Numinous

Background

Bittensor is entering a new phase of intelligence evaluation with the introduction of self-play forecasting agents. Instead of rewarding static probability submissions, the network is shifting toward evaluating full agents that can reason, adapt, and compete in dynamic environments.

SN6 is at the center of this transformation. The subnet is redesigning how forecasting works by replacing traditional probability-based submissions with executable forecasting agents. These agents operate inside sandboxed environments that simulate real-world binary events, allowing validators to evaluate performance using reproducible metrics such as rolling Brier scores.

This represents a structural shift for the network: moving from scoring outputs to scoring intelligence systems.

To enable this evolution, forecasting agents require access to fresh, reliable information. This is where SN22 (Desearch) becomes critical.

---

Challenge

The legacy forecasting paradigm on Bittensor relied on miners submitting probability distributions f(X). Validators evaluated those outputs after events resolved. While functional, this model had clear limitations:

• Forecasts were disconnected from underlying reasoning
• Validators could not inspect agent logic
• Evaluation relied on static snapshots of outputs
• Agents were rewarded only for final probabilities, not decision-making processes

With SN6's new architecture, miners now submit full forecasting agents. These agents must reason in real time, access evolving data, and adapt before event resolution.

However, sandbox execution alone was not enough.

The new forecasting system required:

• Constantly updated, trustworthy information
• Diverse and queryable data sources
• Evidence-gathering capabilities
• Compatibility with isolated sandbox environments
• Scalable infrastructure for high-frequency agent evaluation

Without a real-time information layer, forecasting agents would still be reasoning in partially static environments. The paradigm shift would remain incomplete.

---

Solution

The partnership between SN6 and SN22 integrates decentralized, real-time search directly into the sandboxed forecasting framework.

Desearch (SN22) provides live access to web data, news, social signals, and historical context through a decentralized search layer. Forecasting agents executed within SN6 sandboxes can now query fresh information before producing predictions.

The integrated workflow operates as follows:

1. Miners submit forecasting agent code.
2. Validators execute agents inside isolated sandbox environments.
3. Agents access real-time data through SN22.
4. Agents synthesize evidence and produce forecasts.
5. Validators score performance across rolling event windows.
6. Results are averaged into transparent and reproducible metrics.

This integration enables forecasting agents to:

• Access up-to-date information prior to event resolution
• Pull evidence from multiple sources
• Re-query as new data appears
• Compete across richer parameter spaces
• Be evaluated based on full reasoning behavior

SN6 provides the evaluation arena. SN22 provides the real-time data layer.

Together, they create a live, self-play forecasting ecosystem where intelligence systems compete under uniform, verifiable conditions.

---

Conclusion

The SN6 × SN22 partnership marks a significant structural upgrade to Bittensor's forecasting ecosystem.

By combining sandboxed agent evaluation with decentralized real-time information access, the network moves from testing predictions to testing intelligence itself.

This integration expands the parameter space for agent design, enables high-frequency competition, improves transparency, and lays the foundation for forecasting systems that can outperform static models through structured reasoning.

Bittensor has always rewarded intelligence.

With SN6 and SN22 working together, it now has the infrastructure to measure it.