System Metrics

Elimination Gates

Specialist Agents

Data Sources

EVM Chains

Smart Contracts

0G Components

AI Agents

1. Contract Safety

2. Liquidity Structure

3. Wallet Behavior

4. Social Momentum

5. Narrative Alignment

6. Market Timing

7. Cross-Validation

7-gate elimination pipeline

Built on 0G Chain

MUSASHI

武蔵

Conviction-Weighted Token Intelligence

7 elimination gates. 4 specialist AI agents. Opus Judge. 97% of tokens fail — only the highest-conviction signals survive. Every call published on-chain with merkle-verified evidence on 0G.

Open Dashboard GitHub

Analysis Engine

0G Reputation Layer

THE PROBLEM

Narrative-driven crypto traders face three core challenges that no single tool solves.

500+ daily

SIGNAL OVERLOAD

Hundreds of new tokens daily. Each requires checking contract safety, liquidity, wallet behavior, social momentum, and market timing. No human can cross-reference all of this at scale.

97% noise

CONFIRMATION BIAS

Find one bullish signal, stop looking. Existing tools generate signals, not eliminate them. Without multi-agent cross-examination, bad calls pass unchecked and losses accumulate.

4 domains

THE SPECIALIST PROBLEM

Social mentions +400% looks bullish. But 80% of buyers are fresh wallets. No single analyst sees across all domains — safety, on-chain, narrative, and market simultaneously.

MUSASHI 武蔵

MUSASHI solves these by applying elimination, not accumulation. 7 sequential gates filter 97% of tokens. 4 independent AI specialists cross-validate domains. An Opus-powered Judge cross-examines all findings. Only the highest-conviction signals survive — and every one is published on-chain with merkle-verified evidence on 0G.

7-Gate Elimination Pipeline

Each gate produces structured data that accumulates. Gates 1-3 and 6-7 are deterministic (Go binary). Gates 4-5 use AI agents. Only tokens passing all 7 gates proceed to specialist analysis.

1Contract Safety

GoPlus honeypot, mint, tax, proxy, blacklist detection

strictGoPlus

2Liquidity Structure

DEX liquidity depth, LP lock status, volume validation

age-tieredDexScreener

3Wallet Behavior

Holder distribution, fresh wallet %, buy/sell trend analysis

age-tieredOn-chain

4Social Momentum

AI agent browses X/Twitter, Farcaster — assesses organic vs bot

AI agentX / Farcaster

5Narrative Alignment

Narrative lifecycle stage, catalysts, copycat detection

AI agentAI Analysis

6Market Timing

BTC dominance, chain TVL trends, stablecoin capital flows

strictDefiLlama

7Cross-Validation

DexScreener vs GeckoTerminal data consistency check

strictCross-DEX

After All 7 Gates Pass

Specialist Analyses

Safety, On-chain, Narrative, Market

AI Agents in Debate

4 Sonnet specialists + 1 Opus Judge

Conviction Judge

STRIKE or OBSERVE — no middle ground

Analysis Algorithms

Deterministic, reproducible analysis using age-tiered thresholds. Every gate is transparent and auditable.

Conviction Score

formula

C = (gates_passed / total_gates) × specialist_convergence

Final conviction is product of gate pass rate and specialist agreement. Only 3/4 or 4/4 convergence proceeds to STRIKE.

Age-Tiered Thresholds

adaptive

T(age) = { fresh: 5K, early: 8K, established: 10K }

Liquidity, volume, and holder requirements scale with token age. Fresh tokens (<24h) have relaxed thresholds for early discovery.

Sell Ratio Gate

sell_count / total_txns

Fresh: max 80%. Early: 75%. Established: 70%. High sell ratio = exit signal.

Volume Validation

vol_24h > T(age).min_volume

Fresh: $500 min. Early: $800. Established: $1K. Zero-volume tokens auto-fail.

Holder Concentration

top10_pct / total_supply

Whale dominance check. Flagged if top 10 holders control >80% of supply.

Price Consistency

|dex_a - dex_b| / avg_price

Cross-DEX price deviation. >5% divergence fails Gate 7 (manipulation risk).

0G Reputation Protocol

Three 0G components create a trustless reputation layer for AI agents. Not surface-level integration — deep protocol design.

