forces
Cohesion, separation, alignment, curl, attraction, and bilateral symmetry negotiate each particle's next movement.
signal / live study
NEURALRULES
What if the model evolved the generator instead of generating the image?
Breed a small particle ecology. Choose what survives. The visible work is always produced by deterministic browser rules; an optional local language model can only nudge a bounded genome of forces such as cohesion, curl, branching, diffusion, and decay.
choose an evolutionary pressure · breed · select one offspring to survive
generation 00 / seed organism
genomegeneration 00
- cohesion
- 50
- separation
- 50
- alignment
- 50
- curl
- 50
- attraction
- 50
- branching
- 50
- diffusion
- 50
- decay
- 50
- symmetry
- 50
selection pressure
not a prompt / a constraint on mutation
offspring
three bounded procedural variations
Neural mutation is optional. Enabling it allows the browser to download SmolLM2-135M-Instruct and run it locally through the shared Transformers.js worker. The model receives only nine normalized rule values, four coarse behavior measurements, and the selected pressure. Its free-form text is never executed or displayed: only recognized mutation tokens from a fixed whitelist survive. There is no server inference.
under the surface
The genome is small. The behavior is not.
material
Branching, diffusion, and decay alter how stable paths become, how often trajectories diverge, and how long traces remain legible.
phenotype
The browser measures speed, spread, clustering, and directional order. Those are observations of behavior, not model judgments.
selection
Your choice is the fitness function. The system does not decide which child is better; it only generates constrained alternatives.
system
The model can suggest a mutation. It cannot author the runtime.
- simulation
- Canvas 2D / deterministic particle ecology
- genome
- 9 normalized bounded rule parameters
- selection
- human choice / one survivor per generation
- optional model
- HuggingFaceTB/SmolLM2-135M-Instruct
- preferred model path
- WebGPU / q4f16 / current weight about 118 MB
- fallback model path
- WASM / q8 / current weight about 137 MB
- server inference
- none
- status
- experimental / mobile-first procedural baseline
The model interface is intentionally narrower than text generation. It may return only two to four recognized trait-direction tokens. Invalid, contradictory, or missing suggestions fall back to the same deterministic mutation system that is available before any model download.