Crypta — User Manual

Split your bass. Compress the lows. Twist the guts out of the highs.

What it is

Crypta is a parallel bass processor built for metal production, in the lineage of the reference plugin class for parallel-bus bass processing. As of v0.2.0 it splits your bass signal into three bands — low, mid, and high — with two cascaded 4th-order Linkwitz-Riley ("LR4") crossovers, keeps the low band tight with a parallel compressor, drives the mid band with staged saturation, and runs the high band through a choice of three distortion voicings before summing everything back together through a 4-band EQ and a cabinet-simulation IR loader.

Research-derived rebuild (v0.2.0)

v0.2.0 is a research-driven rewrite of v0.1.x's simpler two-band (low/high) topology, sourced against the reference plugin class's own official user manual, a third-party professional review, the same design lineage's hardware product manual, and general parallel-bus-compression community/vendor consensus — not measured against any reference plugin's actual audio output, DSP source code, or hardware unit by this project. See docs/design-brief.md and docs/research-notes.md for the full sourcing, and the same disclosure v0.1.x already carried: voicing character (drive-gain ranges, mid-filter hump/scoop settings) is engineering-tuned, not yet finalized by ear against reference material.

Where it sits in a heavy-music chain

Crypta is designed to be the bass-specific voicing stage in the Basilica Audio suite:

• Track order: DI/amp sim → Crypta → bus compression/glue → mix bus. It expects a reasonably clean, already-amp-sim'd or DI'd bass signal; it is not itself a full amp sim (no built-in preamp gain staging beyond the input trim and drive controls).
• The low band's parallel compressor is meant to keep the fundamental/sub content of the bass locked in place under a wall of distorted guitars. The mid band adds a distinct "throatier" saturation character that sits in the frequency range most likely to clash with a guitar wall — dial it in deliberately, not just as an afterthought. The high band's voicing adds the upper-mid/high "grind" that lets the bass cut through a dense mix.
• Two independent split points (Split Low, Split High) let you tune both crossover corners across the low-mid register to match the song's tuning (drop-tunings push useful low-end content further up) and to control how wide the mid band's "throat" is.
• The output stage's IR loader — now applied only to the Mid+High path, never the Low band — is meant for quick cabinet-style tone shaping without needing a separate cab-sim plugin later in the chain, though it can also be left off entirely if you're already running a dedicated cab sim elsewhere.

Signal flow

Input Trim → Gate → LR4 Split Low (60–400 Hz, default 120 Hz)
                      │
        ┌─────────────┴───────────────────────────────┐
        │                                              │
     Low band                              Remainder → LR4 Split High (300–2000 Hz, default 600 Hz)
 Parallel Comp → Graaawl → Level                         │
        │                          ┌───────────────────┴───────────────────┐
        │                       Mid band                              High band
        │                    Drive → Level          Tight → Voicing → Drive → Tone → Blend → Level
        │                          └───────────────────┬───────────────────┘
        │                                          Mid+High sum
        │                                                │
        │                                          IR loader (cab sim)
        │                                                │
        └───────────────────────┬────────────────────────┘
                                 │
                       Sum (delay-compensated)
                                 │
                            4-band EQ
                                 │
                       Safety Clip (optional)
                                 │
                               Output

The Mid and High bands share the same oversampling anti-aliasing headroom (each 4x oversampled independently, but identically configured, so they report identical latency); the low band carries a matching compensation delay, plus a phase-alignment allpass filter tied to Split High's own cutoff, so all three bands sum flat and stay time-aligned at the final sum. The IR loader (cabinet simulation) sits after the Mid+High sum and before the final three-way sum — the Low band never passes through it, matching the reference class's "low band bypasses the cabsim" architecture. See docs/architecture.md for the full technical breakdown, including exactly how the latency and phase-alignment compensation work.

