Plant a population of neural arpeggiators. Let evolution find the music.
The first generation is finding its voice.
Start Evolve garden to generate new musical variations. Step +1 runs one generation. The initial population is evaluated automatically, without producing sound.
Turn your device volume down, then press Enable audio. Select any plant, network, or archived champion to audition it. Selections and tempo changes enter on the next four-bar loop. Uncheck Follow the best organism to keep one sound in focus. Use the heart to save a favorite and include it in future generations.
Directed neural graphs evolve connection weights, biases, hidden neurons, edges, and activation functions. Musical genes evolve scale, root, melodic motifs, rhythm, arpeggio order, chord-transition associations, fifths movements, envelopes, swing, and eight oscillator harmonics. This is a NEAT-inspired hybrid, not canonical NEAT.
Genetic selection explores discrete musical and structural changes. Real backpropagation takes gradient steps on a teacher-forced musical surrogate. It does not differentiate through rounded notes, audio waveforms, or human musical taste. Improved surrogate loss does not guarantee a more beautiful song.
Current-generation elites bypass mutation and gradient training. The hall of fame preserves up to ten unique high-quality genomes in the current scoring epoch. Every generation also contributes an immutable champion to the ancestral archive; two rotating ancestral slots give earlier discoveries another chance. A finite population does not play or retain all ancestors simultaneously.
Up to four workers evaluate the population in parallel against the same three deterministic challenges: home key, transposition by a fifth, and a modal shift. Musical quality combines harmony, voice-leading, motifs, rhythm, spectral balance, and a small complexity penalty. Novelty also influences selection, separately from the musical quality score. Changing musical climate begins a new scoring epoch; old champions remain in the all-generation archive.
The audio starts off. Master attenuation starts at −26 dB and cannot exceed −16 dB. Filters, compression, a bounded non-oversampled waveshaper, voice limits, and smoothed envelopes protect the digital signal. The −16 dBFS ceiling is a sample-peak ceiling, not a certified true-peak or hearing-safety limit. A browser cannot know your amplifier, headphone sensitivity, or acoustic sound pressure. Keep device volume low. Escape silences immediately.
WAV renders the selected four-bar phrase and its tail with the same protected synthesis graph. MIDI exports four note tracks, but an external synthesizer determines its own timbre and volume. Save garden exports all genomes, archives, favorites, settings, innovations, and random-generator state. Open garden validates and restores that file, without starting playback. Local autosave is best-effort; your JSON file is the durable backup.
Native HTML5 + Web Audio. No external libraries, accounts, microphone, samples, or network requests. Browser workers fall back to cooperative main-thread evaluation when unavailable. Space toggles evolution outside form controls; Escape always silences audio.
This replaces the current session and local autosave. Save your garden first to preserve every champion.