Ideathon Submission — Subnet Design Proposal

1. Introduction: The Vision for Deliberative Collective Intelligence

The NSED (N-Way Self-Evaluating Deliberation) subnet proposes a fundamental shift in how Bittensor produces inference: from parallel sampling of independent miners to structured multi-round deliberation between miners. Today's reasoning subnets compete on single-pass output quality. This subnet competes on something different — the quality of collective reasoning produced by miners whose models critique, refine, and converge on answers together.

<aside> 💡

if you understand fundamental AI/ML well just read this paper of mine, it shows 30%+ increase in reasoning quality above naive majority MoA systems: https://arxiv.org/abs/2601.16863

</aside>

The thesis is benchmark-validated. A heterogeneous swarm of three open-weight models running through NSED's deliberation protocol consistently produces a 30+ percentage-point accuracy lift over naive majority-voting baselines using the identical model set on hard reasoning benchmarks (AIME, LiveCodeBench Hard, GPQA Diamond). This is not single-pass best-of-N selection from frontier models — it is iterative refinement where each round's proposals integrate feedback from peer cross-evaluation, producing converged answers that don't exist in any single agent's initial response. The result holds across model-size tiers, with the relative improvement preserved whether the swarm runs small, mid, or large open-weight models.

This subnet brings that architecture to Bittensor as a task-competitiveness validation arena. Miners do not implement the deliberation protocol themselves. They consume NSED orchestration as an encrypted SaaS service, where the orchestrator coordinates their declared model swarm through structured rounds while preserving end-to-end privacy of intermediate proposals via cryptographic commit-reveal. The competitive surface for miners is the quality of their declared model configuration and prompt tuning, not the orchestration code. Bittensor provides what NSED needs to validate at scale: an adversarial, economically-incentivized population of independent operators stress-testing the protocol on continuously refreshed problem sets.

The flagship deployment configuration for this subnet is NSED Mid 0v1 — a mid-tier swarm composition built on three open-weight models suitable for self-hosted enterprise deployment. A proprietary Linux kernel reasoning model is in active development as the first specialized adapter, demonstrating how NSED's architecture supports domain experts joining the swarm without modifying the base orchestration layer. The subnet is structured to validate both the general deliberation thesis and the specialized-adapter thesis under live competitive conditions.

What makes this a real "proof of intelligence" rather than another best-of-N sampling subnet: NSED produces answers that don't exist in any single agent's initial response. Multi-round refinement synthesizes solutions through structured argument. This is empirically distinguishable from selection-based approaches, and the gap widens on the hardest reasoning problems — exactly where Bittensor needs a differentiated reasoning capability.

2. Incentive & Mechanism Design

Emission and Reward Logic: Deliberation-Quality-Weighted

This subnet does not adopt a strict winner-takes-all model. Instead, emissions flow according to contribution to deliberation quality, measured along three axes that all participants can verify through the orchestrator's cryptographic attestation:

1. Proposal quality — how often a miner's proposed solution wins the final consensus score across deliberation rounds, weighted by problem difficulty.
2. Evaluation accuracy — how well a miner's cross-evaluations of peer proposals correlate with the final converged answer (rank correlation against ground truth where available, against consensus otherwise).
3. Refinement contribution — measurable quality improvement in a miner's proposals between rounds, capturing the miner's ability to integrate peer feedback.

Top-quintile performers across these three axes receive the bulk of emissions, with a long-tail allocation to mid-tier performers who demonstrate consistent quality. Pure winner-takes-all creates an unstable equilibrium for deliberation subnets — a single dominant miner removes the diversity that makes deliberation valuable. The reward curve here intentionally preserves a competitive ecosystem of 8–15 actively participating miners per query, which the NSED protocol requires for the quadratic-voting consensus mechanism to function (the consensus degrades below 5 participants and shows diminishing returns above 15).

Incentive Alignment for Miners and Validators

For Miners: The competitive surface is intentionally narrowed to make competition substantive. Miners do not write orchestration code — they declare swarm configurations: which models they bring, which prompts they tune, which evaluation strategies they configure. The NSED orchestrator (accessed as an encrypted SaaS service) is identical infrastructure for all miners, the same way all miners on existing subnets access identical Bittensor protocol primitives. This eliminates an entire class of low-value optimization (algorithmic micro-gaming, harness exploitation, evaluation-loop tricks) and concentrates competition on what actually matters: model selection, prompt engineering, and budget management.