Repro Lens

Your refactor runs twice. Did it keep the same results?

Explore Demo · Examples · Contribute

Catch reproducibility risks before a commit, replay an experiment, and compare outputs before and after a change. Built for ML developers and AI coding agents.

Check for reproducibility risks, verify runs, and compare outputs instantly.

Machine Learning code is notorious for silent reproducibility failures. Repro Lens statically analyzes your ML code to find non-deterministic behavior and verifies that refactors preserve your exact outputs.

• Zero overhead: Static analysis takes milliseconds. No ML dependencies or API keys required.
• Broad support: Native rules for scikit-learn, XGBoost, LightGBM, PyTorch, TensorFlow, and Lightning.
• Agent ready: Out-of-the-box skills to let your AI coding assistant audit and fix reproducibility issues autonomously.

Install using uv (or pip):

uv tool install 'git+https://github.com/00200200/repro-lens.git@v0.3.1'

Scan your project instantly:

repro-lens check --root /path/to/project

Example output:

train_test_split(X, y)  # ❌ R101: no explicit random_state

Try a complete experiment validation:

repro-lens init repro-demo --name repro_demo
cd repro-demo
uv sync --locked
repro-lens verify

Pre-commit hook: Prevent reproducibility issues from being committed.

repos:
  - repo: https://github.com/00200200/repro-lens
    rev: v0.3.1
    hooks:
      - id: repro-lens

Coding Agents: Equip your LLMs with the Reproducibility Skill to automatically fix non-deterministic ML code.

"Use the reproducibility skill to refactor this training script while preserving its declared outputs."

• Framework Coverage
• Verification Contract
• Community Support
• Architecture & Rules