Reported latency

Crypta is not zero-latency: oversampling is its source. Reported latency is a function of the sample rate alone, identical across both Drive Engines at every rate (44.1 / 48 / 96 / 192 kHz), and independent of host block size (tests/LatencyTests.cpp, "T14"). This is deliberate: the two engines can have different intrinsic latencies (Circuit's oversampling factor adapts to the host rate — 4x at or below 50 kHz, 2x at or below 100 kHz, 1x above; Classic is always 4x), so rather than re-report latency whenever you automate Drive Engine — which hosts handle poorly mid-transport — the plugin reports the maximum across both engines and pads the Circuit path up to it. Switching Drive Engine therefore never shifts your track by a sample.

One consequence worth knowing: the reported figure is the oversampling delay, not a full group-delay description. On Classic, a test impulse peaks within 1 sample of the reported latency; on Circuit it peaks within 32 samples, because the two extra IIRs Circuit's High band carries (the 10 Hz DC blocker and the drive-tracked pole) move the impulse's peak by up to roughly 25 samples (about 0.5 ms) without changing what gets reported. No single number can describe a frequency-dependent group delay - what is guaranteed instead is that the reported figure matches across engines and rates, and that the three-way band sum stays flat (≤ 1.0 dB deviation) at every split-frequency combination tested.

No absolute sample count is pinned as a constant in this manual - check your host's own plugin-delay-compensation readout for the exact figure at your session's sample rate.

Presets

Crypta ships with a preset system: a horizontal bar at the top of the plugin window lets you step through factory and user presets (< / preset name / >), save/save-as/delete your own, and import/export single presets or preset banks (zip files of multiple presets). Nine factory presets ship in v0.2.0 — see docs/presets.md for what each one demonstrates. User presets are stored per-plugin under:

• macOS: ~/Library/Audio/Presets/Yves Vogl/Crypta/
• Windows: %APPDATA%\Yves Vogl\Crypta\Presets\

A fresh instance loads a user "Default" preset if you've saved one ("Set current as default" in the preset menu), otherwise the factory "Default" preset (matching the plain parameter defaults documented below).

The interface

The plugin window is laid out in signal order, left to right and top to bottom, in ten sections: Input, Noise Gate, Crossover, Low Band, Drive Engine, Mid Band, High Band, Cabinet, EQ, Output. Whatever you see further right or further down happens later in the chain. Every one of the plugin's 54 parameters is on the front panel — nothing is hidden behind a menu.

Knobs. Drag up/down or left/right to turn; hold Shift while dragging for fine adjustment. Knobs with a fixed set of positions (Gate Mode, Low Comp Detector, Drive Engine, High Voicing) click into their detents and show the option's name. Double-click a value box to type an exact number.

Keyboard. Every control is reachable with Tab, in signal order — and Tab never gets stuck inside one section. With a knob focused:

Toggles (the lamp switches) respond to Space and Return. Ctrl/Cmd-modified arrows are left alone, so your host's own shortcuts keep working. The focused control is always marked with a gold ring.

Screen readers. Every control announces its name, its role and its current value with units ("Gate Threshold, slider, −60.00 dB"), and each section is announced as a named group. The four meters can be queried on demand and read out their current value ("Input peak level meter, −8.2 dBFS"); they deliberately do not interrupt you on every update.

Meters. Four needle meters read the metering added in v0.3.0:

The peak meters rise fast and fall slowly, so a transient is actually visible, and the scale from 0 dBFS up is engraved in red — the needle turns red there too, but the position alone already tells you. Gain-reduction needles rest at the right and swing left as reduction deepens. All metering runs on the interface thread; it costs the audio engine nothing.

Resizing. Drag the grip in the bottom-right corner. The window keeps its aspect ratio and scales between 60 % and 180 % — it is a pure zoom, so nothing re-flows or gets clipped at any size. The size you choose is saved with the session and with the plugin instance, so a reopened project comes back the way you left it. A project saved before this version opens at 100 %.

Engines (NEW in v0.3.0)

Three parameters select between the v0.2.0 DSP and its v0.3.0 replacement. A new instance boots into the new engines; any session or preset you saved before v0.3.0 keeps the old ones, so nothing you have already made changes how it sounds. Switch freely — the change is crossfaded, not stepped.

