Circuit modelled gain stages, from subtle preamplifier saturation to a scooped high gain sound, and the rest of a guitar rig behind them: a pedal in front, a power stage, a speaker that loads it, a cabinet and up to two microphones. CLAP and VST3, built with nih-plug.
Nothing here is a filter shaped to sound like an amplifier. Every circuit is a netlist — resistors, capacitors, inductors, valves, transistors, diodes, op-amps, transformers — solved by modified nodal analysis, with Newton–Raphson wherever a part is not linear. The sound comes out of the topology, so a capacitor in the gain leg makes the mid-hump because that is what it does in the hardware, not because someone drew the curve.
PEDAL -> CIRCUIT (preamp) -> POWER AMP <-> SPEAKER LOAD -> CONE -> CABINET -> MIC A / MIC B -> OUTPUT
Each stage is chosen on its own. A pedal can sit in front of any circuit, any circuit can drive any power stage, and any power stage can drive any speaker in any cabinet. The arrow between the power amp and the speaker goes both ways: the speaker's impedance is part of the power stage's own netlist, so its resonance and rising voice-coil impedance change what the valves, the output transformer and the negative feedback do. It is not an EQ after the fact.
Every new stage defaults to what the plugin did before it existed — no pedal, Matched power stage, Legacy cabinet — so old sessions and presets load and sound as they did.
Six numbered bands in the order the signal goes through them, with an arrow at the foot of each pointing into the next. 780 × 912 at its own size, and the button in the strip scales it from 75 % to 150 % — it is drawn rather than pictured, so it is sharp at any of them. The model lists are dropdowns, and the wheel over a closed one steps through it, which is the quickest way to hear what is in it.
| 1 Input | Trim, meter, pedal with drive/tone/level | The meter reads against the level the circuits were voiced at. Its zero is where the rest of the panel means what it says. |
| 2 Circuit | Topology or modelled circuit, clipping, amplifier, iron, power amp | What does the work. Clipping applies to the diode circuits, the amplifier choice to the preamplifier channels, iron to everything. Lists that do not apply grey out rather than vanish. |
| 3 Drive | Drive, master, and the Cali IIC+'s five-band graphic | All the way up is the sound the circuit is named for. Down from there only cleans up. |
| 4 Tone | Stack, bass, mid, treble; reverb, speed, intensity | A passive stack, so it only ever cuts. A modelled circuit's own tone controls take these knobs. The second row belongs to the Twin. |
| 5 Cabinet | Cabinet, speaker, mic A, mic B, placement, blend, polarity, time | Legacy keeps the old baked cabinet filter; any other cabinet switches to the physical path. |
| 6 Output | Mix, level | The dry path is delayed to match, so mixing is a mix and not a comb filter. |
Bypass is the host's: the plugin reports a bypass parameter, so the DAW's own bypass switch and automation lane drive it.
| Circuit | What it is | Measured at full drive |
|---|---|---|
| Clean | One valve stage barely working | 3 % distortion, almost all second harmonic |
| Crunch | Two stages, the second driven by the first | 15 % |
| High Gain | Three stages, all clipping on every note | 40 % |
| Overdrive | Diodes across the feedback resistor | 20 %, and it cleans up when hit harder |
| Distortion | Diodes across the signal to ground | 38 %, a ceiling the output stops at |
| Console | A step-up transformer into a discrete stage | 5 %, essentially all second harmonic |
| Studio | An op-amp on a studio rail | 0.00 % across the band — see below |
The two clipper families are not a matter of degree. In the loop, the diodes lower the gain, so the output never stops following the input — which is why an overdrive on a hot signal barely does anything. To ground, they are a ceiling: past it the waveform is squared off and the harmonics fill out to the top of the band.
