DOSSIER Methods, measurements, contracts
Everything here exists in the codebase today and is enforced by the test suite. Each item states what we did and why it matters to the sound.
The resonators and filters are TPT state-variable structures solved without unit delays. An independent oracle test re-implements every analog prototype from scratch and checks the filters against it to fractions of a decibel, most of the way to Nyquist.
Why you care: the resonances sit where the circuit says, at any sample rate.The kick click is the original four-component pulse-shaper network solved per sample: a trapezoidal companion model of the capacitor with the diode’s transcendental equation resolved in closed form via the Wright omega function. Not a filter that sounds like a click; the circuit that makes one.
Why you care: accent drives that network harder and the click changes shape, exactly like the hardware.The six-oscillator metallic bank runs polyBLEP-corrected squares at the measured frequencies and measured 47.98 percent duty cycle, because that asymmetry is part of the spectrum. The bass companion’s saw and pulse oscillators are polyBLEP with hard sync.
Why you care: bright metallic voices without digital hash.Voices whose pitch moves every sample recompute filter coefficients per sample through a bounded polynomial with relative error under one part in a billion. The audio path never calls a transcendental function; everything expensive happens at prepare time or lives in a precomputed table.
Why you care: smooth knob motion and modulation with no CPU spikes and no drift.Discrete white noise spreads fixed power across sample rate, so a naive port gets quieter snares at 96 kHz. Our sources are power-spectral-density invariant and band-limited at 24 dB/oct to the hardware’s audio bandwidth, so high sample rates gain no unphysical ultrasonic energy.
Why you care: your snare has the same body at 44.1 and at 192.The hardware’s common trigger is a one-millisecond pulse. Integer sample rounding would shift that pulse’s area by up to a quarter percent across rates, which amplitude-sensitive circuits amplify into audible drift; we correct the area exactly while keeping the edge clean.
Why you care: identical attacks at every sample rate.The bass companion’s filter is a four-pole transistor-ladder model with zero-delay feedback and saturating stages, pole-mixed into low-pass, band-pass and high-pass at 12 or 24 dB per octave from one core, stable into self-oscillation.
Why you care: one filter that growls, screams and stays put.Every modeled machine gets a research pass first: service-manual extraction with page references, published circuit analyses, and a register of where sources disagree. Values we could not source are marked provisional in the code itself, next to the number.
Why you care: no folklore constants.The default kit is checked against reference recordings of a documented unit: kick fundamental and full decay-law sweep, snare shell pair, cowbell pair, hat decay ranges and the attack pitch signature. The 909-family models are fitted to two independent units, including eleven-step accent sweeps recorded from individual outputs at 96 kHz/24-bit.
Why you care: the knobs behave like the hardware’s knobs, across their whole range.Where internet numbers disagreed with the circuit, we measured: pitch trajectories by zero-crossing analysis, decay laws by log-domain RMS slopes, oscillator pairs by spectral peaks, accent tapers step by step. Several famous numbers did not survive.
Why you care: the model matches hardware, not forum posts.A tracking allocator is installed as the global allocator in the test suite and counts every allocation across prepare and processing for every machine. The required count is zero.
Why you care: no glitches when your session gets busy.A source-audit test walks the production source on every build, parses every audio-path function body, and fails on any forbidden token. The audit includes a mutation self-test that injects a violation to prove the guard still bites, and discovers new files automatically.
Why you care: the real-time promises cannot silently rot.A render chopped into pathological block sizes, with events landing on arbitrary boundaries, is asserted bit-identical to the same render in one pass.
Why you care: your DAW’s buffer setting cannot change the sound.The only randomness in the engine is a seeded generator whose constants are pinned by a test. Noise reseeds per trigger, so two identical hits are identical; two identical renders null to silence. Unit-to-unit character is planned as seeded component tolerances: expressive, reproducible and testable.
Why you care: recall a session years later and get the same record.Design math in 64-bit, coefficients checked and stored in 32-bit, filter state kept in 64-bit, sub-audible tails flushed deterministically. Resonant low-frequency voices with long decays are exactly where lesser precision policies fall apart.
Why you care: long kick tails that ring out clean.A written policy defines which reference code may be read, which must never be opened, and how ROM-era content may be produced: independent resynthesis with measurement oracles, never sampling anyone’s recording. The clean corpus for every machine is documented before its code is written.
Why you care: the instrument you build music on is built on clean ground.The synthesis crate has zero dependencies. Every state struct is fixed-size and copyable. The plugin layer is CLAP first. Supply-chain checks and a pinned toolchain guard the build.
Why you care: fewer moving parts between the circuit model and your speakers.Each machine ships the way the first one did: research document, per-voice specification with sourced numbers, hardware calibration, the full rate matrix and determinism gates. A machine that cannot meet the bar waits.
Why you care: the badge on the box means the same thing every time.05 / 05 Generated proof
Every plot is regenerated from deterministic engine output or committed reference fixtures. No smoothing changes a conclusion.
Kick pitch trajectory at 44.1 through 192 kHz, plotted against the cents budget.
Two fixed-seed renders and the sample-by-sample difference: negative infinity dBFS.
The kick’s measured reference decay sweep and the implemented anchor law.
Two hardware units and the current 909-family model. Settled pitch remains essentially flat.
Level-normalized spectra reveal the nonlinear timbral change produced by trigger voltage.