Experiments

Lenia: What a Cellular Automaton Cannot Do

The Substrate Ladder (V2–V6)what each variant added, and whether it helped+ content-based kernels+ directed motion+ memory channels+ boundary sensingV2: Lenia EvolutionCurriculum > complexity~V3: Content CouplingFoundation substrate✓V4: ChemotaxisMotility selectable~V5: Temporal MemoryMemory selectable, Φ-stress 2×✓V6: BoundaryBEST: rob 0.969, max 1.651✓V5 to V6: a real membrane wins; the plasticity and signaling variants, since removed, did not help

V2: The Evolution Series

Lenia with resource depletion and regeneration, evolved under drought. The measure is the change in partition-loss Φ\intinfo under severe drought.

VersionΔΦ\Delta\intinfo (drought)Lesson
(naive)−6.2%-6.2\%Decomposition baseline
(homogeneous evolution)−6.0%-6.0\%Selection alone does nothing
(heterogeneous chemistry)−3.8%-3.8\%Diverse viability manifolds help a little
(hierarchical coupling)−9.3%-9.3\% evolved vs +6.2%+6.2\% naiveStress overfitting: more integrated and more fragile
(curriculum)+1.2 to +2.7 points on novel stressorsThe only intervention that generalized

The shifts are small and reported without error bars. What held up: mild stress raised integration and severe stress collapsed it in every condition, most likely by pruning; evolution produced fragile integration; curriculum beat substrate complexity; and the substrate has a ceiling that no amount of resource dependence moves.

Study record: metadata, result path, seeds, source code, and key finding for .

V3–V6: The Substrate Ladder

Seven substrates, each adding one capability; four remain in the catalog. An attention variant, since removed, added evolvable attention and took the fraction of cycles with integration rising under stress from 3 to 42 percent, which stopped at the threshold rather than crossing it, and which later work showed was not the operative variable. added content-based coupling and became the foundation for the measurement passes. added chemotaxis; patterns forage. added memory channels; evolution kept longer memory in two of three seeds, the one substrate addition consistently selected for. A Hebbian-plasticity variant, removed, was the worst substrate: the extra degrees of freedom added noise faster than selection could filter it. A quorum-signaling variant, removed, had two of three seeds evolve to suppress signaling, and its record robustness came at a population of two. added a real membrane and had the best robustness, with internal gain evolving down in all three seeds. Evolution preferred permeable membranes to insulated cores.

VersionMean robustnessMaxCycles above 1.0
(content coupling)0.9231.0523/90
(+ chemotaxis)~0.91~0.95~1/90
(+ memory)0.9071.0703/90
(boundary)0.9691.651~10/90
The Substrate Ladder: mean robustness across V2–V6
The substrate ladder. Mean robustness, Φstress/Φbase\Phi_{\text{stress}} / \Phi_{\text{base}}, across substrates. Stars mark cycles above 1.0. More mechanisms did not mean better results.
V6 internal gain, boundary width, and insulation field evolution
parameter evolution. Internal gain falls in all three seeds; boundary width rises. External sensing was worth more than internal rumination.

Every one of these stayed at ρsync≈0.003\rho_{\text{sync}} \approx 0.003. The FFT kernel mixes each cell's contribution into a global field every step, so nothing a pattern does survives to be re-sensed. There is no medium to close the loop through, internal or external. Within Lenia the architectural space was exhausted.

Study record: metadata, result path, seeds, source code, and key finding for .

Study record: metadata, result path, seeds, source code, and key finding for .

The Measurement Program on V3

Thirteen measurement passes over snapshots, three seeds, seven snapshots each, fifty recording steps per snapshot. Most of them are small, and the table says which.

MeasurementResult
languageMutual information above baseline in 15 of 20 snapshots; topographic correlation zero everywhere. A chemical commons, not a language.
counterfactual detachmentNull. ρsync≈0\rho_{\text{sync}} \approx 0 from cycle zero.
self-modelNull, with one pattern in one seed at one cycle above threshold.
affect geometryStructure-behavior RSA significant in 8 of 19 snapshots, rising in one seed. Same form as the alignment that collapsed under rotation-invariant re-test in . Suspected basis artifact; not re-tested.
normativityNull. No valence asymmetry between cooperative and competitive contexts. Social context raised integration regardless of interaction type.

Eight further passes, substrate engineering, existence, world model, representation, perceptual mode, social-scale integration, entanglement, and a capstone roll-up, were removed from the catalog in November 2026 as uninformative. Their code and results are in git history before commit 111a046.

Study record: metadata, result path, seeds, source code, and key finding for .