1 of 10

REASONFORGE

The Decentralized Marketplace for Verifiable Intelligence

Where reasoning itself is the commodity.

Bittensor Subnet Proposal | Subnet Ideathon — Round I

RS BEJINI| February 2026

2 of 10

The Reasoning Crisis

Frontier AI models fail at the one thing that matters most: trustworthy reasoning.

60%

End-to-end reliability of

a 10-step chain at 95%/step

$0

Revenue going to improve

reasoning competitively

0

Formal verification

guarantees from top models

  • Hallucination compounds exponentially with chain length
  • Zero verifiability — users must blindly trust the black box
  • Centralized bottleneck — 3 companies control all frontier reasoning
  • No competitive incentive for continuous improvement

3 of 10

ReasonForge: Proof-of-Intelligence Protocol

Miners compete to produce verifiable multi-step reasoning. Validators enforce quality. TAO flows to intelligence.

Process Supervision

Score every reasoning step,

not just final answers

Formal Verification

Lean 4 / Coq proof checkers

for mathematical tasks

Adversarial Debate

Consensus protocol for

subjective evaluation

6 Domains

Math, Code, Science, Strategy,

Causal, Ethics

Anti-Collusion

Blind eval, trap problems,

plagiarism detection

Cross-Subnet

SubnetsAPI composability—

other subnets query RF

4 of 10

System Architecture

4-layer design: Task Sources → Miner Network → Validator Network → Reward Distribution

5 of 10

Miner Node Architecture

Any LLM, agent framework, or multi-agent swarm can participate as a miner

Input: Standardized JSON task with domain, difficulty, timeout Output: Structured reasoning trace + formal proof artifacts

6 of 10

Mechanism Design & Task Lifecycle

CMS(m,t) = 0.40·Q + 0.30·A + 0.15·N + 0.15·Eff

R(m) = E_miner × [S_epoch(m) × (1+PEB(m))] / Σ_j [S_epoch(j) × (1+PEB(j))]

Anti-adversarial: 15% trap problems | Blind evaluation | Plagiarism (cos>0.95) | Quadratic slashing

7 of 10

Validator Scoring Pipeline

Hybrid evaluation: 60% automated verification + 40% stake-weighted consensus

Honest (VAS>0.90): 1.5× reward | Lazy (VAS≈0.50): 0.83× + slashing | Malicious (VAS<0.30): Net loss — attack is irrational

8 of 10

Token Economics & Emission Flow

Self-sustaining flywheel: quality reasoning → TAO rewards → more miners → better intelligence

9 of 10

Market & Competitive Landscape

TAM

$47B

AI Reasoning

Services (2028)

SAM

$8.2B

Verifiable /

Auditable AI

SOM

$120M

Year 3 Target

1.5% of SAM

Key Competitors

Solution

Gap

OpenAI o1

Black box, no verify

DeepSeek-R1

No incentive loop

Ridges (SN)

Code-only scope

Perplexity

Shallow reasoning

Why Bittensor Is the Right Substrate

Built-in Incentives

TAO emissions create self-sustaining

competitive marketplace from day 1

Yuma Consensus

Proven validator mechanism as

foundation for hybrid scoring

Cross-Subnet API

Other subnets query ReasonForge

as intelligence primitive

Permissionless

Anyone with reasoning capability

can mine — zero barriers

10 of 10

Roadmap & Call to Action

ReasonForge delivers genuine proof-of-intelligence with rigorous incentives, crystal-clear roles, and a real path to adoption.

  • Full 25-page proposal with mathematical formulas & architectural diagrams
  • 6 professional system diagrams embedded throughout
  • Concrete roadmap: testnet Q1 → mainnet Q2 → enterprise API Q3 → scale Q4
  • Looking for: collaborators, early miners, validator partners, enterprise pilots

Ravi Shankar Bejini| @Rbejini | Let’s build verifiable intelligence together.