Particular circuits, built part by part from their schematics. Each has a
research log in docs/models recording the revision, the
sources, where they disagree and what was approximated.
| Circuit | Built from | Power stage when Matched |
|---|---|---|
| Green 808 | A 1970s green overdrive pedal: input buffer, clipping amp, tone and level, output buffer | none |
| Ram Fuzz | A 1973 four-transistor fuzz | none |
| Cali IIC+ | A 1980s Californian lead channel: six triodes through its lead return and recovery stage, its own stack and five-band graphic | Cali 6L6 |
| American 5150 | A 1990s high-gain lead channel: six triodes, one run cold | American 6L6 High-Gain |
| British 73 | A transformer-coupled class-A microphone preamplifier and line driver | none (its line driver stays) |
| American 312 | A console microphone preamplifier card: input transformer, one discrete op-amp, output transformer | none |
| British 4K E | A 1980s console channel's microphone input: a 1:10 transformer into two op-amps around one gain pot | none |
| Tube 610 | A 1960s valve console channel: four triodes in two feedback loops, a transformer at each end | none |
| American Twin | A 1960s blackface clean channel with spring reverb and optical tremolo | American 6L6 Clean |
| Brit 800 | An early-80s British 100 W master-volume lead preamp: four triodes, a cathode follower into its own stack | Brit EL34 |
| Brit Plexi | A late-60s British 100 W lead amp with no master: the bright channel's three triodes into its own stack, so the Volume decides how hard the power valves work | Brit Plexi EL34 |
| Brit AC30 | A 60s British 30 W combo: a top-boost valve into a stack with no middle control, then self-biased power valves with no feedback loop at all | AC30 EL84 |
| Brit DR103 | A British 100 W head built for headroom: five triodes, a master volume, and an inverter the driver holds still so it cannot shift its bias | DR103 EL34 |
| Cali Rectifier | A 90s American two-channel head, red channel: five triodes with one run cold on 39 k, which is where its bottom end is squared off | Recto 6L6 (or Recto 6L6 Tube) |
| Green 9 | The green overdrive with the later pedal's output resistors | none |
| Rodent | A hard-clipping distortion whose slow op-amp runs out of gain-bandwidth before it runs out of gain | none |
| Round Fuzz | Two germanium transistors and a feedback resistor | none |
| Yellow Dist | One slow op-amp and a pair of germanium diodes to ground | none |
| Heavy Metal | A gated distortion: two germanium diodes in series with the signal, and three gyrators | none |
| Metal Zone | Two gain stages and seven filters, with a three-band equaliser whose middle sweeps | none |
Every pedal is also a circuit. The eight in the pedal slot can each be selected on their own, with nothing behind them — which is how a pedal into a desk was always recorded, and how an HM-2 into an MT-2 becomes expressible: one in the slot, the other as the circuit. The list is grouped under PEDALS, AMPLIFIERS and MICROPHONE PREAMPS, because it now holds three different kinds of thing.
Display names are generic on purpose. Which hardware each one was built from is in What each name is below, and in the research logs alongside the stable ids saved sessions use.
A microphone preamplifier is not a guitar preamplifier turned down. A guitar preamplifier is supposed to run out of room — that is where the sound is — and these are built so that they do not, so everything interesting happens in the last few decibels before they do, or in the transformer rather than the gain stage.
The amplifier list picks what does the work on those channels: a valve, a JFET, or an op-amp. A JFET follows a square law, and a square law has only a second-order term, so it makes second harmonic and essentially nothing else until it runs out of room. An op-amp with enough loop gain contributes nothing of its own at all — measured, 0.00 % across the whole band at every gain setting.
The iron list puts an output transformer after any circuit here, and it is the part of this plugin that a waveshaper cannot be. What saturates in a transformer is the flux, and flux is the integral of voltage — so the same signal makes far more of it low down. Measured on a core at twelve volts:
| 30 Hz | 120 Hz | 500 Hz |
|---|---|---|
| 68 % | 4 % | 0.00 % |
No curve applied to the samples behaves that way. It is why the studio channel, which has no character whatever of its own, reads 13.97 % at 40 Hz with iron on it and 0.000 % without. The core is symmetric, so it makes odd harmonics where a single-ended valve stage makes even ones.
Three materials, and they are not one another turned up: nickel is the only one that colours a quiet signal and the softest when pushed, steel stays out of the way then arrives hard and goes furthest, amorphous is still clean at eight volts where steel is at 40 %.
Silicon, germanium and LED are selectable wherever there are diodes. They differ by forward voltage, which is the whole reason anyone swaps them: measured, germanium is already bending at 20 mV where an LED has not started, and the LED overtakes it when driven hard.
