Agility, Not Throughput
Abstract. Rust gain is 0.21 ns/frame; Sixth is 8.46. Biquad 5.31 vs 37.30. The midside anomaly is 110×, likely a cache bug. None of that is the thesis. Sixth compiles in 0.02 s (GCC 48 ms/binary). A thousand-stage self-generating pipeline: 30 ms versus GCC’s 50 seconds. Generate eight sound variants in 0.16 seconds while audio plays. TPT SVF replaces the biquad. Gesture → sound in under 3 ms.
The engine truth
Per-op, Sixth loses. The compiler is 533 KB; output dylibs 33–84 KB. There is no NEON auto-vectorization (Forth is scalar). fsqrt/fpow/flog/fexp are unexposed; only tanh-approx exists, and 12+ plugins need fpow. 58% of S3 code is stack shuffling. E1 crashes on three of four DSP benches (fover/fdup inside +loop → SIGILL/SIGSEGV). X23 pretends to hold the third stack item. Top-level variable init does not persist into the binary.
W1 compiles flat ARM64 for Hypervisor.framework: fault isolation at 500 ns IPC per trap. ws_audio_vcpu.dylib auto-patches RET → HVC; no source fork between dylib and vCPU.
TPT SVF over biquad
Biquad Direct Form I/II: five coefficients, two states, one output type, static per buffer, cannot self-oscillate cleanly, cramps near Nyquist, clicks under audio-rate FM. TPT SVF: 14 NEON f32 instructions, LP/BP/HP simultaneous, unconditionally stable, clean self-oscillation at Q=∞, per-sample coefficient updates, physically meaningful states (capacitor voltages). Forty percent more work for three times the outputs. Moog ladder (40–60 ops with tanh) and diode ladder (50 ops, asymmetric sat, 303 character) compose from the same TPT integrator. Fast tan for audio-rate FM: x + x³/3 + 2x⁵/15, 7 ops, <0.1% error for fc < 0.4·fs.
What makes producers not go back
Not features. Paradigm: gesture is automation (trackpad drag morphs the wavetable while Metal deforms at 120 fps). Instant variation at compile speed. Effects are 2 KB Forth files, not $200 binaries. Progressive disclosure from tap-to-sound to patching ARM64. One unified canvas — the Forth graph already is a node system; arrangement vs session is a historical accident Wavesmith should refuse. Scene transitions are graph swaps: compile the next graph while the current plays, crossfade at audio rate, 0.02 s hot-swap.
The path is not more features. It is deeper sound, then character, then intelligence, then completeness — in that order.
Source report. Summarized from docs/SYNTHESIS-002.md in the hive tree. This page is the paper. The report remains the primary.