This is the design of the instrument, fixed before the build that the eight weeks then realize one layer at a time. It sets out what Eidolon is trying to sound like, the hardware that taught it, and the reasoning that ties the two together. Where a weekly report documents one milestone, this page documents the intent behind all of them.
The thesis: analog life is deliberate imperfection
A vintage synthesizer sounds alive because every stage of its signal path, from oscillator and mixer to filter, amplifier, and output, adds a little nonlinearity, noise, and cross-coupling that a clean digital path removes. The usual way to sound "analog" in software is to run a clean signal into one master saturation stage at the end. That is why so much virtual-analog sounds sterile: the character is concentrated in a single place instead of accumulated across the whole chain.
Eidolon takes the opposite position. The design goal is not one convincing nonlinearity but dozens of small, always-on imperfections woven through every sample of every voice, and every one of them has a documented physical cause in real hardware. This is a white-box instrument: there is no machine-learning model of analog behavior, only specific circuits, measured and reconstructed. Three commitments follow from this thesis, and they organize everything below. Character comes from distributed imperfection, not a single effect. The imperfection is always on, not a switch. And the parameter set exists because each control has a documented sonic reason, not to fill a feature grid.
Hardware lineage: what each instrument taught
The design is a deliberate synthesis of specific hardware. Each contributes one clear lesson.
The ARP 2600 is the reference for coloration spread across every stage. Its amplifier is a bare CA3080 transconductance amplifier with no linearizing diodes, so it runs the full hyperbolic-tangent nonlinearity: gentle odd-harmonic distortion that grows with level and tips into even harmonics when a small offset breaks its symmetry. Its path is direct-coupled, so envelopes leave a faint thump; its amplifier leaks a little even when closed, so the instrument is never truly silent; its filter tracks the keyboard fully by default. None of these is a defect to remove. Together they are much of what alive means, and each becomes a modeled behavior in Eidolon.
The Soundcraft 600 console, designed by Douglas Self around 1989, is the reference for the master bus. Transformerless, its coloration is gentle and entirely electronic. Its defining trait is that distortion rises with frequency, very low at 1 kHz and about four times higher at 20 kHz, because the amplifier's feedback falls off with frequency, not by any deliberate voicing. That rising signature is the fingerprint of a real console bus, and Eidolon's output stage reproduces it rather than a generic soft clip.
The Sequential Prophet-5 teaches how analog voices differ from one another. Its 2020 reissue shipped a Vintage knob for the express reason that the new circuit was too stable: in effect, the manufacturer confirmed that instability is the feature. The subtler lesson is that voice-to-voice variation is mostly a set of static offsets fixed by component tolerance at manufacture, a slightly different tuning, cutoff, and envelope time per voice, with only a small drift on top. Eidolon fixes those offsets when each voice is created and adds a gentle drift, rather than modeling everything as continuous wandering.
Schwa's Olga (2008) is the closest prior instrument to this thesis, and I studied it directly. Its saturation engine applies gentle overdrive at every stage rather than one hard stage at the end; the developer's own rule, keep the signal low at every stage to stay clean, is the principle stated plainly. Measuring the plugin confirmed that its drive is subtle, a few decibels of compression rather than an aggressive harmonic effect, which is why Eidolon's always-on drive sits low. Olga's oscillator also varies its waveform shape slightly every cycle and never resets its phase when a note begins, so no two notes are identical. Eidolon adopts both behaviors.
The Buchla and Serge modular systems are the reference for the wavefolder, the west-coast alternative to the filter. The design carries several folding topologies rather than one, because a parallel folder (Buchla) and a series folder (Serge/Lockhart) grow their harmonics in audibly different ways. Two further ideas come from this world: the low-pass gate, which darkens a sound as it grows quieter the way a struck object does, and the practice of re-applying a sound's dynamics after folding, which heavy folding otherwise flattens.
