⚡ ArbOS

Automated Prediction Market Arbitrage Engine

The market is wrong. Profit from it.

ArbOS is a low-latency, logic-driven trading system that automatically detects and exploits mathematical inconsistencies in decentralized prediction markets on Polymarket (Polygon PoS).

What is ArbOS?

Prediction markets like Polymarket let people trade the probability of real-world events. Each market has its own isolated order book, and nobody enforces consistency across related markets. That structural gap is the alpha.

Unlike traditional crypto arbitrage (buy low on Exchange A, sell high on Exchange B), ArbOS performs Logical Arbitrage entirely within Polymarket — it finds situations where the math of the order books breaks the rules of logic, then executes delta-neutral trades to lock in a mathematically guaranteed, risk-free profit.

ArbOS is a polyglot monorepo spanning three languages: Rust for the microsecond-sensitive data and execution path, Python for the AI reasoning layer, and TypeScript/React for the real-time control dashboard — unified by a PostgreSQL data plane and an IBM watsonx.ai reasoning core.

Built for the IBM SkillsBuild Hackathon · Fintech Track · February 2026.

Metric	Value
Arbitrage detection latency	~14 ms end-to-end
Simultaneous markets tracked	up to 1,000 live order books
Arbitrage strategies	4 (Implication, Partition, Mutual Exclusion, AI-Chained)
Minimum profit gate	$0.50 net, after fees + gas
Order safety model	Fill-or-Kill only, with atomic panic-dump failsafe
Reasoning engine	IBM watsonx.ai (`mistralai/mistral-large`) via LangGraph

The Alpha: Logical Arbitrage

Consider two markets:

Market A — Will the Fed cut rates in June 2026? → trading at 60¢ (60% implied)
Market B — Will the Fed cut rates in 2026? → trading at 55¢ (55% implied)

June 2026 is a strict subset of the year 2026. It is logically impossible for the Fed to cut in June without also having cut in 2026. Therefore P(A) can never exceed P(B) — yet the market prices A higher than B. The order books have contradicted the laws of logic.

flowchart LR
A["<b>Market A</b><br/>Fed cut in June 2026<br/><b>60¢</b>"]
B["<b>Market B</b><br/>Fed cut in 2026<br/><b>55¢</b>"]
A -- "A ⊆ B, so P(A) ≤ P(B)<br/>VIOLATION: 60¢ > 55¢" --> B
A -. "SELL @ 60¢" .-> S["🔒 Locked-in profit<br/><b>+5¢ / share</b><br/>(risk-free)"]
B -. "BUY @ 55¢" .-> S
classDef mkt fill:#16213E,stroke:#00D4AA,color:#F8FAFC;
classDef win fill:#0A3D2E,stroke:#00FF88,color:#F8FAFC;
class A,B mkt;
class S win;

Why it's guaranteed, not merely probable: a cross-exchange spread can close against you before you execute. A logical violation cannot — if June ⊆ 2026, that is true regardless of what the Fed actually does. The profit is locked in the moment both legs fill.

The Four Arbitrage Strategies

#	Strategy	Axiom	Violation	Action
1	Implication (Monotonicity)	If `A ⊆ B`, then `P(A) ≤ P(B)`	`P(A)_bid > P(B)_ask + fees`	SELL sub-event A, BUY super-event B
2	Partition (Normalization)	`Σ P(Eᵢ) = 1.0`	`Σ P(Eᵢ)_bid > 1.0 + fees`	Basket-SELL every outcome
3	Mutual Exclusion (Contradiction)	`P(A) + P(B) ≤ 1.0`	`P(A)_bid + P(B)_bid > 1.0`	SELL both sides
4	Chained Implication (AI-discovered)	If `A → B → C`, then `P(A) ≤ P(C)`	`P(A) > P(C)`	Multi-hop; mapped by the AI Brain

Worked example — Implication arb (Fed Cut June 60¢ → Fed Cut 2026 55¢, 100 shares):

