The wave decides. The particle obeys.

Melodies are sampled from a small neural network (30→20→15, tanh) running forward passes live in this tab, blended with an order-1 Markov prior learned from its own playing. Every bloom you trigger is a reward signal — a reinforcement nudge on the weights, so the melody bends toward what you touch. Every 16 bars the weights mutate and the generation counter ticks. Nothing is prerecorded. The model that plays for you exists only here, only now.

Nine synthesized voices — pads, bass, plucks, FM bells, lead, and a full drum engine — are composed by a stochastic conductor: weighted section grammar, Euclidean rhythms, voice-led harmony over five modes, pivot-chord and chromatic-mediant modulations, drifting tempo and swing. All synthesis is raw Web Audio. Zero samples. Zero libraries.

Five generative scenes render in raw WebGL2 with a video-feedback core. Every note the engine plays is passed to the GPU as an event — position, hue, magnitude — at its exact scheduled time, so light and sound share one clock. The page's accent color is the current key, mapped through the circle of fifths: when the music modulates, the whole world changes color.

dunl.app · instrument 01 · seed in URL — share it, same universe, different life