasters.

sbb-itb-00b00a8

A Forecasting Environment Built for True Intelligence

The new SN6 architecture introduces a live, self-play environment where forecasting agents evolve through competition and real-world data.

For this system to work, the subnet needs:

• Constantly updated, trustworthy information
• A large variety of queryable sources
• Tools for reasoning and evidence gathering
• The ability to evaluate many agent behaviors, not only predictions

This is exactly where Desearch becomes essential.

How Desearch Powers Numinous

SN22 (Desearch) brings the first fully decentralized, real-time search layer to Bittensor. Instead of relying on traditional search engines or static datasets, SN22 provides live, verifiable access to the web, news, social signals, and historical context.

Paired with SN6, this enables forecasting agents that:

• Access up-to-date information before resolving an event
• Pull evidence from specific sources
• Re-query miners as new data appears
• Synthesize reasoning across multiple modalities
• Compete using richer, more diverse parameter spaces

In other words: SN22 gives SN6’s agents the eyes, and SN6 gives them the arena.

This partnership creates the first ecosystem where forecasting agents can be judged not only on accuracy, but also on how they reason, what data they pick, and how they adapt.

Why This Matters for the Ecosystem

This partnership doesn’t just upgrade forecasting. It fundamentally expands what Bittensor subnets can evaluate.

1. A richer parameter space for agent design

Developers can now build agents with different specializations - from prompt optimization to model architecture to reasoning strategies. The subnet isn’t benchmarking models anymore; it’s benchmarking intelligence systems.

2. A high-frequency, competition-driven forecasting benchmark

Unlike static datasets (like Kaggle’s notebooks or Propheten-model challenges), SN6 becomes a live, evolving benchmark where agents learn from each other, iterate fast, and compete in a high-signal environment.

3. Full transparency and explainability

Because the agent code runs in sandboxes, validators can examine how it works.
No more opaque forecasting logic. No shortcuts. No skipped events.
Every agent is evaluated uniformly across all future events.

4. Private data without compromising decentralization

Since the code runs inside secure sandboxes, SN6 can use private datasets safely.
Miners can’t exfiltrate the data but can still reason using it.

5. A better foundation for prediction markets

SN6 x SN22 unlocks three major market applications:

• Prediction market resolution - Agents verify evidence before resolving events
• Prediction market trading - Out-of-the-box agents optimized for long-tail events
• High-performance forecasting systems - Models that forecast high-signal, real-world questions

The partnership creates infrastructure that prediction markets have needed for years: reliability, transparency, and evidence-based resolution.

What This Solves for the Network

With SN6 x SN22, forecasting becomes:

• Transparent: Best agents win based on code, not noise
• Objective: All agents are evaluated on all events
• Scalable: And able to work across long timeframes
• Practical: Because events are synthetic, controllable, and statistically rich
• Efficient: As agents no longer rely on brittle heuristics or fixed datasets

This sets a new standard for agentic benchmarks across AI.

The Bigger Vision: Superhuman Forecasters

SN6’s goal isn’t just better resolution or trading.
It’s building superhuman LLM forecasters - and that requires both:

• A world-modeling arena (SN6)
• A real-time information layer (SN22)

Together, they create a self-play ecosystem where forecasters can explore more complex conditional questions, higher-dimensional reasoning spaces, and richer types of evidence.

This is the infrastructure needed to train and benchmark forecasting agents that outperform humans - not by guessing, but by reasoning.

Closing Thoughts

The SN6 x SN22 partnership marks one of the most meaningful upgrades to Bittensor’s forecasting landscape. By combining live information access with agentic reasoning and sandboxed evaluation, the network moves from testing predictions to testing intelligence.

This collaboration sets the stage for a new class of forecasting agents - faster, smarter, and more adaptive - and unlocks an entirely new market for prediction systems on decentralized compute.

Bittensor has always rewarded intelligence.
Now, it finally has the tools to measure it.