SELL June YES @ 60¢ = $60.00 collected
BUY 2026 YES @ 55¢ = $55.00 paid
────────────────────────────────────────
Gross spread: +$5.00
Polymarket taker fee: $0.00 (0% on most markets)
Polygon gas (2 txns): -$0.06
────────────────────────────────────────
Net profit: +$4.94 (≈8.2% ROI on $60 risked)

System Architecture

ArbOS is organized into five tiers, each with a single clear responsibility. The three Rust tiers compile into one high-performance orchestrator binary; the Brain runs as an independent Python microservice; the dashboard is a static SPA.

flowchart TB
subgraph EXT["External Services"]
GAMMA["Polymarket Gamma API<br/>(REST · market metadata)"]
CLOB["Polymarket CLOB<br/>(WebSocket · order books)"]
WX["IBM watsonx.ai<br/>(mistral-large LLM)"]
GAS["Polygon Gas Station<br/>(gas oracle)"]
end
subgraph BRAIN["🧠 Tier 2 · Brain (Python / FastAPI)"]
SCHED["APScheduler<br/>(15-min discovery)"]
LG["LangGraph pipeline<br/>validate → classify → persist"]
NX["NetworkX DiGraph"]
PG[("PostgreSQL 15<br/>markets · relationships")]
end
subgraph RUST["🦀 Rust Orchestrator (single binary · crates/bot)"]
ING["Tier 1 · Ingestor<br/>WebSocket feed + VWAP"]
ENG["Tier 3 · Engine<br/>arb math + signals"]
EXEC["Tier 4 · Executor<br/>rate-limit · dedup · FOK"]
LEDGER["Ledger<br/>P&L · positions"]
API["Axum Server :3001<br/>REST + WebSocket"]
end
UI["🖥️ Tier 5 · Dashboard<br/>(React / Vite / TS)"]
GAMMA --> SCHED
WX --> LG
SCHED --> LG --> PG
LG --> NX
PG -- "GET /state (every 60s)" --> ENG
CLOB --> ING
GAS --> ENG
ENG -- "Subscribe / Unsubscribe" --> ING
ING -- "NormalizedOrderbook" --> ENG
ENG -- "ArbSignal" --> EXEC
EXEC --> LEDGER
EXEC -- "ExecutionReport (broadcast)" --> API
API -- "WebSocket + /api/status" --> UI
classDef ext fill:#1a1a2e,stroke:#7B61FF,color:#F8FAFC;
classDef brain fill:#16213E,stroke:#00D4AA,color:#F8FAFC;
classDef rust fill:#2A1A0E,stroke:#DEA584,color:#F8FAFC;
classDef ui fill:#0A2540,stroke:#61DAFB,color:#F8FAFC;
class GAMMA,CLOB,WX,GAS ext;
class SCHED,LG,NX,PG brain;
class ING,ENG,EXEC,LEDGER,API rust;
class UI ui;

Two independent control loops

Brain discovery loop (15-min cadence): APScheduler syncs the latest markets from Gamma, then runs the LLM over all unprocessed pairs to grow the relationship graph.
Engine execution loop (60-s cadence): the Rust Engine re-polls /state, diffs the tracked-asset set, dynamically manages live WebSocket subscriptions, and refreshes the gas oracle — all while continuously evaluating incoming order-book updates in real time.

How It Works: End-to-End Flow

sequenceDiagram
autonumber
participant G as Gamma API
participant B as Brain (Python)
participant W as watsonx.ai
participant E as Engine (Rust)
participant I as Ingestor (Rust)
participant C as Polymarket CLOB
participant X as Executor (Rust)
participant U as Dashboard
Note over B: Discovery loop — every 15 min
B->>G: fetch active markets (volume ≥ 1000)
G-->>B: market list
B->>W: classify each market pair
W-->>B: IMPLIES / MUT_EXCL / INDEPENDENT + confidence
B->>B: persist relationships → Postgres + NetworkX
Note over E: Execution loop — every 60 s
E->>B: GET /state (X-API-Key)
B-->>E: implications, partitions, contradictions, timestamps
E->>I: Subscribe(asset_ids)
I->>C: subscribe order books (WebSocket)
loop Real-time
C-->>I: order book update
I->>I: compute VWAP to $100 depth
I-->>E: NormalizedOrderbook
E->>E: run implication / partition / contradiction checks
alt Violation & net profit > $0.50
E-->>X: ArbSignal (delta-neutral legs)
X->>X: rate-limit + dedup
X->>C: place Fill-or-Kill orders
C-->>X: fills
X-->>U: ExecutionReport (WebSocket broadcast)
end
end

