MVTRX

<div align="center">

# **MVTRX**: Bittensor SN79<!-- omit in toc -->
[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
---


**MVTRX** operates as a [Bittensor](https://bittensor.com) subnet at netuid 79 for decentralised market research and AI model training. It comprises three integrated components:

- **τaos**: agent-based simulation of automated trading strategies in intelligent markets, incentivising risk-managed, high-quality market participation
- **GenTRX**: distributed training of a shared order-book generative model, built on top of τaos simulation data (and future exchange data)
- **MVTRX Exchange**: off-chain limit order book exchange for Bittensor alpha tokens at [mvtrx.exchange](https://mvtrx.exchange), running the same C++ matching engine as the simulation. This release provides the tools and documentation to build and test agents against a **localnet** exchange; the mainnet exchange mechanism is a later release (see [Exchange](#exchange))

[![Website](https://img.shields.io/badge/website-black?logo=googlechrome
)](https://mvtrx.fi)
[![Exchange UI](https://img.shields.io/badge/exchange-black?logo=googlechrome
)](https://mvtrx.exchange)
[![Grafana](https://img.shields.io/badge/grafana-white?logo=grafana
)](https://taos.simulate.trading)
[![Simulation Terminal](https://img.shields.io/badge/dashboard-white?logo=grafana
)](https://mvtrx.simulate.trading)
[![Discord](https://img.shields.io/badge/discord-black?logo=discord
)](https://discord.com/channels/799672011265015819/1353733356470276096)
[![τaos Whitepaper](https://img.shields.io/badge/whitepaper-white?logo=proton
)](https://simulate.trading/taos-im-paper)

---
**_taos_ (/ˈtɑos/)** : To make things out of metal by heating it until it is soft and then bending and hitting it with a hammer to create the right shape.

---
### Table of Contents
</div>

1. [Incentive Mechanism](#mechanism)
    - [Owner Role](#mechanism-owner)
    - [Validator Role](#mechanism-validator)
    - [Miner Role](#mechanism-miner)
2. [Technical Operation](#technical)
    - [Simulator](#technical-simulator)
    - [Validator](#technical-validator)
    - [Miner](#technical-miner)
    - [GenTRX](#gentrx-technical)
3. [Requirements](#requirements)
    - [Validator](#requirements-validator)
    - [Miner](#requirements-miner)
4. [Agents](#agents)
5. [Exchange](#exchange)
6. [Install](#install)
    - [Validator](#install-validator)
    - [Miner](#install-miner)
    - [Docker](#install-docker)
7. [Run](#run)
    - [Registration](#run-registration)
    - [Validator](#run-validator)
    - [Miner](#run-miner)
8. [GenTRX Distributed Training](#gentrx)
---

<div style="page-break-after: always;"></div>

## Incentive Mechanism <span id="mechanism"><span>
The incentive mechanism described here covers the **τaos simulation** and **GenTRX training** components, which are the two that carry emissions on mainnet today. The **Exchange** runs as a separate subnet mechanism whose scoring is additional to these; it is not yet enabled on mainnet, and its reward dimensions will be published before it is. See [Exchange](#exchange) for what is available now.

For the τaos component: the mechanism is designed to promote intelligent, risk-managed trading logic to be applied by agents, in order that we are able to produce valid and valuable datasets mimicing the properties of a variety of different real-world asset classes and market conditions. See the [whitepaper](https://simulate.trading/taos-im-paper) for a detailed exploration of the background, goals and scope.

**Two reward pools.** Miner rewards are split across two incentive pools that run in parallel:

- **Trading pool** (~95% of rewards by default): scored on kappa and PnL from simulation trading. All miners registered on the simulation mechanism participate. Exchange trading is scored under its own mechanism with its own weights, not blended into this pool.
- **GenTRX training pool** (~5% by default, set by `--scoring.gentrx.simulation_share` on the validator): scored on gradient quality, assessed each round against held-out order-book data. Scales with active participation; unused training rewards return to the trading pool. Opt-in for both validators and miners; zero impact on trading rewards when not in use.

### Owner Role <span id="mechanism-owner"><span>
The subnet owners are tasked with ensuring fair, equitable and correct operation of the subnet mechanisms (as in all other subnets), while also being responsible for the design, refinement, tuning and publishing of the simulation parameters and logic.  This involves consistent monitoring, testing and development to expand the capabilities of the simulator and determine parameters which result in the most useful possible outputs being generated through the subnet's operation.  The owner must also ensure that the metrics utilized in determining miner rewards are chosen such that miners are incentivized to act fairly and in such a way that outputs are of optimal value in research, trading strategy development, market surveillance and other applications.

For the GenTRX component, the owners also operate the canonical **aggregator** (uid 0): a gradient server that evaluates all miner and sibling-validator proposals each round, applies the best-scoring delta to the shared model, and publishes the new checkpoint on-chain for all participants to download.

### Validator Role <span id="mechanism-validator"><span>
Validators in the subnet are responsible primarily for maintaining the state of the simulation, and rewarding agents (miners) which achieve the best results over all realizations of the simulated market.  They deploy two components:
- The C++ simulator, which handles all the computation necessary to simulate asset markets
- The Python validator, which receives state updates from the simulator, forwards these to miners, submits instructions received in response back to the simulator, and calculates miner sco