If you preferred the old sound, set the relevant engine to Classic — it is preserved exactly, not approximated.

Parameter reference

Unless noted otherwise, all continuous parameters are smoothed to avoid zipper noise when automated.

IO / Global

Noise Gate (full-band, before the crossover splits)

Sits ahead of both crossovers, so it gates the input signal as a whole rather than per band.

Gate Ratio is Classic-only. Modern is a gate with a range floor rather than a ratio expander, and ignores it.

Split Low / Split High (two cascaded crossovers, NEW topology in v0.2.0)

Split High is always kept at least a fraction of an octave above Split Low internally (a reasoned safety margin against a degenerate near-zero-width Mid band) — if you push the two close together, Split High's effective value will float slightly above whatever you've set Split Low to, rather than collapsing the Mid band to nothing.

Low band: parallel compressor + level

The low band is compressed in parallel: the compressed signal is blended back with its own uncompressed self via Mix, rather than replacing it outright, which is what keeps the low end feeling tight and controlled without ever sounding squashed or lifeless. v0.2.0 re-sources the ballistics defaults to the reference class's own fixed, sourced values — a fast, gentle "glue" bus compressor, not the heavier "New York style" squash v0.1.x's defaults implied (see docs/research-notes.md §3–4 for the full sourcing).

The gain reduction both detectors apply is published to the meters (and to any UI that asks for it) in positive dB. On Smooth RMS the figure is the detector's own gain; on Classic Peak it is estimated from the block's peak either side of the compressor, because juce::dsp::Compressor does not expose its internal gain — accurate on the steady material a meter is read against, worst-case-per-block on fast transients.

Low band: Graaawl (NEW in v0.4.0)

Graaawl puts the woody, vocal, aggressive grind of a Warwick Thumb onto your low end without making it flabby. Off by default.

It is not low-frequency distortion, and that distinction is the whole design. The Thumb's growl is an asymmetric upper-mid character around 700 Hz – 2.2 kHz, with a formant-like resonance near 1 kHz, sitting on top of a low end that stays tight. So Crypta generates that harmonic content in a parallel branch — an asymmetric saturator followed by a band-pass onto the formant window — and blends it on top of the untouched low band. Your fundamental never goes through a shaper. Measured on a 50 Hz probe with Amount at 100 %: the energy in the 20–150 Hz sub region changes by less than 0.01 dB, while the 700 Hz – 2.2 kHz window gains real signal.

It sits after the low-band compressor and before Low Level, so the compressor has already made the input level predictable and the Amount control means the same thing from note to note.

The branch is antialiased arithmetically (ADAA-1) rather than by oversampling, which is why Graaawl adds no latency at all — switching it on mid-session can never shift your track. Measured alias-to-signal on a hot 50 Hz probe at 100 % Amount: −86 dB at Tone 0, −105 dB at Tone 100.

Mid band: drive + level (NEW in v0.2.0)

A dedicated mid band with staged/cascaded saturation, structurally similar to the High band's Wool voicing (two cascaded soft-clip stages) but with no filter, tone, or blend control of its own — matching the reference class's own Mid band exactly ("Mid Drive... Mid Level" only). This band's job is a distinct "throatier" grind character, separate from the High band's own presence/fuzz/harshness-control role.

High band: Tight, voicing, drive, tone, blend, level

Three selectable distortion voicings, each 4x oversampled to keep the nonlinear shaping stage's aliasing out of the audible band.

Voicings:

All three keep their v0.2.0 placement — Gnaw flat, Wool a −6 dB/Q 0.9 scoop at 500 Hz, Razor a +5 dB/Q 1.0 hump at 900 Hz — but on the Circuit engine the nonlinearity and the filtering around it are rebuilt from circuit models rather than being three settings of one curve (src/dsp/CircuitDrive.cpp).

