STELLATE

01The point becomes weights

app/audio/targeting.js · weightsAt

The star map turns a point into a genre-weight blend: inverse-square distance over the ≤4 nearest anchors, plus a continuous snap-to-anchor pull scaled by how unambiguously nearest the closest star is — on a star ⇒ pure genre; between two ⇒ natural blend. Each genre neighborhood also gets its own key — the dominant genre's name hashed with the seed, mod 12 — so crossing into a new dominant genre is a key change.

02resolveMulti — weights into choices

engine/genre-kernel.js · resolveMulti(weights, seed)

One mulberry32(seed) stream — mulberry32 is a tiny seeded random-number generator — with a fixed draw order. Scalars (bpm, swing, humanize, fx) blend by weighted lerp (a linear blend) of each parent's [lo,hi] range, then one sample. Pools (progression, kit, form, voicing) draw one parent proportional to weight, then pick inside that parent's pool — blends combine, never smear. Grouped dimensions draw ONE parent for the whole group so a recipe and its source pool always agree.

Energy-preserving pattern inheritance: for bass and lead patterns, a parent whose pattern-pool energy is ≥1.25 rungs above the runner-up and carries weight ≥.35 takes the draw outright — a dancey parent keeps its groove against a wash-heavy neighborhood. DX7 patches sharing an FM algorithm morph by lerping their ~144-dim param vectors. Meters (3/4 waltz, compound 6/8) parent-pick and never lerp — a bar holds an integer beat count, so a blend keeps ONE parent's bar line and a journey crosses an audible meter flip, not a smear. The MUSIC-MIND axes (adventure, color, voicing, reharm, pipes, rhythm complexity) draw dead last on the same stream.

Then constrain() keeps midpoints songs: from 150 bpm up, any kit that is neither a chopped break nor a straight machine kit snaps to jungle — a declared 3/4 or 6/8 is exempt and keeps its bar line; adventure is capped at .75; a ≤2-chord progression forces reharm off; rhythm complexity is capped at .4 above 165 bpm.

03The state — a song's blueprint

engine/genre-kernel.js · toState(choice, opts)

toState applies the 8 macro sliders (acoustic, density, dust, space, bright, feel, energy, vocal — post-resolution, zero-rng, all-neutral = byte-identical no-op), then builds sections: a form graph walked by one generic interpreter, with a duration solver that lands cycle counts within ±10% of the 180 s target. Past 175 s unevolved, the 3-minute rule splits plateaus and evolves at boundaries — a key walk of +2/−3/+5, a kit or lead re-roll, or a one-time form pivot.

Sampled by default: the state ships with the SF2 sample library (108 sampled instruments, plus 114 DX7 patches), while the nine signature models — tb303, acid, reese, wobble, synclead, modeld, vocoder, hammond, hoover — stay pure synth, because a moving filter, a detune beat, a spinning Leslie or a live speech carrier is not a thing a static multisample can hold. The absent-knob law holds: if a feature wasn't chosen, the blueprint simply doesn't carry that knob — and an absent knob cannot change the output (the byte-identity law).

04buildEvents — the one score brain

engine/csd-engine.js · buildEvents(state)

One function turns the state into every note. It runs in this exact order, each pass on its own seed stream:

1. Theory / reharm — with reharm on, the named progression is only a skeleton: a T→S→D→T functional walk regenerates the chords, adventure gating diatonic → modal interchange → secondary dominants → chromatic mediants and tritone subs; minimal-motion voice-leading. stream seed+40961
2. Sections, voice by voice — per cycle per chord: pads, bass, drums, melody, countermelody — plus found-sound beds, beat-synced break slicing, one-shot hits, glitched vox, rave stabs; transitions at every boundary (fills, hat rush, snare roll, stutter, dropout, the per-genre micro lick).
3. Transform pass — a generic per-bar pass over the resolved pool (rev, ply, degrade, octflip, rest, rot, stutter); hooks left alone. stream seed+31337
4. Feel — jux stereo divergence — the left and right channels play slightly different variations; groove/swing with per-voice push-pull; thunk (soft key-noise thumps); snare ping-pong throws; tb303 accent and slide. groove on seed+777
5. Sample roles + perc lane — generalized sample placements (bed, slice, one-shot, opener, cadence, buried, response) and a percussion lane over the kit. stream seed+9091
6. Rubato — ONE smooth monotonic beat-warp applied to all four arrays; every consumer inherits the same musical clock.
7. Pipes — the resolved pipe chain (≤3 of 10: harmonize, echoCanon, strum, ghost, callResponse, densityArc, sweepArc, vibratoSwell, throwFx, octavePump) applied at the one true choke point; densityArc never drops a kick or snare. pipe i on seed+71000+i×97
8. Snare-law, dead last — a kernel-wide mandate: no bar's snare or hat signature repeats three times, measured in one rolling window across the whole timeline; machine kits vary with machine moves, human kits with ghosts, drags and drunk displacement. stream seed+5150

05mapEvents — notes into knobs

engine/faust/voices/state-engine.js · mapEvents

The event fabric becomes per-note Faust parameter sets: freq, gain and decay everywhere; the NOTE_PARAMS whitelist maps pipe annotations to real DSP sliders only where the module exposes them (cutoffMul ⇒ cutoff/tone, vib ⇒ vibrato) and silently drops the rest. Blue-note bends render only on sampler units; sampled drums emit one-shots with real tom repitch; found sounds get an auto-tune spec and a speech-rate cap for intelligibility.

06Faust WASM voices

engine/faust/voices/state-engine.js · voiceUnits

Per-voice modules are picked from the synth fleet, dx7_algN cartridges and native PCM samplers — the SF2 default — with per-voice insert chains, pool-level dry/reverb/delay/ping-pong sends, and an fx bus: zita or a per-genre reverbColor module, plus a 3-band master compressor. The synth fleet and the DX7 cartridges are Faust compiled to WebAssembly; the sampled instruments are native PCM zone playback — everything mixed on one bus, live in your browser.

07The live chord-bar scheduler

engine/faust/live/live.js · makeWalk

ONE authoritative walk stands behind both live output paths, and a gate holds its per-bar output to what buildEvents would produce for the whole song. Each call builds exactly one chord bar — it polls the current target (retarget and glide just mutate what it reads), reseeds seed + serial×7919 per bar while instrument identity stays on the song seed, fills only on a section's last cycle, sweeps only on its first and last. Two output paths: the ring/worklet path (desktop — same voice topology ⇒ glide by feeding bars into the running stream; a topology change ⇒ bridge stream and equal-power crossfade) and the WAV-first path (mobile — rolling rendered WAV segments on a real audio element, pocket-proof).