零纸AI研究工作流

Prerequisites

```bash

Easiest setup (no git or CLI knowledge needed)

If you already have Claude Code installed, open it in any empty folder and paste this in:

Set up an autonomous finance research project in this folder.

1. Clone https://github.com/alejandroll10/zeropaper into a temp location
2. From there, run ./setup.sh my-paper --variant finance
   (or --variant finance --ext empirical if I want CRSP/Compustat data)
3. Move the resulting my-paper/ folder here
4. Check that I have the prerequisites installed (python3, uv, git; bubblewrap on Linux).
   If anything is missing, walk me through installing it on my machine (Mac or Linux).
5. When setup is done, tell me to cd into my-paper and say "Run the pipeline."

Claude Code will handle the clone, setup, and prereq checks for you. Works on Mac and Linux.

Git identity (one-time, used by every setup.sh run)

git config --global user.email "you@example.com" git config --global user.name "Your Name" ```

Edit .env with your API keys (created by setup.sh)

nano .env ```

Project structure (after setup)

my-paper/
├── CLAUDE.md                 # Claude Code orchestration (assembled by setup.sh)
├── AGENTS.md                 # Codex orchestration (assembled by setup.sh)
├── GEMINI.md                 # Gemini CLI orchestration (assembled by setup.sh)
├── .env                      # API keys (gitignored)
├── dashboard.html            # Live progress dashboard
├── .claude/
│   ├── settings.json         # Sandbox config
│   ├── agents/               # Claude subagents (.md)
│   └── skills/               # Claude skills
├── .codex/
│   └── agents/               # Codex custom agents (.toml)
├── .gemini/
│   ├── settings.json         # Gemini config
│   └── agents/               # Gemini subagents (.md)
├── .agents/
│   └── skills/               # Shared skills (Codex + Gemini)
├── output/                   # Pipeline outputs by stage
├── paper/                    # LaTeX paper
│   ├── main.tex
│   ├── sections/
│   └── simulated_referee_reports/
├── code/
│   ├── analysis/             # Analysis and verification scripts
│   ├── download/             # Data download helpers
│   ├── explore/              # Exploration scripts and diagnostics
│   ├── tmp/                  # Scratch files
│   └── utils/                # Utility scripts (including codex-math; more with extensions)
└── process_log/
    ├── pipeline_state.json   # Current stage, scores, history
    └── history.md

Quick start

Step 3: Configure credentials (if using extensions)

```bash cd my-paper

System packages

uv (Python package manager)

curl -LsSf https://astral.sh/uv/install.sh | sh

Combine extensions

./setup.sh my-paper --variant finance --ext empirical --ext theory_llm

Manual mode (research toolkit — no autonomous pipeline, you drive the agents)

./setup.sh my-toolkit --manual

Extensions

Extensions are additive and combinable — they inject extra agents and skills without changing the core pipeline. Use multiple --ext flags to combine them.

Pipeline stages

Stage 0: Problem Discovery   → Gate 0: Problem Viability
Stage 1: Idea Generation     → Gate 1: Idea Review (iterates)
                                Gate 1b: Novelty Check on idea
                                Gate 1c: Idea Prototype (tractability)
Stage 2: Theory Development  → Gate 2: Math Audit (structured + free-form)
                                Gate 3: Novelty Check on theory
                                Stage 2b: Theory Exploration (compute, verify, plot)
                                Gate 3a-feasibility: Empirical Feasibility (optional)
Stage 3: Implications
Stage 3a: Full Empirical Analysis (optional, if --ext empirical)
Stage 3b: LLM Experiments         (optional, if --ext theory_llm)
Stage 4: Self-Attack          → Gate 4: Scorer Decision
Stage 5: Paper Writing
Stage 6: Referee Simulation   → Gate 5: Referee Decision
Stage 7: Style Check
Stage 8: Bibliography Verify
Stage 9: Polish               → Done (eight parallel polish agents — consistency,
                                 formula, numerics, institutions, equilibria,
                                 identification, bibliography, prose — triaged
                                 + applied; max 2 rounds)

Each gate is adversarial. Failed theories get revised, reworked, or abandoned. The system loops until it produces a paper that passes simulated referee review.