Finally, a survey of thirty-six filters, envelopes, and amplifiers produced the design's single most useful finding: the Moog ladder, the EMS diode ladder, and the Curtis chip filter share the same underlying linear response. Their famous differences come entirely from how and where each one distorts. So Eidolon keeps one correct linear filter, a topology-preserving state-variable filter, and makes character a matter of where the nonlinearity sits: inside every stage for Moog warmth, in the feedback for a cleaner Roland voice, in the forward path for the aggressive Korg scream. The amplifier extends the idea, pairing the classic transconductance stage with a recent diode-amplifier model (built on the Wright omega function) whose distortion rises as its gain falls, so the cascade is richer than either stage alone.
The sound target, fixed before code
A principle underlies the whole contract: the sound target is set before the architecture, and the architecture follows from the sound. The target is not "vaguely analog" but a set of measured references. The oscillator's shape is calibrated against published measurements of a real Minimoog sawtooth. The noise floor is set from measured vintage instruments, which sit around sixty to seventy-five decibels of signal-to-noise, noisier than a first guess assumed, so the design floor was lowered accordingly. The distortion targets come from published console and amplifier measurements rather than intuition.
Perception sets the tolerances. Aliasing, the spurious inharmonic tones that digital shortcuts create, is inaudible below a measurable masking threshold, and that threshold, not zero, is the goal. The ear is most sensitive between two and four kilohertz, so the harmonics that land there matter most and are held to tighter accuracy. Two oscillators a few cents apart beat slowly and read as warmth; the same detuning stretched wider reads as roughness, so the instability is kept inside the warmth window. And a recurring result shapes how the instability is built: varying a waveform's shape a little each cycle is far more forgiving than varying its pitch.
Method: judge by ear, then measure
The tuning method is deliberate and, I think, the real content of the contract. Measurements are necessary but not sufficient: a metric catches only what it was built to catch. A single-sample click at a note's onset is invisible to a waveform-continuity metric yet obvious to the ear; conversely, distortion that a meter flags is often inaudible in music. So the loop is build from the best theory, listen against a reference, then measure to confirm, never measure-first-and-trust. Every change that touches the sound is auditioned against the previous version before it is accepted.
Two disciplines support this. Measured hardware figures are treated as fixed baselines, loosened only with a written justification and a regenerated reference. And every source of per-voice randomness (the offsets, the drift, the noise) is independent between voices but fully reproducible from a fixed seed, so a given sequence of notes always renders identically and any change in the sound can be traced to the edit that caused it.
The design decisions that follow
The thesis, the lineage, and the target converge on a set of concrete decisions. These are the most consequential:
| Decision | Why |
|---|---|
| One correct linear filter; character by nonlinearity placement | The classic filters share one linear response; the difference is where they distort |
| Saturation distributed across the chain, always on, kept gentle | Analog color accumulates; a single hard stage sounds concentrated and digital |
| Per-voice offsets fixed when a voice is created | Real voice variation is mostly static component tolerance, not continuous drift |
| Always-on warmth and drift floors that cannot be switched to zero | Stability is the defect; the reissued classics add instability back deliberately |
| Hybrid amplifier: transconductance stage into a diode-amplifier model | The cascade is harmonically richer than either stage on its own |
| Full keyboard tracking of the filter by default | Matches the semi-modular reference; keeps timbre even as pitch rises |
| Noise injected before the amplifier, not after | So the envelope gates it into silence between notes, as the hardware does |
| A modeled console output stage rather than a generic limiter | The rising-with-frequency distortion is the master-bus fingerprint |
| Clean band-limited oscillators, colored downstream | A dirty oscillator's aliasing would compound through every later nonlinearity |
A handful of rules are treated as fixed, each learned from a specific failure: inject noise before the amplifier, not after; retire a voice by watching its envelope, not its output level; never reset the oversampling filters when a note begins; scale polyphony by a fixed factor rather than the live voice count; and always prime a nonlinearity's memory at a note's onset. Each of these prevents a click, a dropout, or an inconsistency that the ear notices immediately.