Green 808, Ram Fuzz, Green 9 (the 808 with the later pedal's output resistors), Rodent (a hard-clipping distortion whose slow op-amp runs out of gain-bandwidth and slew rate, as the original's does), Round Fuzz (two germanium transistors), Yellow Dist (one slow op-amp and a pair of germanium diodes to ground), Heavy Metal (a gated distortion with two tone controls) and Metal Zone (two gain stages and seven filters, with a three-band equaliser whose middle sweeps), each with the knobs that pedal actually has, in front of whichever circuit is selected — so a Green 808 into the American Twin keeps both sets of controls. The pedal is its own netlist, solved before the circuit it feeds, so a pedal can also sit in front of itself.
A pedal's knobs are the ones that pedal has, named the way its box names them: most have drive, tone and level, the Heavy Metal has two tone controls (its Colour Mix pair), the Metal Zone has four including a swept mid, and the panel grows a knob for each. The Heavy Metal is also the only gated distortion here -- two germanium diodes sit in series with the signal rather than across it, so quiet playing is held back entirely and the pedal cuts off rather than fading.
The slot is level matched the way the circuit list is. Each pedal's Level knob at noon is unity through that pedal — its own measured position, a little above half on all of them, because the level control of a pedal is an attenuator after a stage with gain. So switching pedals with the knobs where they are moves the level by a few decibels rather than by ten, and what is left is the pedals being different pedals: a fuzz at half its drive really does make more than a clean boost does. And the pedal is fed a guitar whatever follows it, while the circuit behind it is fed the level it was calibrated at — a guitar for the amplifiers, up to a volt for the clean stage and the consoles — so a pedal in front of the clean circuit sounds like a pedal rather than like a fault.
Nominal, 90 %, 80 % or 70 %. A valve amplifier's supplies all come from one transformer, so running it from a variac brings every rail down together: the valves have less room, the bias follows them, and the amplifier goes soft and compressed at a setting where it used to be loud and clean. It is a selection rather than a knob because that is what it was — how the rig was wired, not something anyone swept while playing. The Brown '78 and Brown '84 presets use it, because the engineer who recorded those records says the amplifier ran at 80 to 85 volts.
| Choice | What it is |
|---|---|
| Matched | The selected amplifier's own power stage; none for pedals, topologies and the studio preamplifiers |
| Bypass | No power stage: phase inverter, output valves, feedback and output transformer all skipped |
| Cali 6L6 | Two 6L6s, long-tailed-pair inverter, feedback and presence |
| American 6L6 Clean | Four 6L6s with a 12AT7 inverter and light feedback — the clean one |
| American 6L6 High-Gain | Four 6L6s, feedback and presence |
| Brit EL34 | Four EL34s from a 1981 British 100 W master-volume drawing; the Brit 800's own |
| Brit Plexi EL34 | Four EL34s from a 1970 drawing of the non-master head: no master, four times the 2203's feedback, a 5 k presence; the Brit Plexi's own |
| AC30 EL84 | Four EL84s sharing a 50 ohm cathode resistor, no feedback loop, and a cut control across the inverter; the Brit AC30's own |
| DR103 EL34 | Four EL34s on 22 k grid stoppers and a tight loop, with the inverter direct-coupled to the preamplifier; the Brit DR103's own |
| Recto 6L6 | Four 6L6s on a cold -51 V bias from the manufacturer's own drawing; the Cali Rectifier's own |
| Recto 6L6 Tube | The same stage with its rectifier switch on valve: two 5U4GB, so the rail sits lower and sags under a chord |
A power stage is a complete netlist: master, inverter, bias, grid coupling, output valves with their screen supplies, a centre-tapped transformer with a nonlinear core, and the feedback loop around all of it. Choosing one behind a different circuit is level matched against that circuit's own, so trying combinations does not mean chasing the fader.
Cabinet: Legacy (the old resistor load and Combo/Stack filter, the default), Bypass (the driver on an open baffle), or one of ten cabinets with real dimensions and driver layouts: Brit 1960 4x12, Cali Oversized 4x12, Brit Closed 4x12, Brit Green 4x12, Brit V30 4x12, Oversized 4x12, American Open 2x12, American Open 1x12, Closed 1x12, Closed 2x12.
Bypass there means no box, not no speaker: the driver still radiates, on an open baffle, and a microphone still picks it up, which is why it does not sound like Legacy with its filter off. The way past the acoustic path is the speaker row's own Bypass, and that is bit-identical to Legacy/Off (a test says so).
