Zhanli-Li / Data-World

Scaling Data Environments from Scratch 
 Specify the world behind the files. Generate the data. Ground the supervision. 
 English · 中文 
 Overview ·
 Quick Start ·
 Documentation 
---
## Overview
**Data World is a mechanism-first, multi-agent engine for synthesizing data-analysis environments, tasks, and rubrics from scratch.** Rather than exploring existing files and writing criteria for whichever findings happen to be discovered, it specifies the generative mechanisms first, builds a coherent canonical state, and then produces imperfect analyst-visible observations. An independent auditor checks the specified relationships and whether targeted findings can be recovered from those files.
The result is more than a collection of synthetic tables: it is a **data world with grounded analytical targets**, supporting evaluation and rubric-based rewards for data agents.
 From a dependency DAG to observed files, audited targets, and task-specific supervision. 
| Design | What it provides |
|---|---|
| **Mechanisms before files** | A world contract fixes entities, timelines, generative relationships, invariants, and observation systems before data generation. |
| **Layer-wise canonical state** | DAG layers are built in topological order; committed upstream state is read-only input to subsequent layers. |
| **Imperfect observations** | Missingness, delays, identifier mismatches, duplicates, and drift arise from explicit observation mechanisms, not arbitrary noise alone. |
| **Independent audit** | A separate agent checks relationships, global invariants, and target recoverability, recording an audit and a world guide. |
| **Grounded rubrics** | Tasks follow upstream causes or downstream consequences; atomic criteria separate `data_computation` from `insight_discovery`. |
 Why specify the world before generating its files? 
Exploring existing data can miss valid findings and leave them without matching evaluation criteria. Data World uses the generative record to establish selected targets and ground their rubrics, then audits their recoverability from observed files. This does not guarantee exhaustive coverage of every valid insight.
## Quick Start
### 1. Install
Run from the repository root (the directory containing `pyproject.toml`):
- **Python 3.9+** for the engine.
- **Node.js 18+ and npm** for a working Claude Code CLI (`claude`).
- A model service supporting the **OpenAI-compatible contract call** and the **Claude Code agent stages**; a Claude Code login alone does not configure the contract call.
 [code] 
Optional trajectory capture requires Node.js 22+ and `@musistudio/claude-code-router@3.0.20`; see trajectory capture and endpoint configuration.
### 2. Configure your model service
 [code] 
Edit `.env` and fill in `ANTHROPIC_AUTH_TOKEN`, `ANTHROPIC_BASE_URL`, and `ANTHROPIC_MODEL`. The CLIs load this file automatically; do not put credentials in command-line arguments or commit them. See `.env.example` for optional judge overrides.
### 3. Generate