Provepy: Towards Autoformalizing python code using AI and Lean

This package provides a decorator @provable that tries to automatically formalize and prove your python functions.

Functions which fail to be proved will raise a VerificationError. Functions which are proved to be correct will just run normally.

Only works for simple functions without other decorators for now.

For best results, use a frontier model.

This package is in alpha and contributions are welcome.

Caution

This is experimental software and the proofs may not be perfect. Do not rely on this for critical applications.

Installation

Step 1: Install the package

To install from source (recommended for latest updates), clone the repository and run:

cd provepy
pip install -e .

To install from PyPI, run:

pip install provepy

Step 2: Setup the lean environment

Navigate to your project folder ( from where you will run the python command) and run:

provepy init

Then add provepy_lean_project to your gitignore file if you have one as provepy_lean_project will contain a new git repository.

This will setup a lean environment in your project directory. It involves a large download so might take a while.

Usage

Add the decorator @provable to every function you want to prove and pass the spec you want the function to follow and the context (other functions used) to the decorator. Then just run the python command normally. Functions passed as context should be implemented in python, not C.

Examples:

from provepy import provable

@provable(claim="This function returns the sum of its inputs")
def add(a: int, b: int) -> int:
    return a+b

from provepy import provable

def addTwo(a: int) ->int:
    return a+2

@provable(claim="This function returns the sum of its inputs plus two",context=[addTwo])
def add(a: int, b: int) -> int:
    return addTwo(a+b)

Type annotations are highly recommended. Currently, using only builtin types is also highly recommended.

Configuration

The package supports using LLMs either via OpenRouter or through a OpenAI compatible api or directly through Google's Gemini API. You can toggle between these services using the LLM_PROVIDER environment variable. By default, the package uses the gemini api.

Provider Selection

Set the environment variable LLM_PROVIDER to choose your backend:

• gemini (default)

• custom

• openrouter

---

Google Gemini Configuration

To use Google's official API, ensure LLM_PROVIDER is unset or is set to LLM_PROVIDER=gemini. You can create a free API key from Google AI Studio. However generation on the free tier is very slow.

• GEMINI_API_KEY: Your Google API key (Required).
• GEMINI_MODEL_NAME: The exact model name used in the Google API (e.g., gemini-3-flash-preview). If not set, it defaults to the Gemma-31B model.
• BETTER_GEMINI_MODEL_NAME: (Optional) Specify a more powerful model (e.g., gemini-3.1-pro-preview) to use as a fallback when the primary model fails to prove a function.

---

Custom Configuration

To use a custom Open AI compatible api, ensure LLM_PROVIDER is explicitly set to custom.

• CUSTOM_API_KEY: Your custom API key (Required).
• CUSTOM_API_URL: Your custom API URL (Required).
• CUSTOM_MODEL_NAME: The specific model routing string to use.(Required)
• BETTER_CUSTOM_MODEL_NAME: (Optional) Specify a more powerful model to use as a fallback when the default, smaller model fails to prove a function.

---

OpenRouter Configuration

To use OpenRouter, ensure LLM_PROVIDER is explicitly set to openrouter.

• OPENROUTER_API_KEY: Your OpenRouter API key (Required).
• OPENROUTER_MODEL_NAME: The specific model routing string to use. If not set, this defaults to google/gemma-4-31b-it:free.
• BETTER_OPENROUTER_MODEL_NAME: (Optional) Specify a more powerful model to use as a fallback when the default, smaller model fails to prove a function.

Logging

To turn off Logs which are printed to stdout, set the environment variable PROVEPY_LOG to 0

TODO:

1. Support other LLM Providers
2. Improve pulling context from outside the function
3. Allow other decorators on functions
4. Add support for repls
5. Add tests
6. Resolve name clashes between provided python functions and builtin mathlib objects
7. Store proofs for manual review
8. Improve prompts and use custom system prompts