LifeManifold

Explore cellular-automata “worlds” as points in rule space — simulate them, search diverse high-quality niches with MAP-Elites, optionally guide generation with an LLM + surrogate, and inspect archives in a Streamlit UI.

Core loop: WorldSpec (rules + parameters) → run_world → behavioral metrics → archive or metrics JSONL.

Requires Python ≥ 3.11 and uv.

Install

make install                 # creates .venv and syncs dependencies
source .venv/bin/activate    # optional; or prefix commands with uv run

Dashboard extras are included in make install. To sync only the dashboard group later: make install-dashboard.

Quickstart (MAP-Elites)

Primary path — illuminate a behavior archive from a scheduler YAML:

uv run python -m worldspace --illuminator mapelites \
  --scheduler worldspace/specs/map_elites_scheduler_mini.yaml \
  --output-dir artifacts/map_elites_smoke \
  --seed 42 --steps 200 --grid 8

Writes artifacts/map_elites_smoke/map_elites_archive.jsonl. Same smoke as CI:

make smoke-map-elites

Dashboard

streamlit run dashboard/Home.py

Default scan picks up the smoke archive above. Paths and discovery: dashboard/config/config.yaml — see docs/DASHBOARD.md.

Batch exploration (legacy)

Generate and simulate a fixed set of worlds (PCA / k-means layout, not the MAP-Elites archive):

uv run python -m worldspace \
  --generator random \
  --worlds 30 \
  --steps 200 \
  --grid 40 \
  --metrics-trace results/metrics.jsonl

Documentation

Start here
Architecture	Package map, two CLI paths
MAP-Elites	Illuminator, archives, schedulers
WorldSpace	Simulator, WorldSpec, generators
Surrogate	Hints, buffer, acquisition (shadow / filter)
Dashboard	Streamlit setup and pages

Also
Domains	Maze / dungeon / sphere runners (not main CLI)
Formulas	Metrics and fitness definitions
Protocol freeze timeline	Lock / extension / reporting ledger (not in the journal PDF)

License

MIT — © 2026 Artem Starovoitov