• Gnaw — Classic: an op-amp-style hard clip, symmetric and unforgiving, pushing hard into a square-ish waveform at high drive. Circuit: the same character, gained through a pre-emphasis shelf (1200 Hz, +6 dB) ahead of the clipper and its exact algebraic inverse behind it — the pairing collapses to unity at Drive 0 structurally, not approximately, and concentrates clipping on the upper harmonics instead of the whole band.
• Wool — Classic: cascaded soft-clip fuzz with a mid scoop and a touch of fixed asymmetry. Circuit: the asymmetric diode shunt clipper's own DC curve — two series diodes one way, one the other, solved from a small-signal-diode SPICE card by damped Newton at every table point, at prepare time, never on the audio thread — plus a dynamic bias side chain that leaves the clipper offset for roughly 20 ms after a loud passage. This makes Wool genuinely history-dependent: the same quiet probe measures at least 3 dB differently depending on what played immediately before it, against under 1.5 dB for the memoryless voicings (tests/CircuitDriveTests.cpp, "T5"). Note: the shipped behaviour is a decaying bloom after a loud passage, not a sag — the opposite sign to this feature's original design prediction. See Known limitations below; describe Wool as history-dependent / touch-sensitive, not as "sag."
• Razor — Classic: a tighter overdrive — a comparatively mild clipper with a mid-hump filter afterwards. Circuit: rebuilt as a feedback-clipper factorisation ("unity clean + clipped difference"), with the clipped-path pre-emphasis corner moved from a guitar-pedal original's 720 Hz down to 330 Hz for the bass register, and a softer-kneed tanh-fit clipping curve that better matches the modelled topology.

Gnaw and Razor also share a drive-tracked post-clip pole on Circuit: it opens to 61 kHz at Drive 0 (transparent) and slides down to 5.1–6.9 kHz at full drive, moving continuously in between rather than sitting at a single fixed corner.

Starting points, not final voicing: the drive-gain ranges and mid-filter hump/scoop settings for the Classic engine, and the Circuit voicing constants alike, are engineering defaults - tuned for musical usefulness and mathematically bounded (no runaway output at any drive setting), not yet finalized by ear against reference material. The suite's open ear-tuning gates (issues #15/#16/#17, #34) still apply to the Circuit engine specifically.

Post-sum 4-band EQ

Applied after all three bands are summed back together (and after the IR loader). Off by default; when off, the EQ stage is skipped entirely (guaranteed transparent, not just set to unity gain). v0.2.0 re-anchors the default corner frequencies to a sourced bass-tone-stack frequency set from the same design lineage as the reference class (v0.1.x's defaults were unsourced placeholders).

IR loader (cabinet simulation)

A convolution-based cab-sim stage that now processes only the Mid+High post-sum signal (relocated in v0.2.0 - it sat post-everything in v0.1.x). Off by default. With no impulse response loaded, this stage is a guaranteed bit-exact passthrough at every session sample rate, so turning it on before loading an IR never changes your sound. The Low band is structurally never passed through this stage, matching the reference class's "low band bypasses the cabsim" architecture - your fundamental/sub content stays uncolored regardless of what cab IR you load.

Loading impulse responses: the IR-loading engine and its factory-IR slot mechanics are complete — decoding an embedded WAV, installing it, and clearing back to the passthrough identity IR are all implemented and tested (getNumFactoryImpulseResponses() / loadFactoryImpulseResponse() / clearImpulseResponse()). The Cabinet section of the editor currently exposes the two parameters above and nothing else: there is no in-plugin file browser yet, and a host or wrapper that calls the plugin's IR-loading entry point works today.

What is not shipped is content: Crypta bundles no cabinet IRs, and that is a deliberate licensing decision rather than an unfinished feature. Bundling a cab IR means redistributing someone else's recording inside every copy of the plugin, so the bar is a licence that is beyond doubt — CC0, an explicit public-domain dedication, or our own capture. "Free download" and "royalty free" are not licences. Nothing has cleared that bar yet; the bar itself is enforced in code (src/dsp/FactoryIRs.h) and asserted by the test suite, so an unverified IR cannot slip into a release build. Until then, load your own IRs — the stage is a guaranteed bit-exact passthrough with nothing loaded, at every session sample rate.

