autoresearch-gollyx-python

Autonomous optimization of the gollyx-python two-color toroidal Game of Life simulator, using the Karpathy autoresearch pattern.

How it works

An LLM agent iteratively modifies program.py to reduce wall-clock simulation time, validated against known-good checkpoint data at every step.

File	Role
program.py	Self-contained ToroidalGOL simulator. Only file the agent modifies.
bench.py	Benchmark harness + correctness validation. Immutable.
program.md	Prompt that guides the LLM agent.
results.tsv	Experiment log (appended automatically).

Running a benchmark

python bench.py
# => PASS  1000 generations in 180.00s
# => FAIL  at generation 5  (if correctness breaks)

Correctness

Checkpoints are hardcoded in bench.py and validated at generations 0, 1, 2, 3, 4, 5, 60, 100, 200, 300, 400, 500, 600, 700, 800, 900, 1000, 2000, 3000, 5000 — earliest first, so broken changes fail fast.