Skillforest

Skillforest Subnet (GEP-Oriented Agent Capability Market)

One-line Description

A Bittensor subnet where miners earn by shipping measurable improvements to reusable agent capabilities ("genes/capsules"), validated through reproducible benchmark execution.

What the Project Does

Skillforest Subnet turns agent improvement into a competitive, verifiable game.

Instead of rewarding miners for serving a static endpoint, the subnet rewards miners for publishing better agent skill packages that perform better on benchmark tasks. Validators execute submitted artifacts in controlled environments and score them using a transparent composite score:

outcome quality / correctness

trajectory quality (how the agent solved it)

safety / policy compliance

efficiency (token and latency normalized)

reliability (timeouts/errors)

This creates a public, replayable record of which miners are actually improving useful agent behavior.

Core Idea (Why It Matters)

Most "AI skill" ecosystems are hard to trust:

demos are selective

outputs are hard to reproduce

capability claims are often unverifiable

Skillforest Subnet addresses this by treating agent capabilities as evolving units (genes/capsules) that must pass standardized evaluation. The subnet is designed to reward validated capability improvement, which is closer to a practical "proof of effort" and "proof of capability" than simple API uptime or self-reported metrics.

How It Works (High-Level)

A miner packages an agent capability improvement as a versioned artifact/container.

The miner publishes the artifact reference for validator discovery.

Validators assign benchmark task batches and execute the artifact.

Validators capture outputs, telemetry, and trajectory events.

Validators compute deterministic scores and aggregate miner performance.

Validators set on-chain weights based on normalized, smoothed performance.

Miners receive rewards for demonstrated capability gains.

Why Bittensor

This use case benefits from Bittensor because it requires:

continuous open competition

transparent validator scoring logic

reward allocation based on measurable performance

a long-term transition from subsidy-driven discovery to market-driven demand

Skillforest is designed to use chain emissions initially to bootstrap participation and popularity, then gradually shift miner rewards toward revenue generated by premium skills and paid custom-skill improvement subscriptions.

GTM / Sustainability Vision

Stage 1: Use miner emissions to attract builders and establish benchmark credibility.

Stage 2: Introduce premium skills and paid custom capability improvement subscriptions.

Stage 3: Shift miner rewards from emissions toward revenue share.

End State: 70% of revenue pays miners, 30% supports operations/buybacks, and miner emissions are fully burned/neutralized so incentives are directly market-driven.

Current Status

This repository is a subnet scaffold with:

validator and miner templates

scoring logic

benchmark and evaluation specs

mechanism and GTM draft

Production features still to implement include:

real miner discovery and submission retrieval

production sandbox/container execution in validator path

hotkey-to-UID mapping and safe weight setting

persistent EMA scoring and audit/challenge flows