Miserere

Anti-aliasing without oversampling. Textbook antiderivative anti-aliasing costs half a sample of delay and a cos(πf/fs) droop — fatal when four busses have to sum against a dry path. Miserere splits every curve into an exactly-aligned linear part plus a nonlinear residual and filters only the residual. The distortion products get the full treatment; the signal gets none of the smear. Flat within 0.1 dB to 20 kHz, group delay 0, and the impulse-alignment tests pass unmodified.

A limiter that is a loop, not a lookup. The CRUSH bus's sidechain is driven from its own output through a single-capacitor RC network — attack in the charge path, release in the discharge path, two fixed-point iterations per sample so the fastest settings act sub-sample at 44.1 kHz. Ratio creep, attack/release interaction, knee ordering and ALL's overshoot are emergent behaviours of that loop, each pinned by its own measurement — ALL is a genuine fifth setting, never an interpolation of the other four.

A leveler with no ratio control, because the circuit has none. SANDWICH's Opto Leveler feeds an electroluminescent-panel law into a two-state carrier-density model in double precision. The two-stage release, the memory effect — hold it longer or harder and it lets go more slowly — and the program-dependent attack all fall out of the physics: 50% recovery in 30–120 ms with the last 10% taking over 400 ms, and a measured ≥1.2× slower recovery after a ten-second hold than after half a second. The sidechain is deliberately unrectified and never decimated.

Passive EQ from component values. SANDWICH's LF boost/cut network is the ladder's exact transfer function, so the non-cancelling boost-and-cut curve, the cut corner sitting above the boost corner, and the gain/bandwidth coupling on the HF bell are consequences of the circuit rather than of chosen filter settings. The cramped filters — 12 kHz shelf, HF bell, HF shelf — are magnitude-matched, so the top octave keeps its analog shape at 44.1 kHz instead of being squeezed toward Nyquist.

A tape transport, not a filtered delay. SLAP reads its delay line with cubic Hermite interpolation, saturates on the record side ahead of the write, carries a fixed head bump, ages its spacing loss, and derives Wobble from two quasi-periodic oscillators plus a leaky random drift — measured at 0.9 Hz and 5.2 Hz, with a flutter-lag autocorrelation of 0.936 that proves it never locks into a repeating cycle. At Wobble 0 the generator is structurally off, and two renders come out bit-identical.

One tonal trade-off, kept honest. SPREAD's two crossfading taps read a single delay line at a fixed offset, so a sustained pure tone meets a comb whose depth depends on the note; broadband programme material is unaffected. Correlation-aligned splicing that removes the comb is on the roadmap, not in this release.

Honest engineering, enforced by CI. 161 test cases and 72,589 assertions on macOS and Windows, in Debug and Release: zero heap allocations on the audio thread under a replaced allocator, renders independent of host block size below −120 dBFS, a −120 dBFS bit-transparent wire null with every section off, impulse alignment held to 1e-5, and alias measurements pinned to goldens that may not regress by more than 3 dB.