The Five Tiers

🦀 Tier 1 — Ingestor (`crates/ingestor`)

Maintains a real-time, normalized view of the market. A Tokio actor connects to Polymarket's CLOB WebSocket via the official polymarket-client-sdk, dynamically subscribes/unsubscribes to per-asset order books (capped at 1,000), and computes VWAP up to a $100 liquidity target. Books that can't supply $100 of depth are flagged UNTRADEABLE. Reconnects with exponential backoff.

🧠 Tier 2 — Brain (`services/brain`)

Discovers the "Hidden Graph" of logical relationships between markets. A FastAPI service that uses IBM watsonx.ai (mistralai/mistral-large) orchestrated by a LangGraphStateGraph (validate → classify → persist) to classify each market pair as IMPLIES / MUTUALLY_EXCLUSIVE / INDEPENDENT with a confidence score and reasoning. Relationships persist to PostgreSQL (SQLModel + asyncpg) and an in-memory NetworkX DiGraph. An APScheduler job discovers new relationships every 15 minutes. Exposes /state — the asset-level graph payload the Rust Engine consumes.

🦀 Tier 3 — Engine (`crates/engine`)

The high-frequency calculator. Caches the entire logical graph in memory, polls the Brain's /state every 60 s (trusting only edges ≥ 0.90 confidence), and runs all three arbitrage strategies on every order-book update. Models taker fees (1.56% conservative) and live Polygon gas (with a 50% safety margin), firing an ArbSignal only when net profit clears $0.50. Enforces a max-duration filter (skips markets resolving > 30 days out).

🦀 Tier 4 — Bot / Executor (`crates/bot`)

The primary orchestrator entry point and atomic settlement layer. main.rs spawns all Rust tiers as Tokio tasks wired by MPSC channels. The executor applies a token-bucket rate limiter (10 req/s), canonical-hash deduplication (30 s cooldown), and places Fill-or-Kill orders via the Polymarket SDK — with automatic panic-dump of a filled leg if its pair fails. An in-memory ledger tracks P&L and positions; an Axum server exposes REST + WebSocket endpoints to the dashboard. Supports Demo (paper-trading) and Live modes.

🖥️ Tier 5 — Dashboard (`frontend`)

A React 18 + Vite + TypeScript SPA styled with Tailwind CSS and shadcn/ui. Features a d3-force relationship graph, an AI chat graph-builder, a live control room (P&L ticker, trade log, open positions, system health), and a risk-configuration wizard. Guided operator flow: Landing → Connect → Graph → Configure → Dashboard → Guide.

🦀 Tier 0 — Core (`crates/core`)

Shared Rust foundation: canonical domain types (ArbSignal, NormalizedOrderbook, ExecutionReport, TradeAction) and tuning constants (fees, gas model, thresholds, channel buffer sizes). All money math uses rust_decimal fixed-point — never floats.

Tech Stack

mindmap
root((ArbOS))
Rust
Tokio async runtime
Axum web/WebSocket
polymarket-client-sdk
rust_decimal
reqwest
tracing
Python
FastAPI
LangGraph + LangChain
langchain-ibm watsonx.ai
SQLModel + asyncpg
NetworkX
APScheduler
loguru
TypeScript
React 18 + Vite
shadcn/ui + Radix
TanStack Query
d3 + Recharts
Tailwind CSS
framer-motion
Platform
Polymarket CLOB + Gamma
Polygon PoS
PostgreSQL 15
Docker + Compose
GitHub Actions CI

