Aureate

A feedback glue law with two circuit classes. One mono-summed sidechain, and a detector that sees the signal after the gain cell, one sample old. VCA is a dB-domain timing network in that loop — the knee, the ratio-dependent attack (63% at τ/(1+k): measured 4.96 / 2.46 / 0.98 ms at the 10 ms position for 2:1 / 4:1 / 10:1) and the program dependence are consequences of the topology, not features bolted on top. Vari-Mu swaps in a softplus dead-zone sidechain, a current-limited rectifier, a trapezoidally-discretised three-capacitor release network and a gain cell derived from the analytic transconductance of a published triode law. Its attack is a genuine slew, and the test proves it the hard way: peak d(GR)/dt grows only 1.22× when the step size doubles, against 2.00× for the VCA law's exponential — a bound no exponential can satisfy at any time constant.

An Auto release that remembers, without measuring anything. The slow reservoir capacitor is only ever charged by release-phase ripple, so sustained programme fills it and a lone transient does not. A 50 ms peak is forgotten in 0.28 s; ten seconds of sustained gain reduction takes 2.37 s — 8.3× longer — with a tail that never bounces. The release switch's printed markings are nominal, by design: measured circuit-class timing runs longer than the label at every position, and the manual states the real numbers rather than the print.

Iron: a transformer stage where the low end is Faraday's law, not an EQ curve. Integrate to flux, saturate the core, differentiate back. Because flux is the integral of the signal, the core saturates far harder at low frequencies for the same level, and third-harmonic content rises steeply toward the bottom end on its own — measured at roughly 10 dB per octave, with no coefficient anywhere fitted to produce it. The integrator/differentiator pair is backward-Euler matched, not bilinear, because the bilinear inverse carries an undamped pole at Nyquist that a transient would park forever; the shipped pair is a pure one-zero FIR that decays to true zero, verified after ten seconds of silence.

HQ: a lower alias floor, and the same voicing. First-order antiderivative anti-aliasing on the same three Character curves, in the same oversampled region, at the same reported latency. Measured 24 dB lower non-harmonic floor on every Character at 10 kHz / 0 dBFS / 24 dB Drive — while the audible harmonic structure stays within ±1 dB of Classic and the low-drive magnitude response within 0.1 dB to 5 kHz. The antiderivatives are verified against their curves before a single aliasing number is claimed. HQ carries an inherent half-sample delay at the rate it runs at internally — far below the reported-latency's integer granularity and inaudible — so it is an alias-floor option, not a re-voicing.

Proof, not adjectives. 152 tests in Release on macOS and Windows, plus strictness-10 format validation on both and AU validation on macOS. Zero heap allocations on the audio thread, under an instrumented allocator, with every new path hot. Bit-identical output at block sizes 1, 513 and 4096. Identical timing in absolute seconds from 44.1 to 192 kHz. Factory presets pinned by SHA-256. Zero subnormal samples after ten seconds of silence.

Stays out of the way until you ask. Every new control is neutral by default, and each neutral default is a branch-skip rather than a transparent setting — the Glue section returns before touching a sample, the Iron stage is stepped over, Classic quality still runs the previous release's own code. An existing session plays back bit-identically, with reported latency unchanged at every setting and every sample rate.