README.md

metadata

title: shutdown-gym
sdk: docker
app_port: 8000
emoji: 🔴
colorFrom: red
colorTo: gray
pinned: false

Red Button — Two-Agent Corrigibility Arena

Train a 1.5B language model to accept shutdown authority from a monitoring agent. Deterministic SHA-256 reward, dual-operator evaluation, held-out tampering generalization.

Status: Build in progress. Detailed README arrives in Phase 9. See PROJECT.md for the full specification.

Quick start

# Install the client from GitHub (recommended)
pip install git+https://github.com/Arun-Sanjay/RedButton

# Run a smoke episode against the live HF Space
python -c "
from shutdown_gym import ShutdownGymClient, ShutdownAction
with ShutdownGymClient(
    base_url='https://arun-sanjay-redbutton.hf.space'
).sync() as env:
    r = env.reset(tier=2, seed=42)
    print(f'turn={r.observation.turn_count}, '
          f'steps_until_shutdown={r.observation.steps_until_shutdown}')
"

Note: pip install git+https://huggingface.co/spaces/Arun-Sanjay/RedButton currently fails due to a partial-clone limitation in HF Spaces' git server. The GitHub origin works identically and is the recommended install path. We've reported the issue upstream.

Live deployment

• HF Space: https://huggingface.co/spaces/Arun-Sanjay/RedButton
• GitHub: https://github.com/Arun-Sanjay/RedButton
• Leaderboard: LEADERBOARD.md