Layer	Technologies
Languages	Rust (edition 2024, toolchain 1.93), Python ≥ 3.13, TypeScript 5.8
Rust runtime & web	Tokio 1.49, Axum 0.8, tower-http, tokio-util, futures
Blockchain / market	`polymarket-client-sdk` 0.4.2 (CLOB WS + EIP-712 signing), reqwest 0.12
AI / LLM	IBM watsonx.ai (`mistralai/mistral-large`), LangGraph 1.0, LangChain-Core 1.2, langchain-ibm 1.0
Python web/data	FastAPI 0.133, Uvicorn 0.41, SQLModel 0.0.37, asyncpg 0.31, Pydantic 2.12, NetworkX 3.6, APScheduler 3.11, httpx 0.28
Frontend	React 18.3, Vite 5.4, TanStack Query 5.83, Radix/shadcn, d3 7.8, Recharts 2.15, Tailwind 3.4, framer-motion 11, zod 3.25
Database	PostgreSQL 15
Package managers	Cargo (Rust), uv (Python), npm (frontend)
Testing	`cargo test`, pytest + pytest-asyncio, Vitest + Testing Library
Tooling	rustfmt, clippy, ruff, ESLint, pre-commit, Dependabot
Infra & CI/CD	Docker + Docker Compose, GitHub Actions

Getting Started

Prerequisites

Docker & Docker Compose (for the full-stack launch)
Rust 1.93+ with Cargo (for Rust development)
Python 3.13+ with uv (for the Brain)
Node.js 18+ with npm (for the frontend)
Credentials: an IBM watsonx.ai API key + project ID, and (for live trading) Polymarket API key / passphrase / secret

Quick start (Docker Compose)

# 1. Clone
git clone git@github.com:Harikeshav-R/Arb-OS.git
cd Arb-OS
# 2. Configure — copy the template and fill in your keys
cp .env.example .env
# edit .env → WATSONX_APIKEY, WATSONX_PROJECT_ID, POLYMARKET_* , etc.# 3. Launch the stack (Postgres + Brain + Ingestor + Frontend)
make up # or: docker compose up -d# 4. Watch it run
make ps # container status
make logs # tail all logs (make logs S=brain for one service)

Service	URL
Brain API (FastAPI)	http://localhost:8000 · docs at `/docs`
Frontend dashboard	http://localhost:5173
Bot API + WebSocket	http://localhost:3001
PostgreSQL	`postgresql://postgres:password@localhost:5432/arbos`

Common Make targets

make up # start all services in the background
make down # stop and remove containers
make build # rebuild images and start
make logs # tail logs (S=<service> for one)
make db # open psql against the database
make test# run all tests (Rust + Brain)
make clean # stop and remove volumes (⚠ deletes DB data)

Running tiers individually

# Rust orchestrator (spawns Ingestor + Engine + Executor + API server)
cargo run -p bot
# Brain (Python / FastAPI)cd services/brain && uv run uvicorn main:app --reload
# Frontend (Vite dev server)cd frontend && npm install && npm run dev

Configuration

All configuration is via environment variables (see .env.example).

Polymarket & infrastructure

Variable	Description
`POLYMARKET_API_KEY` / `PASSPHRASE` / `SECRET`	CLOB API credentials (required for live trading)
`DATABASE_URL`	PostgreSQL DSN (`postgresql://…`)
`REDIS_URL`	Reserved for future cross-service signal bus

IBM watsonx.ai (Brain / Tier 2)

Variable	Default	Description
`WATSONX_APIKEY`	—	watsonx.ai API key (required)
`WATSONX_URL`	`https://us-south.ml.cloud.ibm.com`	watsonx.ai region endpoint
`WATSONX_PROJECT_ID`	—	watsonx.ai project ID (required)
`WATSONX_MODEL_ID`	`mistralai/mistral-large`	LLM model