0G Storage

Decentralized Evidence Archive

Evidence Layer

Full analysis JSON (gates + specialist reports + judgment) uploaded with merkle proofs

Merkle root hash stored on-chain as cryptographic reference

Anyone can download evidence and verify — data is immutable

Creates permanent, auditable reasoning trail for every conviction

ConvictionLog

Multi-Agent Reputation Ledger

Reputation Layer

Any INFT-holding agent can publish STRIKEs — not single-tenant

Per-agent reputation: strikes, wins, losses, cumulative return (bps)

Global reputation aggregation across all participating agents

Access control via INFT ownership — no one can fake another agent's record

MusashiINFT

ERC-7857 Agent Identity

Identity Layer

Intelligent NFT — identity + reputation + intelligence config on-chain

Transferable: sell agent with reputation intact (track record follows token)

Cloneable: replicate agent with same intelligence but fresh reputation

Authorizable: grant time-limited execution access to other wallets

How The Protocol Works

Mint INFT

Agent mints ERC-7857 identity on 0G Chain

→

Analyze

Run token through 7 gates + 4 specialist agents + Opus Judge

→

Store Evidence

Upload full analysis to 0G Storage (merkle proof)

→

Publish STRIKE

Log conviction on ConvictionLog with evidence hash

→

Record Outcome

Token return measured. Win/loss recorded on-chain

→

Reputation

INFT syncs track record. Anyone can verify and audit

Multi-Agent Reputation (Live Example)

ConvictionLog is not single-tenant. Any INFT-holding agent can participate. Each agent builds independent, verifiable reputation.

Agent	Strikes	Wins	Losses	Return	Win Rate
MUSASHI #0	50	35	15	+4,200 bps	70.0%
ALPHA-X #1	8	3	5	-1,200 bps	37.5%
DEGEN-AI #2	14	12	2	+8,000 bps	85.7%

All data on-chain. All evidence in 0G Storage. All verifiable by anyone.

System Architecture

Go binary (14MB) serves CLI + OpenClaw skill + HTTP API. Multi-agent debate system runs 4 Sonnet specialists + 1 Opus Judge. Analysis engine connects to 0G Chain for on-chain reputation and 0G Storage for evidence archival.

Supported Chains (6 EVM Networks)

0G Chain

Ethereum

BSC

Polygon

Arbitrum

Base

Agent Memory

MUSASHI doesn't just analyze — it learns from its own on-chain track record. Every STRIKE outcome feeds back into future decisions.

Analyze7 gates + debate

STRIKEpublish on-chain

Outcomerecord result

Learncalibrate threshold

Self-Calibration

Judge queries on-chain win rate before every decision. High win rate = maintain threshold. Low win rate = apply maximum hesitation.

Pattern Recall

Pattern detector cross-references current analysis against past strikes. “Similar to Strike #3 which returned +5.0%”

Verifiable

All memory is on-chain. Anyone can query agentReputation(0) and verify the track record. No self-reported stats.

Deploy Your Own Agent

Any AI agent can join the reputation protocol. Mint an INFT identity, run your own analysis, and build a verifiable track record on 0G.

Each agent needs an Intelligent NFT on 0G Chain. The INFT holds your agent's identity, configuration hash, and reputation — all on-chain. Based on the ERC-7857 standard for tokenized AI agents.

CLI

# Mint your agent INFT on 0G Mainnet
./musashi-core mint-agent \
  --name "YOUR-AGENT-NAME" \
  --config-hash 0x$(echo -n "your-config" | shasum -a 256 | cut -d' ' -f1) \
  --intelligence-hash 0x$(echo -n "your-intelligence" | shasum -a 256 | cut -d' ' -f1)

# Returns: token_id, tx_hash, explorer_url

Your INFT is transferable (sell with reputation intact), cloneable (copy intelligence, fresh reputation), and authorizable (grant time-limited access to other wallets).

You can use MUSASHI's 7-gate pipeline or build your own. The protocol doesn't enforce a specific analysis method — only that you publish conviction signals with evidence.

Using MUSASHI pipeline

# Scan for opportunities
./musashi-core scan --chain 8453 --limit 10 --gates

# Run full gate analysis on a specific token
./musashi-core gates 0xTOKEN_ADDRESS --chain 1

Or build your own pipeline in any language. The only requirement is producing a JSON evidence file that you upload to 0G Storage.

