Lenia Swarm

1. Read the Lenia docs map.
2. Open the Compendium Inspector to browse published creatures. For schema, taxonomy, and artifact details, use the schema and taxonomy docs.

• March 1, 2026
  Lenia cabinet docs rebuilt with current contracts, internals, and universe lanes.
• February 3, 2026
  Taxonomy/ecology/morphometrics synthesis reflected in docs and generated analysis notes.
• January 20, 2026
  IMGEP and activity-tracking surfaces were documented in generated implementation analysis.

Flow Lenia generalizes cellular automata to continuous state, time, and space. Two functions define the dynamics: a kernel K for neighborhood sensing and a growth function G for updates. Every persistent structure emerges from those two, from gliders and walkers up to self-replicators. The flow variant adds mass-conserving transport, so a creature stays a coherent object rather than a wave of growth and decay. We use it to search parameter space and index what stays alive.

Under perturbation, a Flow Lenia creature either returns to its baseline trajectory or settles into a different attractor. Each trial runs matched control-versus-intervention pairs across two environments, flat and high-theta, with three perturbation families. init_state re-seeds the initial mass placement. flow_regime halves the integration timestep. param_noise re-rolls the growth function, holding architecture fixed but resampling rule parameters.

Recovery is graded across the families. init_state perturbations leave the trajectory intact in nearly all cases. flow_regime perturbations recover in some. param_noise perturbations recover in few, and the failed recoveries sit far from baseline.

A separate open question, prompted by Cool et al. 2026, is whether position in the morphospace predicts coupling between perturbation and heading change, the property that lets three of their four Chan-catalog creatures swerve around occluded regions that carry no sensory information. They leave this open at four creatures. The compendium gives us a population to put the question to.

When a Lenia creature runs into a patch of the world it cannot sense, some creatures turn away and some die on contact. A recent Tufts paper finds the difference is hardwired into each creature, and cannot say from the rules alone which is which. We have thousands of creatures with their rules attached, so we can ask whether the swervers cluster on our map.

The morphospace track is separate from the creature-discovery track. Here the question is not whether every specimen looks beautiful; it is whether Flow Lenia's genotype-to-phenotype map has measurable topology, how that synthetic cloud sits near developmental and biological shape clouds, and where loop transport leaves residue.

The approach is deliberately comparative. We put Flow Lenia specimens, developmental model snapshots, and biological outline cohorts into a shared descriptor layer so the same questions can be asked of each cloud: where loops appear, which neighborhoods sit near external shape families, and which regions deserve denser follow-up.

Topological data analysis gives the first pass at loop structure. Local cohort comparison asks whether particular Lenia neighborhoods lean toward developmental or biological shapes. Transport then walks around those neighborhoods and checks whether the underlying state returns with the phenotype, or whether hidden residue remains in the rule layer.

Controller-worker topology over a SWIM gossip cluster, with an MLX Swift engine for GPU-accelerated Flow Lenia across Apple machines.

• LeniaCore: physics engine, distributed actors, search infrastructure.
• LeniaCLI: headless workflows under discover, orchestrate, index, analyze, intervene, publish.
• LeniaStudio: macOS SwiftUI app for local exploration, host mode, and worker mode.

For deeper internals, see the cabinet docs.