Brain tuning

Variable	Default	Description
`BRAIN_LLM_TEMPERATURE`	`0.0`	LLM sampling temperature
`BRAIN_LLM_MAX_TOKENS`	`1024`	Max tokens per LLM call
`BRAIN_CONFIDENCE_THRESHOLD`	`0.7`	Min confidence to persist a relationship
`BRAIN_GAMMA_API_BASE_URL`	`https://gamma-api.polymarket.com`	Gamma API base
`BRAIN_GAMMA_MIN_VOLUME`	`1000`	Min market volume to track
`BRAIN_GAMMA_FETCH_LIMIT`	`100`	Markets fetched per page

Bot / Engine (Rust)

Variable	Default	Description
`ARBOS_LIVE_MODE`	`false`	`false` → Demo (paper); `true` → live execution
`ARBOS_INITIAL_ASSETS`	—	CSV of asset IDs to seed subscriptions
`ARBOS_WS_PORT`	`3001`	Bot REST + WebSocket port
`ARBOS_DEDUP_COOLDOWN`	`30`	Signal dedup cooldown (seconds)
`ARBOS_MAX_HISTORY`	`500`	Ledger trade-history ring-buffer size
`BRAIN_API_URL`	`http://localhost:8000`	Engine → Brain endpoint
`ADMIN_API_KEY`	—	`X-API-Key` for protected Brain endpoints
`VITE_ALLOWED_ORIGINS`	`http://localhost:5173`	CORS allow-list for the Bot API

Key thresholds (crates/core/src/constants.rs): TARGET_LIQUIDITY = $100, MIN_PROFIT_THRESH_USDC = $0.50, MIN_CONFIDENCE_THRESHOLD = 0.90, MAX_TRACKED_ASSETS = 1000, EXECUTOR_RATE_LIMIT_PER_SEC = 10. Note the dual confidence gate: the Brain persists edges at ≥ 0.70 but the Engine only trades edges at ≥ 0.90.

API Reference

Brain (FastAPI · `:8000`)

Method	Path	Auth	Purpose
`GET`	`/health`	—	DB + LLM configuration status
`POST`	`/markets/sync`	admin	Upsert markets from Gamma
`GET`	`/markets`	—	List cached markets (paginated)
`POST`	`/analyze`	—	Analyze a single market pair
`POST`	`/scan`	admin	Scan top-N markets, analyze all unprocessed pairs
`GET`	`/relationships`	—	List active relationships (filter by `logic_type`)
`GET`	`/relationships/{condition_id}`	—	Relationships involving a market
`GET`	`/graph/stats`	—	Graph summary (edge/component counts)
`GET`	`/state`	admin	Full asset-level graph payload for the Engine

Admin endpoints require an X-API-Key header when ADMIN_API_KEY is set.

Bot (Axum · `:3001`)

Method	Path	Purpose
`GET`	`/api/health`	Liveness probe
`GET`	`/api/status`	JSON snapshot: mode, cumulative P&L, signals executed, uptime, open positions, recent trades
`GET`	`/ws`	WebSocket — sends a `SNAPSHOT` on connect, then streams live `EXECUTION_REPORT` messages

