MicroFish/simulator

MiroFish × BMNR Simulator

Agent-based stock simulator for BitMine (BMNR) powered by LLM swarm intelligence.

A fork of MiroFish adapted for financial market simulation. Instead of predicting outcomes with formulas, 21 AI agents with distinct personas reason collectively about BMNR's future — and predictions emerge from their interactions.

Not financial advice. This is an educational simulation tool. Not affiliated with BitMine Immersion Technologies, Tom Lee, ARK Invest, or any entity mentioned. Agent personas are fictional archetypes.

---

What It Does

Feature	Description
Live Market Data	Fetches BMNR price, ETH price, mNAV, holdings from bitminetracker.io on every load
21 AI Agents	ARK/Cathie, Tom Lee, short sellers, Reddit bulls, WSB degens, arb bots — each with detailed personas
LLM Reasoning	Each quarter, agents reason via LLM about the market state — not programmed formulas
25 Market Stimuli	Toggle events (ETH rally, SEC crackdown, MAVAN launch) with adjustable intensity
Flywheel Math	Real ATM issuance mechanics: NAV calc → breakeven check → share issuance → ETH/share accretion
3 Scenarios	Bear / Base / Bull with overlaid comparison charts
Multi-LLM	Pluggable providers: Anthropic, OpenAI, Google Gemini, OpenRouter, Ollama (local), or custom
Formula Fallback	Instant results via network-propagation math when LLM is unavailable

---

Quick Start

Option 1: Claude.ai Artifact (zero setup)

Copy the contents of src/App.jsx into a Claude.ai artifact. It runs immediately — the Anthropic API is available inside artifacts with no API key.

Option 2: Self-Host

git clone https://github.com/mikema-rgb/BMNR-Mirofish.git
cd BMNR-Mirofish
cp .env.example .env        # Edit with your API key
npm install
npm run dev                  # Opens at http://localhost:3000

Option 3: Any LLM

Edit .env to use your preferred provider:

# OpenAI
VITE_LLM_PROVIDER=openai
VITE_OPENAI_API_KEY=sk-...

# Google Gemini
VITE_LLM_PROVIDER=gemini
VITE_GEMINI_API_KEY=AIza...

# Ollama (local, free, no key)
VITE_LLM_PROVIDER=ollama
VITE_OLLAMA_BASE_URL=http://localhost:11434

# OpenRouter (100+ models)
VITE_LLM_PROVIDER=openrouter
VITE_OPENROUTER_API_KEY=sk-or-...

# Any OpenAI-compatible endpoint
VITE_LLM_PROVIDER=custom
VITE_CUSTOM_BASE_URL=https://your-endpoint.com
VITE_CUSTOM_API_KEY=your-key

---

Architecture

BMNR-Mirofish/
├── src/
│   ├── App.jsx                 # Main component (monolith — runs standalone)
│   ├── main.jsx                # React entry point
│   ├── config/
│   │   └── index.js            # Simulation config, scenario metadata, fallback data
│   ├── data/
│   │   ├── agents.js           # 21 agent personas (the core of MiroFish)
│   │   └── stimuli.js          # 25 market event definitions
│   ├── engine/
│   │   └── prompts.js          # LLM prompt builder (fork this for other stocks)
│   └── providers/
│       └── index.js            # LLM provider abstraction (Anthropic, OpenAI, etc.)
├── .env.example                # Environment config template
├── package.json
├── vite.config.js
└── README.md

How MiroFish Works (vs. Traditional Models)

	Traditional Monte Carlo	Formula Agent Model	MiroFish LLM Agents
How price moves	price × exp(drift + vol × dW)	Sentiment formulas drive returns	LLM reasons about what agents would do
Agent behavior	None — just noise terms	Programmed bias + susceptibility	Natural language persona → LLM reasoning
Emergent effects	None	Limited network propagation	Herding, contrarianism, cascade failures
Stimulus response	Hardcoded impact numbers	Type-weighted sensitivity	LLM interprets event for each persona
Surprise factor	None — deterministic given seed	Low — formulas are predictable	High — LLM reasoning can produce unexpected cascades