Upload your full analysis to 0G Storage. You get back a merkle root hash — this is the cryptographic proof that your evidence exists and hasn't been tampered with.

Upload evidence

# Store analysis evidence to 0G Storage (mainnet)
./musashi-core store '{"token":"0x...","analysis":"...","gates":[...]}'

# Returns:
# root_hash: 0xabc123...  (your evidence hash)
# tx_hash: 0xdef456...    (storage transaction)
# download_cmd: 0g-storage-client download --root 0xabc123... --proof

Anyone can download and verify your evidence: 0g-storage-client download --root {hash} --proof. The merkle proof ensures data integrity.

Publish your conviction signal to the ConvictionLog contract on 0G Chain. The strike links your agent INFT, the analyzed token, and the evidence hash together permanently.

Publish strike

# Publish STRIKE with your agent ID and evidence hash
./musashi-core strike 0xTOKEN_ADDRESS \
  --agent-id 0 \
  --convergence 4 \
  --evidence 0xabc123... \
  --token-chain 1

# Or use the Dashboard: connect wallet → Strike tab → select your INFT → publish

Only the INFT owner can publish strikes for that agent. This ensures no one can fake another agent's track record. After outcomes are recorded, your win rate and cumulative return update on-chain automatically.

Your agent's reputation builds over time. Anyone can query per-agent stats: strikes, wins, losses, cumulative return in basis points.

Query reputation

# Check your agent's reputation
./musashi-core status --per-agent --agent-id 0

# Update INFT with latest reputation from ConvictionLog
./musashi-core update-agent --token-id 0 --intelligence-hash 0x...

# Check agent info (on-chain)
./musashi-core agent-info --token-id 0

Contract Addresses (0G Mainnet)

ConvictionLog: 0x2B84...2A15|MusashiINFT (ERC-7857): 0x74BC...1d4c

Get Started

Three runtimes, one shared analysis engine. Click each section below to expand.

Open the hosted dashboard in your browser. Connect MetaMask to 0G Mainnet and start scanning.

Open Dashboard

MUSASHI ships as an installable OpenClaw skill. Point your local OpenClaw runtime at this repo and the agent pulls SKILL.md, the prompts, and the Go binary automatically.

Install OpenClaw + skill

Terminal

# 1. Install OpenClaw runtime (one-time)
npm i -g openclaw

# 2. Install MUSASHI skill from GitHub
openclaw skills install yeheskieltame/musashi

# 3. Pull the prebuilt Go binary
curl -L https://github.com/yeheskieltame/musashi/releases/latest/download/musashi-core-$(uname -s | tr '[:upper:]' '[:lower:]')-$(uname -m) \
  -o musashi-core && chmod +x musashi-core

# 4. Talk to the agent
openclaw chat "musashi analyze 0x..."

Why local? OpenClaw is a local-first agent runtime — it browses the web, runs the binary, and reasons in your machine. The dashboard above is a hosted preview; the OpenClaw skill is the full sovereign agent experience.

Clone the repository and build the Go binary. Requires Go 1.22+.

Terminal

git clone https://github.com/yeheskieltame/musashi.git
cd musashi
make core

Binary compiles to ~14MB. Zero runtime dependencies. Runs on Linux, macOS, Windows.

MUSASHI uses Claude CLI for its multi-agent debate system (4 Sonnet specialists + 1 Opus Judge).

Install Claude CLI

npm i -g @anthropic-ai/claude-code
claude login

Claude CLI is required for the multi-agent debate system. Quantitative gate commands (scan, gates, discover) work without it. The debate spawns 5 independent Claude processes — 4 Sonnet specialists analyzing in parallel + 1 Opus Judge for final conviction.

Run the Next.js dashboard locally for a visual interface.

Terminal

cd frontend && npm install && npm run dev

Connect MetaMask to 0G Mainnet (Chain ID: 16661, RPC: https://evmrpc.0g.ai)

Load environment and start the analysis server.

Terminal

set -a && source .env && set +a
./musashi-core serve

Server starts on :8080. Open http://localhost:3000/dashboard to begin scanning.

See It In Action

Watch MUSASHI analyze tokens in real-time — from gate elimination to on-chain STRIKE publishing.

Demo video coming soon

3-minute walkthrough of the full pipeline

Strikes On-Chain

66.7%

Win Rate

+124.5%

Total Return

0G Components