Crypta

One oversampling region, not two. The previous engine ran two independent 4x oversamplers — one for the mid band, one for the high band. The Circuit engine upsamples once, runs the second Linkwitz-Riley split at the oversampled rate, processes both bands there, sums and downsamples once. The region that gets saved is what pays for the per-voicing pre/de-emphasis networks and the drive-tracked post-clip pole. The factor adapts to the host: 4x up to 50 kHz, 2x up to 100 kHz — and dropping to 2x at 96 kHz is measured to be at least as clean as 4x at 48 kHz, tone for tone.

Antialiasing done arithmetically, on top of oversampling. First-order antiderivative antialiasing with closed forms where the antiderivative is elementary — tanh becomes ln cosh — and a 2048-point cubic table with a Simpson-integrated antiderivative where it is not, so the curve and its integral stay mutually consistent. The ill-conditioned branch and the out-of-range extrapolation are each pinned by their own test, because a drifting fallback would make slow material subtly wrong while fast material stayed right. Measured 25–30 dB less aliasing than the engine it replaces on every tone, and −80 dB or better through the whole bass register.

Three clippers, three circuits. The hard-clip voicing gets a pre-emphasis shelf and its exact algebraic inverse behind the clipper, so drive 0 collapses to unity structurally rather than approximately. The fuzz voicing is an asymmetric diode network's own DC curve, solved by damped Newton at every table point at load time — plus a dynamic bias side chain that leaves the clipper offset for about 20 ms after a loud passage, with the shaper's own image of that offset subtracted so what survives is the change in the curve's local slope and not a thump. It is measurably not memoryless: the same quiet probe comes out at least 3 dB different depending on what was played 5 ms earlier, against under 1.5 dB on the memoryless voicings — described here as history-dependent and touch-sensitive, not as hardware-style "sag", because what the measurement shows is a decaying bloom rather than a dip.

A pole that moves with the drive control. The feedback clipper's compensation pole slides down as drive opens — post-smoothing inside the nonlinearity, which a static waveshaper misses entirely. Measured: down to 5.1–6.9 kHz at full drive, above 12 kHz at half drive, monotone in between, and out of the audio band entirely at zero.

A gate with hysteresis, and the test that proves it. Drive a tone dithering ±1.5 dB around the threshold at 3 Hz for two seconds. A single-threshold gate chatters continuously. With a 4 dB hysteresis window this one opens once and then makes exactly zero transitions for the rest of the render. Add retriggering hold so it survives 16ths at 120 BPM, a detector-only sidechain highpass so a ringing low string cannot hold it open, and a release that is a straight line in dB — fitted to better than R² = 0.99 against the ideal slope, so it closes in the time the control says instead of asymptotically never arriving.

A safety clip that is a wire until it isn't. Naive antiderivative antialiasing of a clipping curve degenerates, on sub-ceiling material, into a two-tap average — about −8.3 dB at 18 kHz, applied to the whole mix whenever the clip is merely armed. This one applies the treatment to the clipper's residual instead, which is algebraically the antialiased clipper plus an exact compensator for that droop and delay. Measured deviation across 40 Hz – 20 kHz: 0.13 dB. Past roughly 10 dB over the ceiling the safety clip's final hard bound engages on most samples and the antialiasing advantage goes away — a deliberate priority: a safety clip that lets some peaks through is not a safety clip.

Latency is not zero, by design. Crypta oversamples, so it reports a positive latency to the host — on the order of a millisecond or less — and relies on host plugin delay compensation. The reported figure is a function of sample rate alone, the maximum across both engines, specifically so that switching Drive Engine never re-reports latency mid-transport.

Nothing you already made changes. A fresh instance boots into the new engines; every session and every preset saved before this release gets the old ones injected, through two independent migration paths — because presets never pass through the session-state path at all. The proof is committed pre-release renders that the migrated processor has to reproduce sample-exactly on macOS, with the fixtures themselves asserted to predate the new schema so the test cannot quietly stop testing anything.

Honest engineering, enforced by CI. 152 test cases across 30 files on macOS and Windows, plus pluginval at strictness 10 and auval -strict: zero heap allocations on the audio thread on both engines with every feature live, renders independent of host block slicing, reported latency identical across engines at every sample rate, a bit-exact bypass, a low band that is structurally excluded from the cabinet-sim path, and a three-band sum flat to ±0.1 dB — including exactly at the minimum-split-gap boundary.