Each note is a pluck burst convolved with an impulse response synthesised for that exact pitch. The burst is half a millisecond to three and a half of low-passed noise, dominated by a leading impulse, much shorter and brighter than any mallet in this roster, because a release is not a contact: the tine leaves the nail, and the only thing limiting the bandwidth is how cleanly it slips off. The impulse response is a bank of exponentially-decaying sinusoids at the clamped-free beam ratios 1, 6.267, 17.55 and 34.39, with the upper modes down 10 to 26 dB and gone within tens of milliseconds. Convolving one with the other is what excites them, which is why the Pluck control changes the balance of partials rather than only the brightness.
The music box's twin is a second, detuned copy of the whole bank at 1.0035 times each mode, about six cents, a beat every second or so at C6. Rendering both and letting them interfere is real interference between two detuned partials rather than a chorus applied afterwards, the same mechanism the handpan's ding uses for its octave shoulder and modal-ir.ts uses for a gamelan's ombak. It is added at every mode rather than only at the fundamental, because the shared comb root answers the tooth's whole bank.
The buzz layer is rendered from the note's own decay rather than mixed in beside it. Bounce rate and bounce amplitude both follow the fundamental's exponential, so a loud board throws the bead further and sooner and a fading one stops throwing it at all, the rattle dies at twelve per cent of the initial swing, well before the note does. Each bounce is a short ping at a pitch tied to the note, roughened with seeded noise so no two are the same object hitting the same spot. The layer rides the note's own amp gain, so damping the tine stops the rattle exactly as a thumb on the board does. At exactly zero no buffer is rendered, no PRNG draw is taken and no node is built.
The mechanism tick is the one thing here that precedes its note. A cylinder pin drags a tooth aside and then lets go, so the drag, the release click and the gear-train clatter belong twelve milliseconds before the pitch, which is why raising the Mechanism control delays the note rather than adding noise on top of it, and why setting it to exactly zero removes that delay along with the sound. A kalimba has no mechanism at all, so the control is inert on that family by construction rather than by taste.
Which note a key plays is decided before any of that, by the board's tine list. A MIDI note either has a tine or does not; there is no nearest-note substitution, because substituting would turn a seventeen-note diatonic instrument into a chromatic one with duplicates, which is precisely what a kalimba is not. An unmapped key allocates nothing and reports that it made no sound, and the face acknowledges the strike so a silent key reads as a board with no tine there rather than as a bug.
Every random sample in the path, the pluck burst, the rattle, the mechanism, comes from the SDK's seeded generator, forked by the cache key that asked for it, so a given setting always produces the same buffer and a cache hit is the same audio as a fresh render. Impulse responses are cached per family, note and the parameters that change the samples, and dropped wholesale when one of those moves. Voices are capped at twelve with oldest-first stealing, and a voice's nodes are freed on an eight-second tail cap with a short fade, or as soon as a note-off's release has run, whichever comes first.