State migration

v0.2.x → v0.3.0

Opening a session or loading a preset saved before v0.3.0 sets Drive Engine, Low Comp Detector and Gate Mode to their Classic values, so your saved work sounds exactly as it did. This applies to host sessions and to your own saved presets alike, including a user preset named "Default" — so if you had saved your own default, that is still what you get on a new instance.

The one deliberate exception is the safety clip: if you had it switched on, v0.3.0's clipper is not bit-identical to v0.2.0's. It aliases far less and is transparent below the ceiling; the difference is confined to material that was actually being clipped.

New parameters (High Bias, the knee and auto controls, the Modern gate's controls, Clip Ceiling) are either neutral by default or are only read when their engine is selected, which legacy state never selects.

v0.1.x → v0.2.0

If you open a Crypta v0.1.x session, the old single Crossover Frequency value is migrated to the new Split High parameter, clamped into its new 300–2000 Hz range (v0.1.x's own shipped default, 250 Hz, is below that floor, so an untouched v0.1.x session lands exactly at 300 Hz on reopen). Split Low and every new Mid-band/Tight parameter fall back to their v0.2.0 defaults. Any low-band compressor settings you had explicitly changed away from v0.1.x's old defaults are preserved as-is — only the shipped default changed, not your own deliberate settings. This is a best-effort, lossy, one-directional migration; re-check your low/mid/high balance after reopening an old session.

Under the hood

Everything below is measured in-tree and runs on every CI push (macOS + Windows, Release, .github/workflows/ci.yml, plus pluginval --strictness-level 10 on both platforms and auval -strict on macOS). See docs/architecture.md for the full derivations.

One shared oversampling region, not two

v0.2.0 ran two independent 4x oversampling instances — one inside the Mid band, one inside the High band. The Circuit engine collapses them into one: the remainder is upsampled once, split by a second Linkwitz-Riley crossover running at the oversampled rate, processed there, summed and downsampled once (src/dsp/CircuitDrive.{h,cpp}). The saved region pays for the extra per-voicing filtering. The oversampling factor adapts to the host rate — 4x at or below 50 kHz, 2x at or below 100 kHz, 1x (ADAA only) above — and that trade is tested, not assumed: 2x at 96 kHz measures at least as clean as 4x at 48 kHz, tone for tone (tests/AliasingTests.cpp, "T1: dropping to 2x at 96 kHz costs nothing measurable"). Both engines stay prepared at all times, so Drive Engine is automatable; a switch crossfades over 256 samples (5.3 ms at 48 kHz) and flushes the incoming engine's state first, because the idle engine's oversampling history otherwise releases a stale-audio burst on switch-back (measured: a 1.96 peak before the flush existed, 1.39 after, tests/CircuitDriveTests.cpp, "T18").

Antialiasing done arithmetically, on top of oversampling

src/dsp/ADAAShaper.h implements first-order antiderivative antialiasing (Parker, Zavalishin & Le Bivic, DAFx-16): closed forms where the antiderivative is elementary (tanh → ln cosh, hard clip → piecewise), and a 2048-point cubic-interpolated table with a Simpson-integrated antiderivative where it is not, so the curve and its integral stay mutually consistent — pinned directly by a test, because a mismatch would show up as a DC step under overload rather than as a small error (tests/AliasingTests.cpp, "T2"). Measured result: the Circuit engine beats the engine it replaces by 25–30 dB across a 1.2–10 kHz sweep at full drive (recorded: Circuit −81.9 / −64.0 / −57.1 / −51.9 dB vs. Classic −48.0 / −36.2 / −27.9 / −22.7 dB at 1244 / 2489 / 4978 / 9956 Hz), and clears −80 dB or better through the whole bass register (311, 622, 1244 Hz) (tests/AliasingTests.cpp, "T1").