Speaker: Matched (the cabinet's own driver), Bypass (the power stage's output, a DI), or Brit V30, Brit Green 25, Brit T75, American Vintage 12, American Vintage 10, American Ceramic, American Alnico. Each is a Thiele–Small electrical and mechanical model with a lossy voice coil, fitted to the manufacturer's impedance and response data where it is published and estimated where it is not, and its impedance is solved inside the power stage. A closed box stiffens the suspension in that same solve, so a 4x12 moves the resonance the valves see.
Microphones: eleven profiles — dynamic, ribbon, condenser, valve condenser and FET condenser — each with its own polar pattern, proximity effect, response and off-axis loss. Mic A is always there (its Bypass is an ideal omni); mic B can be off. Each has position (centre to edge of the cone), distance and angle, placed in three dimensions against every driver in the cabinet: close up the nearest cone dominates, further back the array adds up at low frequencies and beams at high ones, an open back cancels its own bass, and two microphones at different distances comb-filter unless time is set to Aligned. B polarity inverts mic B, and blend mixes the two.
The models, the equations and what each number is based on are in
SPEAKER_MODEL.md, CABINET_MODEL.md
and MICROPHONE_MODEL.md.
The panel's names are generic; the hardware behind them is not. This is the whole mapping, with the research log for each — the log is where the revision, the sources, the places they disagree and everything approximated are written down. Every model is built from published drawings and data sheets, part by part; a name in the right-hand column is a statement about what was measured against, not a claim of any connection with the manufacturer. All trademarks belong to their owners, who have neither endorsed nor been involved in this.
Circuits — the ones marked generic are designed topologies with no hardware behind them at all.
| Panel | Modelled from | Log |
|---|---|---|
| Clean / Crunch / High Gain | generic: one, two and three cascaded ECC83 stages | — |
| Overdrive / Distortion | generic: diodes in the feedback loop, and diodes to ground | — |
| Console / Studio | generic: a step-up transformer into a discrete stage; an op-amp on studio rails | — |
| Green 808 | Ibanez TS808 Tube Screamer (JRC4558D, 1N4148 pair, BC549 buffers) | green_808.md |
| Ram Fuzz | Electro-Harmonix Big Muff Pi, 1973 Ram's Head | big_muff.md |
| Cali IIC+ | Mesa/Boogie Mark IIC+ lead channel, 60 W | cali_iic_plus.md |
| American 5150 | Peavey EVH 5150 lead channel | american_5150.md |
| American Twin | Fender Twin Reverb AB763 (blackface), vibrato channel | american_twin.md |
| Brit 800 | Marshall JCM800 2203 100 W, 1981 drawing | brit_800.md |
| Brit Plexi | Marshall 1959 Super Lead 100 W, bright channel, 1970 drawing | brit_plexi.md |
| Brit AC30 | Vox AC30/6 Top Boost, brilliant channel | brit_ac30.md |
| Brit DR103 | Hiwatt DR103 Custom 100, brilliant channel, Issue 4 | brit_dr103.md |
| Cali Rectifier | Mesa/Boogie Dual Rectifier, two-channel Rev F, red channel modern | cali_rectifier.md |
| British 73 | DIY Recording Equipment 73P v1.1 (a 1073-style 500-series card) | project-local drawings |
| American 312 | API 312 microphone preamplifier card (2622, 2520, 2503) | american_312.md |
| British 4K E | SSL SL 4000 E channel amp, 82E01 microphone amplifier | british_4k_e.md |
| Tube 610 | Universal Audio 610-A modular console preamplifier | tube_610.md |
| Green 9, Rodent, Round Fuzz, Yellow Dist, Heavy Metal, Metal Zone | the pedals above, selectable as circuits in their own right | as listed under Pedals |
Pedals (the slot in front of whatever is selected)
| Panel | Modelled from | Log |
|---|---|---|
| Green 808 | Ibanez TS808 Tube Screamer | green_808.md |
| Green 9 | Ibanez TS9 — the 808 with its output resistors | green_9.md |