The signal chain as an argument
The per-voice path makes the thesis concrete. The oscillators are kept as clean as possible, band-limited with minBLEP but varied slightly in shape each cycle and free to start at any phase. From there a structural stage and an always-on imperfection alternate all the way to the output. In the diagram below, those imperfections are the amber stages: each has a documented cause in a specific circuit, and none can be switched off.
- 3 × OscillatorBand-limited; waveform shape varies slightly each cycle, phase free to start anywhere.
- Per-oscillator saturationThe first gentle nonlinearity, once per source.
- Wavefolder A → BWest-coast harmonics; the sound's dynamics re-applied after folding.
- Pre-filter saturationDrives the filter and sets much of its character.
- Filter A → Filter BOne correct linear filter, with a saturation placed between the two.
- Thermal noiseInjected before the amplifier, so the envelope gates it into silence between notes.
- Hybrid amplifierTransconductance stage into a diode-amplifier model; richer than either alone.
- Output saturationA final per-voice nonlinearity before the voices sum.
- Voice sum → console bus → masterThe summed voices take further saturation, mid-side width, and the modeled Soundcraft 600 output stage.
Structural / linear stageAlways-on analog-life imperfection
The order of construction mirrors the argument. The first weeks build the instrument; a later stage makes it imperfect on purpose; the stages after that make those imperfections click-free and alias-free; and the discipline throughout is to protect each improvement, so a fix in one place never silently breaks the sound somewhere else.
Open questions the build raised
Most of Eidolon is careful assembly of known techniques. In a few places, building it carefully kept running into questions the published work seemed to leave open. I raise these as honest gaps I did not find filled, not as claims of priority, and none is the point of the instrument.
One is editing a waveshaping curve in its integral rather than its direct form, which makes an alias-free result true by construction; I found the idea in industry practice but not as a general, user-facing tool. Another is writing a hardware-style datasheet, with inputs, limits, distortion, and cost, for a software signal-processing module, for which I found no existing framework. A third is laying out the whole engine's data for the parallelism of a modern processor and measuring the result, which the field discusses informally but rarely publishes. These are simply the places where the work touched the edge of what I could find written down.
References
The design rests on a body of published virtual-analog research. The core sources:
- T. Stilson and J. O. Smith. Alias-Free Digital Synthesis of Classic Analog Waveforms, Proc. International Computer Music Conference (ICMC), 1996.
- A. Huovilainen. Non-Linear Digital Implementation of the Moog Ladder Filter, Proc. DAFx-04, pp. 61–64, 2004.
- V. Välimäki and A. Huovilainen. Antialiasing Oscillators in Subtractive Synthesis, IEEE Signal Processing Magazine, vol. 24, no. 2, pp. 116–125, 2007.
- V. Zavalishin. The Art of VA Filter Design, Native Instruments, rev. 2.1 (freely available), 2012–2020.
- J. Pekonen, V. Lazzarini, J. Timoney, J. Kleimola, and V. Välimäki. Discrete-Time Modelling of the Moog Sawtooth Oscillator Waveform, EURASIP Journal on Advances in Signal Processing, vol. 2011, Article 785103.
- E. Brandt. Hard Sync Without Aliasing, Proc. ICMC, 2001 (the minBLEP technique).
- F. Esqueda, H. Pöntynen, V. Välimäki, and J. D. Parker. Virtual Analog Buchla 259 Wavefolder, Proc. DAFx-17, pp. 192–199, 2017.
- J. D. Parker, V. Zavalishin, and E. Le Bivic. Reducing the Aliasing of Nonlinear Waveshaping Using Continuous-Time Convolution, Proc. DAFx-16, 2016 (the antiderivative-antialiasing method).
- C. Pines. Real-Time Virtual Analog Modelling of Diode-Based VCAs, Proc. DAFx-25, 2025.