Per-voicing circuit topologies

See the Voicings note above the parameter tables for what changed in Gnaw, Wool and Razor. The drive-tracked post-clip pole Gnaw and Razor share uses a square-law (audio-taper) mapping rather than a linear one, specifically so the pole stays above 12 kHz at half drive instead of collapsing to roughly 9 kHz and making the middle of the drive range audibly duller than the circuit being modelled (tests/CircuitDriveTests.cpp, "T3"). Every automatable scalar in the Circuit engine — drive gains, blend, bias offset, band levels, and the two one-pole filter coefficients themselves — is ramped across the block rather than held constant across it; a per-block-constant parameter measured as roughly −27 dBc of broadband non-harmonic spurs on a fast drive sweep before this existed (src/dsp/CircuitDrive.h, RampedScalar).

A low-band detector that stops the bass tremoloing

v0.2.0's low band used a peak detector with the sourced 6 ms "glue" release, and a 6 ms release follows the individual half-cycles of a bass fundamental, so the gain reduction rippled at twice the note frequency. Low Comp Detector = Smooth RMS (src/dsp/LevelDetector.h) is a log-domain RMS detector with a soft knee, program-dependent release and auto-makeup. Measured ripple on an 80 Hz tone 6 dB over threshold: over 1 dB → under 0.5 dB, while still genuinely compressing (tests/LevelDetectorTests.cpp, "T7"). The auto-makeup sign is worth knowing: −0.5·T·(1 − 1/R) with a negative threshold gives a positive gain — the easy slip (−0.5·T·(1/R − 1)) produces attenuation instead, which the test suite pins at three anchors to 0.1 dB, plus an assertion that the result is a boost at all ("T8").

A gate with hysteresis, and the test that proves it

Gate Mode = Modern (src/dsp/GateEngine.h) runs its control path per sample and adds hysteresis (0–12 dB), retriggering hold (0–500 ms), a detector-only sidechain highpass (20–400 Hz) and a dB-linear release. The cleanest proof point: a 500 Hz tone dithering ±1.5 dB around the threshold at 3 Hz for two seconds — a single-threshold gate chatters continuously through that; with a 4 dB hysteresis window this one opens once and then makes exactly zero transitions for the rest of the render (tests/GateEngineTests.cpp, "T11"). The release is fit by least squares against the ideal −range/release slope at R² > 0.99. Gate Ratio is Classic-only; Modern ignores it — it is a gate with a range floor, not a ratio expander.

A safety clip that is transparent until it isn't

The v0.2.0 safety clip was a raw per-sample std::tanh on the full mix, which lowpassed everything whenever it was merely armed. src/dsp/OutputClipper.h applies ADAA to the clipper's residual instead of the whole signal — algebraically the antialiased clipper plus an exact compensator for the droop and delay that naive ADAA would otherwise apply to material nowhere near the ceiling. Measured deviation across 40 Hz – 20 kHz while armed but not clipping: 0.13 dB (tests/OutputClipperTests.cpp, "T13"). One documented wrinkle: the compensator is a first difference and can push a sample back over the ceiling on fast material (measured 1.15 against a ceiling of 1.0), so a final hard bound is applied - it never engages below the ceiling, so transparency below it is untouched.

Nothing you already made changes

A fresh instance boots into Circuit / Smooth RMS / Modern; every pre-v0.3.0 session and preset gets the legacy engines injected, through two independent migration paths because presets never pass through setStateInformation() at all (see State migration, above). The proof is against real committed output, not just the migration code: tests/GoldenRenderTests.cpp renders four committed v0.2.0 state fixtures (Gnaw / Wool / Razor / Gnaw-with-clip) and asserts sample-exact equality (memcmp == 0) on macOS, the golden platform; on Windows the bar is ≤ −60 dB relative, with the worst of three fixtures measured at −73 dB. The fixtures are asserted to carry no stateVersion attribute, so a regenerated fixture cannot silently turn the migration test into a tautology.

