Lenia Swarm
Flow Lenia: perturbation experiments that test whether spatially localized patterns cohere, recover, and show goal-directedness, and cartography of the genotype-phenotype morphospace they live in.
Flow Lenia generalizes cellular automata to continuous state, time, and space. Two functions define its dynamics: a kernel K for neighborhood sensing and a growth function G for updates over time. Every persistent structure emerges from those two, from gliders and walkers up to self-replicators. The flow variant we use conserves mass for these spatially localized patterns (SLPs) and allows local, position-dependent rules, which is what makes multi-species evolution possible.
Because every rule is known and cheap to rerun, this is a clean place to study morphogenesis and emergence. We grow a creature, perturb it, and watch whether it holds together, the same competency-under-stress question Levin's TAME framing asks of living matter. And because a creature's genotype is just its rule parameters, we can map the whole space of forms instead of studying one creature at a time.
The perturbation harness matches control-versus-intervention pairs across two environments, flat and high-theta, with three families. init_state re-seeds the initial mass placement, flow_regime halves the integration timestep, and param_noise re-rolls the growth function, holding architecture fixed but resampling rule parameters. Recovery falls off across them: init_state perturbations leave the trajectory intact in nearly all cases, flow_regime recovers in some, and param_noise in few.
A related question, prompted by Cool et al. 2026, is whether a creature's position in the morphospace predicts how perturbation couples to its heading. Three of their four Chan-catalog creatures swerve around occluded regions that carry no sensory information while the fourth dies on contact, and they find the difference is hardwired per creature without saying which rules cause it. Our first replay pilot exposed a cohort problem: centroid displacement alone admitted diffuse and weakly moving patterns. We are replacing it with an assay that qualifies spatial localization, persistence, and locomotion before testing obstacle response.
The morphospace work measures the structure of Flow Lenia's genotype-to-phenotype map and compares its synthetic shapes with developmental and biological data. Flow Lenia specimens, developmental-model snapshots, and biological outlines all enter the same descriptor space. Topological data analysis finds loops. Local cohort comparison identifies Lenia neighborhoods that lie closer to one biological cohort. Transport follows a loop and checks whether the measured shape and the rule state both return.
Repository path: dossiers/lenia-swarmFive temporally consistent, coherent movers completed fifteen matched obstacle encounters. All retained their matter, but their motion and organization changed in different ways: bending, arrest, reversal, and loss of coherence.
The same twelve measurements find persistent gaps in 25,167 Flow Lenia specimens, 859 EmbryoMaker snapshots, and 232 fish outlines. One Lenia region selected without fish data is unusually fish-near. Rare closed paths also retain rule-state differences, giving the obstacle assay a concrete prediction problem.
Our experiments begin while a seeded field is still reorganizing, then follow what happens as broad steerability narrows and the response to an intervention becomes more specific to one developmental history. The synthesis asks whether the first clear sign of an organism may therefore appear in its changing possibilities before it is obvious in its shape.
The anatomical compiler begins with a body and searches backward for Flow Lenia rules that can sustain it. The report follows the inverse from a 5,000-form fiber study through conditional proposals, simulation-in-the-loop refinement, and an explicit form demonstration, while keeping the failed simple landscape story and the shape-versus-behavior boundary visible.
The compendium is the materialized morphospace. 13,390 indexed creatures, 7,664 stable under standard initial conditions, spread across 698 species in 8 families. Each row ties a genotype to its phenotype, provenance, and a replay config.
The system is a controller-worker topology over a SWIM gossip cluster, with an MLX Swift engine for GPU-accelerated Flow Lenia across Apple machines, split into three components:
LeniaCore: physics engine, distributed actors, search infrastructure.LeniaCLI: headless workflows underdiscover,orchestrate,index,analyze,intervene,publish.LeniaStudio: macOS SwiftUI app for local exploration, host mode, and worker mode.