Why a soft clipper sits on the master bus
EV·ENGINE trims each preset to match loudness across twenty engines, but a trim is a single gain applied to the whole waveform, peaks included. Without anything else in the signal path, twelve of the fifteen presets measured still peaked above full scale, so a soft clipper sits on the master bus to catch the overshoot the trim cannot remove.
Nineteen units apart, then half a unit
The twenty presets in EV·ENGINE began nineteen loudness units apart under ITU-R BS.1770 K-weighted measurement. A per-preset trim closed that gap to half a unit, and the method is covered in matching loudness across twenty engines. That trim fixed how loud each preset sounds averaged over a drive. It says nothing about how loud any single sample gets.
A trim is one number multiplied against every sample in the waveform, whether that sample sits near silence or at the top of a peak. The number is chosen so the long-run measured loudness lands near a shared target, and a spike lasting a few milliseconds barely moves that average. The trim can shrink the whole signal and leave the tallest peak in exactly the same place relative to everything around it.
Checked against the fifteen presets logged in src/audio/masterLimiter.ts, twelve peaked above full scale with the trim applied and nothing further in the signal path. The trim had done the job it was built for and left the transients precisely where synthesis put them, above 0 dBFS.
Where the overshoot comes from
Some of the twelve are the presets that lean hardest on the exhaust. The anti-lag rally engine runs overrun pops at maximum and noise at 145, and every pop is a short burst that has to punch through the mix rather than sit inside it. The same shape shows up quieter in the turbo hot hatch, which pairs overrun pops at 70 with sub level at 66.
The exhaust itself, modelled as the comb filter described in the exhaust is a comb filter, rings after each firing rather than stopping dead. Comb feedback pushes that ring higher, and the alien craft preset runs it at 90 against a brightness of 190, which stacks harmonic content on top of a resonance that is already reinforcing itself.
Drive into the waveshaper adds a different kind of overshoot: it shapes the waveform rather than scaling it, pushing harder into a nonlinearity that adds harmonics without adding proportional loudness, and a preset with high drive and a lively waveshaper character can generate a peak nowhere near its average level. These presets were built to do exactly this, and the peak is the sound working correctly.
A knee close to full scale
A hard limiter clamps anything above a ceiling to that ceiling, which is simple and audible: the top of the waveform goes flat and the flat section reads as a buzz layered onto the note. The clipper on EV·ENGINE's master bus is soft instead: the waveform bends increasingly as it nears full scale, rather than being marched flat above a threshold, so the loudest samples round off instead of slicing off.
The knee sits close to full scale rather than partway down the range, so it only engages on the overshoot a trim could never remove, and leaves everything below that point exactly as synthesis produced it. Of the fifteen presets checked, three never peak above full scale at all, and those pass through the clipper unchanged.
Listen for it on the steam locomotive, whose single chuff every half revolution runs noise at 200, or on the big single, whose one bang every two revolutions carries a firing unevenness of 55, the highest of any preset here. Each transient lands as a clean thump rather than a harsh crackle, because the sample that would have clipped hard against full scale was rounded off before it reached the output.
The same bass that broke RMS
The decision to measure loudness with ITU-R BS.1770 K-weighting rather than plain RMS was made for a related reason: RMS over-weights bass, and would have mis-ranked signals this bass-heavy rather than levelled them. The muscle car V8, with a firing unevenness of 38 and everything that implies about sub oscillator and low-end emphasis, is exactly the kind of preset RMS would have flattered.
K-weighting attenuates low frequencies before measuring, which is closer to how loud a bass-heavy signal sounds against everything else. It also means a preset can measure moderately loud under K-weighting while its low end still stacks up to a hot peak, because the peak meter downstream of that measurement does not apply the same weighting. A signal can be judged correctly for loudness and still overshoot on sample value.
That gap between a loudness measurement that discounts bass and a peak meter that does not is exactly what the clipper closes. The trim gets the average right by the standard that matches how the ear works. The clipper gets the sample right by the standard that matters to whatever is converting that sample to a voltage.
The clipper does not fix a mistake in any preset. It exists because a gain trim and a peak are two different quantities, and getting the first one right, to half a loudness unit, was never going to move the second one at all.