Four lo-fi layers drawing on one shared voice pool.

Single, Layered, Round-Robin, Random-Robin and MIDI-Split. Three key split points, per-layer enable, level and pan, and channel-tagged routing carried on a per-note bitmask.

The important part is what they share. Voice stealing is global, so a four-layer stack and a four-zone split both draw from the same eight voices, and the instrument runs out of polyphony the way the hardware does.

Each layer is a full voice: two multi-mode oscillators with PolyBLEP band-limited saw and square, a sub, a ring modulator, a ladder filter in LPF24, LPF12, HPF12 or BPF12, three envelopes, tanh drive and glide.

Then Dirt, a per-layer noise character mixed in ahead of the filter, in five flavours: Static, Crackle, Geiger, Click and Burst. It is the difference between a clean digital layer and one that sounds like it came off a table.

The mangle section is where it breaks properly. Five drive waveshapers (PNP, Tube, Pickup, Diode, Crunch), bit crush from 1 to 16 bits, sample-rate reduction up to 64 times, and a free-running stutter that does not care about your grid.

Enzyme passes auval on both architectures, reports correct tail and latency, smooths automation, and guards its output bus with a DC blocker and a soft clipper above a 0.9 knee. Parameter IDs are frozen on an L{0..3} scheme, so presets and automation survive future versions.

It deliberately has no step sequencer. Layer targeting was ambiguous, the serialization risk was real, and the arpeggiator turned out to be the cleaner and higher-value half of that idea.

What is still open is honest: an eight-slot-per-layer modulation matrix, a settings page, and per-parameter smoothing inside the voice, which is audible today on fast cutoff automation.