| Ram Fuzz | Electro-Harmonix Big Muff Pi, 1973 Ram's Head | big_muff.md |
| Rodent | Pro Co RAT, the LM308 revision | rodent.md |
| Yellow Dist | MXR Distortion+, the 741 version | yellow_dist.md |
| Round Fuzz | Arbiter Fuzz Face, germanium (1966–68) | round_fuzz.md |
| Heavy Metal | Boss HM-2 Heavy Metal, the Japanese original | heavy_metal.md |
| Metal Zone | Boss MT-2 Metal Zone | metal_zone.md |
Power stages
| Panel | Modelled from | Log |
|---|---|---|
| Cali 6L6 | Mark IIC+ 60 W: ECC83 inverter, two 6L6GC, feedback and presence | cali_iic_plus.md |
| American 6L6 Clean | Twin Reverb AB763: ECC81 inverter, four 6L6GC, 820 Ω feedback | american_twin.md |
| American 6L6 High-Gain | 5150: ECC83 inverter, four 6L6GC, 39 k feedback | american_5150.md |
| Brit EL34 | JCM800 2203: four EL34 at −42 V, feedback from the 4 Ω tap | brit_el34.md |
| Brit Plexi EL34 | 1959 Super Lead: the same iron with no master, 47 k feedback | brit_plexi.md |
| AC30 EL84 | AC30 Top Boost: four cathode-biased EL84, no feedback loop, cut control, GZ34 rectifier | brit_ac30.md |
| DR103 EL34 | Hiwatt DR103: four EL34, inverter direct-coupled to its driver | brit_dr103.md |
| Recto 6L6 / Recto 6L6 Tube | Dual Rectifier: four 6L6 on the silicon setting, or two 5U4GB valve rectifiers | cali_rectifier.md |
Cabinets, speakers and microphones — dimensions from the manufacturers, Thiele–Small parameters and response curves from data sheets where they are published and estimated where they are not (cabinets.md, speakers.md, microphones.md).
| Cabinet | Modelled from | Speaker | Modelled from |
|---|---|---|---|
| Brit 1960 4x12 | Marshall 1960A, angled | Brit V30 | Celestion Vintage 30 |
| Brit Closed 4x12 | Marshall 1960B, straight | Brit Green 25 | Celestion G12M-25 Greenback |
| Brit Green 4x12 | Marshall 1960AX | Brit T75 | Celestion G12T-75 |
| Brit V30 4x12 | Marshall 1960AV | American Vintage 12 | Jensen P12R |
| Cali Oversized 4x12 | Mesa/Boogie Rectifier Standard | American Vintage 10 | Jensen P10R |
| American Open 2x12 | Fender Twin Reverb combo | American Ceramic | Jensen C12N |
| American Open 1x12 | Fender '65 Deluxe Reverb | American Alnico | Jensen P12N |
| Closed 1x12 | Marshall 1912 | ||
| Closed 2x12 | Marshall 1936 | ||
| Oversized 4x12 | generic | ||
| Combo / Stack (Legacy) | generic filters, kept for old sessions |
| Microphone | Modelled from | Microphone | Modelled from |
|---|---|---|---|
| Dynamic 57 | Shure SM57 | Ribbon 38 | Coles 4038 |
| Dynamic 421 | Sennheiser MD 421 II | Condenser 87 | Neumann U 87 Ai |
| Dynamic 906 | Sennheiser e 906 | Condenser 414 | AKG C414 |
| Dynamic 409 | Sennheiser MD 409 U3 | Tube Condenser 67 | Neumann U 67 |
| Ribbon 121 | Royer R-121 | FET Condenser 47 | Neumann U 47 fet |
| Ribbon 160 | beyerdynamic M 160 |
Everything else on the panel — Diode, Amplifier, Iron, Tone, Mains — is generic:
a material or a device type rather than a particular product.
docs/MODEL_INVENTORY.md carries the same mapping
with the stable ids saved sessions use, and what is approximated in each.
One channel at 48 kHz, as a fraction of the time available, with each preset set up exactly as it ships — pedal, power stage, cabinet and microphones included.
| Studio Preamp | Valve Colour | Blues Crunch | Scooped Metal | British 73 Pre | Plexi Crunch | Brit Crunch | Blackface Clean | Blackout '80 | Experienced '67 | Ultra Rhythm | Boutique Lead | Texas Storm '83 | Puppet Master '86 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 3.0 % | 4.7 % | 9.3 % | 13.7 % | 17.9 % | 20.2 % | 20.3 % | 20.8 % | 20.8 % | 22.2 % | 30.6 % | 31.5 % | 33.0 % | 37.7 % |
The heaviest is Puppet Master '86 at 38 %: the Cali IIC+'s six triodes and graphic, a
speaker-loaded power stage and two microphones. The microphone preamplifiers are among
the cheapest (2.8 to 11 %). Measured with cargo run --release --example presetcost.