The Simulation Loop

Each quarter:

1. Agent Selection — 8 core agents + 3 rotating guests are selected
2. Prompt Construction — Market state, stimuli, prior quarter narrative, and agent personas are assembled into a prompt
3. LLM Call — The prompt is sent to the configured LLM provider
4. Response Parsing — JSON output is extracted with markdown fence stripping and brace-counting fallback
5. Market Update — LLM-derived ETH price change and mNAV change are applied (clamped to ±50% per quarter)
6. Flywheel Engine — If mNAV > breakeven, ATM shares are issued, ETH is purchased, ETH/share updates
7. State Propagation — The quarter narrative feeds into the next round as context

---

How to Extend

Add a New Agent

Edit src/data/agents.js:

{
  id: "your_agent",
  name: "Your Agent Name",
  type: "Institutional",      // or Retail, Quant, Macro, Analyst, etc.
  icon: "🎯",
  bias: 0.2,                  // -1 (bearish) to 1 (bullish) for formula fallback
  influence: 0.5,             // 0-1: how much they sway others
  susceptibility: 0.3,        // 0-1: how much they follow the herd
  memory: 0.6,                // 0-1: how much past rounds carry forward
  persona: "You are... [detailed natural language persona]. You believe... You react to..."
}

The persona is the most important field — write it like you're briefing a method actor.

Add a New Stimulus

Edit src/data/stimuli.js:

{
  id: "your_event",
  name: "Your Event Name",
  cat: "ETH",                  // Category for filtering
  icon: "⚡",
  impact: 0.5,                // -1 to 1 for formula fallback
  desc: "Short description"
}

Add a New LLM Provider

Edit src/providers/index.js:

export const YourProvider = {
  id: "your_provider",
  name: "Your Provider",
  models: ["model-1", "model-2"],
  defaultModel: "model-1",
  requiresKey: true,

  async call(messages, { model, maxTokens, apiKey } = {}) {
    // Make your API call here
    // Return { text: string, raw: any }
    const res = await fetch("https://your-api.com/v1/chat", { ... });
    const data = await res.json();
    return { text: data.output, raw: data };
  }
};

// Register it:
export const PROVIDERS = {
  ...existingProviders,
  your_provider: YourProvider,
};

Fork for Another Stock

The simulator is designed to be adapted. To fork for a different stock:

1. Replace agent personas in src/data/agents.js with archetypes relevant to your stock
2. Replace stimuli in src/data/stimuli.js with events that affect your stock
3. Update src/config/index.js with fallback market data for your stock
4. Modify src/engine/prompts.js to reference your stock's specific dynamics (the flywheel mechanics are BMNR-specific — replace with whatever drives your stock's value)
5. Update the live data fetcher in src/App.jsx to pull from relevant data sources

---

Acknowledgments

• MiroFish — the original swarm intelligence engine by Shanda Group. This project adapts their agent-based simulation concept for financial markets.
• OASIS — the underlying social simulation platform that powers MiroFish.
• Wedge — the BMNR flywheel model and design system.
• bitminetracker.io — live BMNR market data.

---

Contributing

PRs welcome. The most impactful contributions:

• Better agent personas — if you can write a more realistic persona for any of the 21 agents, that directly improves simulation quality
• New stimuli — market events we haven't thought of
• New LLM providers — expand the provider abstraction
• Prompt engineering — improvements to src/engine/prompts.js that produce more realistic agent reasoning
• UI improvements — better charts, mobile responsiveness, dark mode
• Other stocks — fork and adapt for MSTR, COIN, or any other stock with similar dynamics

---

License

MIT — same as MiroFish.

---

WEDGE × MIROFISH × BMNR
Educational only. Not financial advice.