Engineering hygiene

Zero heap allocations on the audio thread, on both engines, with every v0.3.0 feature live simultaneously — Circuit drive, Modern gate, Smooth RMS detector with auto-makeup, the safety clip and the EQ (tests/RobustnessTests.cpp, [realtime]). Reported latency is a function of sample rate alone (see Reported latency, above). reset() flushes every stage (crossover memory, gate/compressor envelopes, oversampling state, the latency-compensation delay line, EQ biquad history, the convolution engine), so a host transport stop/loop/rewind never leaves stale state ringing into whatever plays next (tests/ResetTests.cpp). The low band is structurally excluded from the cab-sim: its own isolated output is bit-exact identical whether the IR loader is on or off, regardless of which IR is loaded (tests/LowBandIsolationTests.cpp).

Known limitations

• Wool's dynamic bias has the opposite sign to what the original design predicted. The design brief expected a loud passage to suppress ("sag") the following quiet probe; what shipped is a decaying bloom instead. The mechanism is understood and is faithful analogue behaviour: the bias makes the clipping asymmetric, asymmetric clipping generates real DC, and the 10 Hz blocker downstream then restores it over its own ~16 ms constant, so the probe rides a bloom rather than a dip. History-dependence itself is confirmed as intended (11 dB on Wool against 1 dB on the memoryless voicings). This is flagged for the suite's open ear-tuning gate (issues #15/#16/#17, #34) - describe Wool as history-dependent / touch-sensitive, not as "sag."
• The Circuit voicing constants remain engineering-derived starting points, not yet finalized by ear against reference material - the same disclosure v0.1.x and v0.2.0 already carried for the drive-gain ranges and character-filter settings.
• The alias floor is stated per tone, not as one flat −80 dB bar, and the reason is arithmetic. Gnaw is a 40x hard clip whose harmonic series falls off as 1/n with no bandwidth limit, so the floor is set by which harmonic order folds back into the band, and that order drops as the fundamental rises. Raising the engine to 8x oversampling was measured and still misses −80 dB at the upper tones, while doubling the stage's cost. Delivered instead: 25–30 dB better than the previous engine on every tone (against a 10 dB brief requirement), −80 dB or better through the bass register, and a −49 dB in-band floor everywhere.
• The two engines are not identical above 3 kHz at drive 0, and this is not claimed. Parity holds to ±0.5 dB up to 3 kHz; above it Circuit is up to 2.5 dB brighter (measured +0.5 dB at 8 kHz, +2.5 dB at 14 kHz), because its tone lowpass runs at the oversampled rate and escapes the bilinear frequency warping the base-rate Classic filter has.
• The drive-tracked pole opens to 61 kHz at drive 0, not the 24 kHz the original design brief called for - a one-pole at 24 kHz is already about −1.9 dB down at 18 kHz and would not have met this feature's own transparency requirement, while 61 kHz (−0.36 dB at 18 kHz) is also the figure the circuit research gives for the real thing.
• CPU is a design intent here, not a benchmark. Collapsing two oversampling regions into one is what funds the extra per-voicing filtering, and the release notes state the trade as landing at roughly the CPU v0.2.0 spent. No CPU measurement ships in the test suite - do not put a percentage or an "x% lighter" figure anywhere.
• The safety clip trades its antialiasing advantage for its ceiling guarantee at extreme overdrive. The delta-form compensator can push a sample back over the ceiling on fast material, so a final hard bound is applied; it never engages below the ceiling, but roughly 10 dB or more past it the antialiasing advantage goes away. Deliberate priority ordering - a safety clip that lets material through is not a safety clip, and heavy clipping belongs to the drive stages, which are oversampled and ADAA'd for exactly that.
• One deliberate departure from v0.2.0's output: the engaged safety clip. If you had it switched on, v0.3.0 is not bit-identical - it aliases far less and is transparent below the ceiling. The difference is confined to material that was actually being clipped and is bounded (measured −26.5 dB relative on a fixture driven 12 dB past the ceiling). Everything else about a pre-v0.3.0 session or preset is sample-exact.
• Cross-toolchain bit-exactness is unattainable and is not claimed. macOS is the golden platform; on Windows the bar is −60 dB relative, with the worst of three fixtures measured at −73 dB. The drift is not last-ulp noise - the gate and the low-band compressor both take decisions off a detector level, so a 1-ulp difference near a threshold can shift a transition by a sample.
• The GUI is a functional generic editor plus a plain labelled meter readout row. The photoreal M3 GUI is a later milestone and will consume the same MeterTaps struct; the preset bar is a plain functional strip.
• No bundled factory cabinet IRs, by decision. The slot mechanics, the decoder and the licence guard all ship and are tested; the asset table is empty because no IR has been sourced with a licence worth staking a redistributed binary on. The in-plugin file browser is the GUI milestone's work. The convolution engine is a guaranteed bit-exact passthrough with nothing loaded, at every session sample rate.
• Graaawl's voicing constants are engineering-derived starting points. The shaper drive and asymmetry, the formant window and its resonance are chosen from what the growl demonstrably is, and every claim about them in this manual is a measurement — but the final voicing is an ear decision against a real Thumb 5 DI, and that gate (issue #36) is still open.
• Graaawl's sub protection is band-limiting, not magic. The branch's own harmonics leak through the formant highpass's stopband: on a pure 50 Hz sine the third harmonic at 149 Hz sits about 51 dB below the fundamental. In aggregate the 20–150 Hz region moves by under 0.01 dB, and real bass material carries its own third harmonic some 30 dB louder than that, so this is inaudible — but it is not zero, and it is not claimed to be.
• Deliberately out of scope for v0.3.0, tracked openly: factory IRs / IR browser / IR trim-align; the stereo strategy (low mono-sum toggle, Mid/High width); full per-sample Newton DK circuit simulation (v0.3.0 ships the calibrated factorised models; a full simulation is a possible later "HQ Circuit" mode); a lookahead gate and time-varying auto-release; linear-phase / HQ-offline oversampling modes and a shared suite-level oversampler module.
• The voicing is research-derived, never measured against hardware or against any reference product's audio, DSP source, or unit. Where a model is a deliberate simplification, the docs say so.
• Pre-1.0 and AGPLv3 - breaking changes possible until v1.0.0. The v0.1.1 rename (plugin code Cryp, bundle id com.yvesvogl.crypta) means DAWs treat this as a new plugin relative to v0.1.0-era sessions.