Swapping the Legacy cabinet for a physical one moves an amplifier's cost by a
few points either way. On the same settings, the Cali IIC+ went from 20 to
22.6 % and the Twin from 15 to 17 %, while the 5150 dropped from 33 to 30 %
because its power stage needs fewer passes into a real speaker. The cabinet and
microphone stage itself costs about half a percent with one microphone and
three quarters with two (examples/acousticcost.rs).
The Quality control is part of the sound rather than only of the cost, and the presets set it. Measured against an eight times reference at 3 kHz, a diode clipper to ground aliases 31 dB down at two times and 71 at four, so the clipper presets ask for four; a three stage valve cascade is already 55 dB down at two, and it is the cascade that costs anything, so paying for four there buys nothing audible.
The modelled circuits are the ones that alias most -- a cold-clipped high-gain
channel at 48 kHz puts 18.6 % of its fundamental back as hash that is not the
note -- and they are also the most expensive to solve. They follow the control
as far as 2x, which takes about two thirds of that away for about double
the work, and stop there: at four times every one of them costs more than the
time there is, which is a DAW missing its deadline rather than a cleaner sound.
The row shows the factors it can deliver and lights the one in use. Every
shipped preset on a modelled circuit asks for host rate, so none of them costs
more than it did. Measured with cargo run --release --example oversampling.
Seventy-three, ordered quietest first within each group so the list reads as a
range: Studio, Preamp, Crunch, High Gain, Overdrive, Distortion, Amplifier, and
five groups of chains aimed at particular records — Classic Rock,
Psychedelic / Lead, Alternative, Metal / Heavy and Blues. The Studio group holds the console and microphone-preamplifier
sounds, with the cabinet off throughout — a guitar speaker in front of a
microphone preamplifier makes no sense at all, and a test enforces it. The
whole catalogue is level matched, and a test holds it to that; cargo run --release --example presetlevel prints where each one sits.
Each one is a full set of panel positions — loading one and looking at the panel
tells you how the sound is made. Every guitar sound comes out of a physical cabinet,
speaker and microphone setup; the old baked cabinet filter is still there for older
sessions, but no shipped preset uses it. The record-inspired presets are built only
from parts the plugin actually has; which parts of each rig are documented and
which are approximated is written down in PRESETS.md.
Save your own with the button beside the name; they go to
~/.config/gainstagefx/presets as readable JSON and appear under their own
Saved heading at the foot of the list. Only those can be deleted — saving
under a shipped preset's name writes a new file beside it rather than replacing
it, so the shipped one never becomes unreachable. A dot next to the name means
the panel has been moved since the preset was loaded; move the control back and
the dot goes, because that state is compared rather than remembered.
Loading a preset goes out as ordinary parameter gestures, so it lands in a DAW's automation lanes and undoes in one step like any other edit. Saved presets store stable model ids as well as positions.
Built archives for Linux, macOS and Windows are attached to each
release, built from the
tag by GitHub Actions. Download the one for your platform and run the installer
inside it — install.sh, Install.command or install.exe. README.txt in
the archive covers installing by hand.
cargo xtask bundle gainstagefx --release # CLAP and VST3 in target/bundled
cargo test --release # the measurements
cargo run --release --features standalone # the panel without a DAW
./install.sh # into the usual placesEvery claim in this README that has a number in it is checked by a test that
measures it. cargo test --release runs 355 of them, at 44.1, 48,
88.2, 96 and 192 kHz where the rate matters, and the audio path is tested not
to allocate.
src/dsp is the engine and knows nothing about audio plugins:
netlist— circuits as named nodes, validated before use. A dangling node or an unreachable output is aFault, not a silent wrong answer. Parts a control moves (pots, and a speaker's electrical equivalent) are adjustable without rebuilding the circuit.ac— stamps the matrix with complex admittances and solves per frequency. Exact, in microseconds, no audio involved. Refuses to run on a circuit containing a device, because a device has no single frequency response.time— trapezoidal companion models, Newton–Raphson with junction voltage limiting, and a separate DC matrix for the operating point with capacitors open and inductors shorted. Linear parts of a circuit are reduced out once, so each sample only solves what is actually nonlinear.device— the nonlinear parts: Koren triodes, pentodes with screen dependence, Shockley diodes, bipolar and JFET transistors, an op-amp, and a transformer core whose flux integrates. Parts that impose a voltage rather than a conductance get a branch unknown of their own.measure— bin-aligned probe tones with phase continuous across the settle boundary.
src/circuits holds the netlists: the topologies, the pedals, the preamplifier
channels and power, which builds every power stage with either a resistor or a
speaker as its load. src/acoustics holds the speaker profiles, the cabinets,
the microphones and the stage that places them. src/voice.rs wires the chain
together.