Project Structure

Arb-OS/
├── Cargo.toml # Rust workspace: core, ingestor, engine, bot
├── docker-compose.yml # postgres + brain + ingestor + frontend
├── Makefile # Docker lifecycle + test targets
├── .env.example # All environment variables
├── PROJECT.md # Architecture & theory spec
├── REPORT.md # Detailed engineering report
├── AGENTS.md # AI-agent contributor guide
│
├── crates/ # ── RUST WORKSPACE ──
│ ├── core/ # Tier 0: shared domain types + constants
│ ├── ingestor/ # Tier 1: WebSocket feed + VWAP normalization
│ │ └── src/{main,actor,clob_client}.rs
│ ├── engine/ # Tier 3: strategy math + signal generation
│ │ └── src/{main,actor}.rs, strategies/{implication,partition,contradiction}.rs
│ └── bot/ # Tier 4: orchestrator + executor + server
│ └── src/{main,config,executor,ledger,dedup,server}.rs
│
├── services/
│ └── brain/ # ── PYTHON (Tier 2: AI) ──
│ ├── main.py # FastAPI app + APScheduler + endpoints
│ ├── graph.py # LangGraph pipeline (validate→classify→persist)
│ ├── gamma_client.py # Polymarket Gamma API client
│ ├── relationship_graph.py # NetworkX DiGraph manager
│ ├── models.py # SQLModel tables + API schemas
│ ├── database.py # Async SQLAlchemy/asyncpg engine
│ ├── config.py # Pydantic settings
│ ├── prompts.py # LLM prompt templates
│ └── tests/ # pytest suite
│
├── frontend/ # ── REACT/VITE/TS (Tier 5: Dashboard) ──
│ └── src/
│ ├── App.tsx # Router: /, /connect, /graph, /configure, /dashboard, /guide
│ ├── pages/ # Index, Connect, Graph, Configure, Dashboard, Guide
│ ├── components/ # ForceGraph, Navbar, WebGLCanvas + shadcn/ui
│ └── hooks/
│
└── .github/workflows/ci.yml # 3 CI jobs: rust-check, python-check, frontend-check

Development

Testing

make test# everything
make test-rust # cargo test (engine, ingestor, core)
make test-brain # pytest suite for the Braincd frontend && npm test# Vitest

Linting & formatting

cargo fmt --all && cargo clippy --all-targets --all-features -- -D warnings
cd services/brain && uv run ruff check .cd frontend && npm run lint

Pre-commit hooks (.pre-commit-config.yaml) run rustfmt, clippy, ruff, and the frontend lint/typecheck automatically:

pip install pre-commit && pre-commit install

CI pipeline

Every push/PR to main runs three parallel GitHub Actions jobs:

flowchart LR
P["push / PR → main"] --> R["rust-check<br/>fmt · clippy -D warnings · cargo test"]
P --> Y["python-check<br/>uv sync · ruff · pytest"]
P --> F["frontend-check<br/>npm ci · tsc --noEmit"]
R --> M{"all green?"}
Y --> M
F --> M
M -->|yes| OK["✅ mergeable"]
classDef job fill:#16213E,stroke:#00D4AA,color:#F8FAFC;
class R,Y,F job;

Git workflow

ArbOS uses Conventional Commits with scoped branches (type/scope/description):

Types:feat/, fix/, chore/, refactor/
Scopes:engine, ingestor, bot, brain, web, infra

Example: feat(engine): add VWAP calculator · fix(brain): gamma-api filter

Safety & Risk Engineering

Guaranteed-at-resolution does not mean guaranteed-before-resolution. ArbOS hardens against every real-world failure mode:

Hazard	Defense
Single-leg exposure	Panic-dump: close the filled leg immediately; pause after 2 consecutive
Fakeout / MEV liquidity	Fill-or-Kill only — never GTC; VWAP pre-check to `$100` depth
Oracle settlement delay	Max-duration filter: refuse markets resolving > 30 days out
API rate limiting (429)	Token-bucket Governor capped at 10 req/s
Slippage	VWAP-adjusted spread check; FOK cancels if unfillable
Duplicate firing	Canonical-hash dedup with a 30 s cooldown + cache eviction
Cascading failures	Circuit breaker: 3 consecutive failures → auto-pause + alert
WebSocket drop	Pause + exponential backoff (1→2→4→8→30 s) + full state re-sync
False relationship	Dual confidence gate: Brain persists ≥ 0.70, Engine trades ≥ 0.90

All tiers emit structured logs — tracing (Rust), loguru JSON (Python), pino/winston (TS).

