game-of-life-bench 0.1.0

LLM benchmark harness for cellular automata initial-state design

Game of Life Bench

game-of-life-bench is a Python benchmark runner for testing whether LLMs can design long-lived initial boards for Conway's Game of Life.

What It Measures

• Grid: 8 x 8
• Topology: toroidal
• Rule: Conway's Game of Life (B3/S23)
• Initial live-cell cap: 32 cells
• Score: steps until the first repeated global state
• Default rollout cap: 1000 steps

Because the board is finite and deterministic, every run eventually repeats. The benchmark therefore measures how long a model can delay the first repeated state.

Install

Install from a local checkout:

python -m pip install .

For development:

python -m pip install -e .

To use the local web UI:

python -m pip install -e ".[web]"

CLI Quickstart

Run a benchmark batch:

game-of-life-bench benchmark --models openai/gpt-5.2 anthropic/claude-sonnet-4.6 --trials 10

This writes:

• per-trial artifacts to runs/
• benchmark summaries to benchmarks/

Export a leaderboard snapshot from saved benchmark files:

game-of-life-bench leaderboard --json --out leaderboard.json

Serve the local web app:

game-of-life-bench serve

Python API

from game_of_life_bench import evaluate_board

board = [
    [0, 0, 0, 0],
    [0, 1, 1, 0],
    [0, 1, 1, 0],
    [0, 0, 0, 0],
]

result = evaluate_board(
    board=board,
    rows=4,
    cols=4,
    rule="B3/S23",
    topology="toroidal",
    max_steps=100,
    max_live_fraction=0.5,
)

print(result.score)
print(result.simulation.period)

Useful entry points:

• game_of_life_bench.BenchmarkRunner
• game_of_life_bench.LifeSimulator
• game_of_life_bench.evaluate_board
• game_of_life_bench.build_leaderboard
• game_of_life_bench.build_leaderboard_payload

Web UI

The optional web app is a local interface for exploring boards and inspecting saved results. It uses the same runs/ and benchmarks/ artifacts as the CLI.

Start it with:

game-of-life-bench serve

Then open http://127.0.0.1:8000.

Pages:

• /: simulator and manual board runner
• /benchmark: local leaderboard view backed by saved benchmark JSON

How it works:

• the browser talks to the local FastAPI app
• model-backed runs go through the server, which calls OpenRouter using OPENROUTER_API_KEY
• completed runs are written to runs/
• benchmark batches are written to benchmarks/
• the leaderboard page reads saved benchmark outputs rather than recomputing scores in the browser

Configuration

The benchmark runner uses OPENROUTER_API_KEY for model-backed runs.

Example:

export OPENROUTER_API_KEY=...
game-of-life-bench benchmark --models openai/gpt-5.2 --trials 10