The two solvers overlap deliberately. On a linear network they must agree, and
tests/agreement.rs is the only real check either of them has.
More: ARCHITECTURE.md, DSP.md,
MODELS.md, docs/MODEL_INVENTORY.md,
IMPLEMENTATION_PROGRESS.md.
The discipline here is that a number nothing can disagree with is not a measurement. Several times the thing that turned out to be wrong was the measurement, and several times it was me:
- A linear network read 2.6 % distortion because the probe tone's phase restarted between settling and capture. Nothing was wrong with the circuit.
- A gain control moved the Clean voice 6 dB end to end, on the reasoning that a gain control sets how much local feedback each stage keeps. It does not: a cathode bypass is bounded by the cathode resistor. It is a volume pot now, and the range is 80 dB.
- A linear rheostat cannot sweep evenly in decibels — for a span of 48 it put 18 of its 30 dB into the first quarter of the travel and 2 into the last. The track has to be shaped for the span it covers.
- The oversampler handed the circuit four samples for every one the host sent while the circuit still solved with the host's timestep, putting every corner frequency four times too high. It cost the overdrive 11 dB and the valve voices almost nothing, so nothing but the clipper would have shown it.
- An operating point converged, satisfied its own tolerance, and was wrong, because the matrix was stale. Kirchhoff caught it; convergence could not.
Device::advancesat on the trait uncalled from the beginning and nothing noticed, because every device until the transformer core was memoryless. The core integrates, and unadvanced its flux never accumulated past one sample: it drew a hundred-thousandth of the current it should have and read 0.00 % at every frequency, level and material.- A channel's gain control sat between the gain stage and the output transformer, which looks like a reasonable place for it. A pot's wiper impedance is highest in the middle of its travel, and a transformer colours according to what feeds it, so the channel measured 18 % distortion at half gain and 5 % at full — the knob inverted the character as it crossed the middle.
- The overdrive pedal's tone control fell more than three decibels short of what its own drawing gives, because two resistors had been swapped on the strength of a reference file this repository never had. Three independent sources agreed on the original values, and the tone reach is now a test.
- The first speaker load made a feedback power stage oscillate at 96 and 192 kHz: a pure inductance for the voice coil turned the feedback into a resonator above the band. A lossy coil fitted to the measured impedance, and the output transformer's own winding capacitance, fixed it. Every power stage is now driven into clipping into three speakers at the lowest and highest rates in a test.
- The Californian amplifier's graphic equaliser was wired as five tracks from the signal to ground with a fixed make-up after it. It could cut twelve decibels and boost less than three, so every V came out as a dip in the middle with nothing added at the ends. Both drawings put the tracks between the two inputs of a feedback amplifier; built that way it cuts and boosts by the same amount, and centred it is flat.
- The Californian lead channel stopped at its lead output and handed that straight to the equaliser, which left out the two triodes both of its drawings send the lead signal back through. One of them clips on every note. With them, the channel measures 61 % distortion at a guitar's level instead of 39 %, and a preset aimed at a famously saturated record stopped sounding clean.
- A power stage's master stayed where another circuit had left it, so the Twin on its own matched power stage came out 27 dB down with only the reverb audible. A test now switches the Twin's power stage behind three other circuits and back.
GPL-3.0-or-later — the full text is in LICENSE.
The VST3 bindings this links are GPL, so the whole is. Every crate it links is
listed with its own licence and copyright notice in
THIRD-PARTY-NOTICES.md; regenerate that with
python3 tools/third-party-notices.py.
VST is a trademark of Steinberg Media Technologies GmbH, registered in Europe and other countries. CLAP is a trademark of its respective owners. Neither this project nor its authors are affiliated with or endorsed by them.