Deployment

ArbOS ships fully containerized via docker-compose.yml (Postgres + Brain + Ingestor + Frontend), with multi-stage, non-root Dockerfiles for the Brain (uv) and Ingestor (rust:1.93-slim). The container-native design maps cleanly onto a managed, highly-available cloud topology:

flowchart TB
DNS["Route 53"] --> CF["CloudFront + S3<br/>(dashboard SPA)"]
DNS --> ALB["Application Load Balancer<br/>(TLS + WebSocket)"]
ALB --> B["ECS Fargate · Brain"]
ALB --> I["ECS Fargate · Ingestor"]
ALB --> BOT["ECS Fargate · Bot"]
B --> RDS[("RDS PostgreSQL<br/>Multi-AZ")]
BOT --> REDIS[("ElastiCache Redis")]
ECR["Amazon ECR"] -. images .-> B
ECR -. images .-> I
ECR -. images .-> BOT
SM["Secrets Manager<br/>API + wallet keys"] -. secrets .-> BOT
classDef aws fill:#232F3E,stroke:#FF9900,color:#F8FAFC;
class DNS,CF,ALB,B,I,BOT,RDS,REDIS,ECR,SM aws;

See REPORT.md for the full reference architecture (Terraform IaC, Prometheus/Grafana observability, CloudWatch, ECR image scanning).

Documentation

Document	Contents
`PROJECT.md`	Architecture & theoretical foundations (the "alpha"), tier specs, launch checklist
`REPORT.md`	Detailed engineering report + inferred production cloud topology
`AGENTS.md`	Context & rules for AI-agent contributors
`CONTRIBUTING.md`	Contribution guidelines
`CODE_OF_CONDUCT.md`	Community standards

Contributing

Contributions are welcome! Please read CONTRIBUTING.md and follow the Conventional Commits + scoped-branch workflow. Never push directly to main — open a clean, modular pull request. Ensure cargo fmt, clippy, ruff, and the frontend typecheck all pass before submitting.

License

Built for the IBM SkillsBuild Hackathon · Fintech Track · February 2026

Powered by IBM watsonx.ai & Granite

Name		Name	Last commit message	Last commit date
.github		.github
crates		crates
frontend		frontend
services/brain		services/brain
.dockerignore		.dockerignore
.env.example		.env.example
.gitignore		.gitignore
.pre-commit-config.yaml		.pre-commit-config.yaml
AGENTS.md		AGENTS.md
CODE_OF_CONDUCT.md		CODE_OF_CONDUCT.md
CONTRIBUTING.md		CONTRIBUTING.md
Cargo.toml		Cargo.toml
LICENSE		LICENSE
Makefile		Makefile
PROJECT.md		PROJECT.md
README.md		README.md
REPORT.md		REPORT.md
docker-compose.yml		docker-compose.yml

Latest commit

History

Folders and files

Repository files navigation

⚡ ArbOS

Automated Prediction Market Arbitrage Engine

Table of Contents

What is ArbOS?

The Alpha: Logical Arbitrage

The Four Arbitrage Strategies

System Architecture

Two independent control loops

How It Works: End-to-End Flow

The Five Tiers

🦀 Tier 1 — Ingestor (crates/ingestor)

🧠 Tier 2 — Brain (services/brain)

🦀 Tier 3 — Engine (crates/engine)

🦀 Tier 4 — Bot / Executor (crates/bot)

🖥️ Tier 5 — Dashboard (frontend)

🦀 Tier 0 — Core (crates/core)

Tech Stack

Getting Started

Prerequisites

Quick start (Docker Compose)

Common Make targets

Running tiers individually

Configuration

Polymarket & infrastructure

IBM watsonx.ai (Brain / Tier 2)

Brain tuning

Bot / Engine (Rust)

API Reference

Brain (FastAPI · :8000)

Bot (Axum · :3001)

Project Structure

Development

Testing

Linting & formatting

CI pipeline

Git workflow

Safety & Risk Engineering

Deployment

Documentation

Contributing

License

About

Resources

Code of conduct

Contributing

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

🦀 Tier 1 — Ingestor (`crates/ingestor`)

🧠 Tier 2 — Brain (`services/brain`)

🦀 Tier 3 — Engine (`crates/engine`)

🦀 Tier 4 — Bot / Executor (`crates/bot`)

🖥️ Tier 5 — Dashboard (`frontend`)

🦀 Tier 0 — Core (`crates/core`)

Brain (FastAPI · `:8000`)

Bot (Axum · `:3001`)