[ How it's built ]

The terrain is a line-mesh grid: 96 frequency bins across, 160 rows of scrolling history receding to the horizon. Each animation frame reads the AnalyserNode (2048-point FFT), maps bins on a log-ish curve so the bass band stays wide the way the ear hears it, applies attack-fast / release-slow smoothing, and pushes a new row onto the ring buffer.

// history recedes: shift everything one row back, write the new frame at row 0
history.copyWithin(COLS, 0, history.length - COLS);
history.set(values, 0);

The sound is pure WebAudio, no samples: a generative dark-ambient drone of detuned sine pairs a fifth apart through a lowpass filter wobbled by a slow LFO, plus a noise wash. Mic mode is analysis-only — the input is never routed to the speakers, so there's no feedback. Dragging across the terrain plucks it: raycast position maps to a minor-pentatonic triangle-oscillator note, so everything you play is in key.

// position across the terrain → pentatonic note over ~2.5 octaves
const semis = 12 * Math.floor(steps / PENTA.length) + PENTA[steps % PENTA.length];

Sound, made visible.

SOUND & CODE — EXHIBIT TWO

DRAG THE TERRAIN — IT PLAYS BACK