Tips

• Start with the low band tight, then dial in the mid band's throat, then the high band's grind. Set Low Comp Mix and Makeup first so the fundamental feels locked in, then dial in Mid Drive for the "throatier" cutting character, and only then pick a High voicing and drive amount.
• Split Low and Split High are tone decisions, not just technical ones. Pushing Split Low up moves more note body out of the (compressed-only) Low band; pushing Split High up widens the Mid band's own passband, giving the "throatier" character more room before the High band's own fuzz/presence character takes over.
• High Tight is your main "fuzz vs. tightness" control, independent of which voicing you've picked - pull it toward its 20 Hz floor for maximum fuzz, push it up toward 500 Hz for a tighter, more controlled top end. It also tames harshness on hot Drive settings.
• High Blend is your "how much" knob, High Drive is your "how hard" knob. If a voicing feels too extreme, try lowering Blend before lowering Drive - you'll often keep more of the character that way, just at a lower overall intensity, rather than flattening the nonlinearity itself.
• Graaawl is a character layer, not a distortion amount. Start at 30–40 % with Tone near the middle and listen for the vowel rather than for grit; if it starts sounding like a fuzz on the bottom end, you have gone past what the Thumb actually does. Because it sits after the compressor, changing Low Comp Threshold changes how hard the growl is driven — set the compressor first.
• Leave the safety clip off during tracking/mixing, and only reach for it as insurance against unexpected automation or a hot input on a specific pass - it's a safety net, not part of the intended tone-shaping signal path.