What the audition cannot hear

One assumption runs through every page of this book: that preference measured under a fixed gesture is the same thing as musical taste. This page is where that assumption is examined and found to be doing more work than it can carry.

The argument is not new here. It is already in this book, made once, in the page that replaced the v1 phrase with the v2 one:

So the grammar could express patches the audition could never reveal, and the taste model was being asked to learn preferences over evidence that was not in $\varphi$. No amount of model improvement fixes that; it is a measurement problem.

— The standard phrase

That reasoning was applied to four holes — slow pads, sub-Hz modulation, anything above Eb4, polyphonic stacking — and each was closed with the cheapest segment that revealed it. Then it stopped. Everything below is the same sentence, still true, about the things the v2 phrase did not reach.

Why this is a separate register

Unraised directions holds possibilities the design never considered. This page holds something different and less comfortable: one premise the design commits to everywhere, and the eight places that commitment costs something. An entry there is an opportunity; an entry here is a limit on what any amount of modelling can currently buy.

Seven and eight are not limits of the audition; they are places where the same premise has quietly shaped the product rather than the measurement. They are here because they come from the same root.

1. A patch is a function from performance to sound, and it is sampled once

NoteSpan — the whole of what the renderer knows about a note — is:

pub struct NoteSpan {
    pub voct: f64,      // pitch, V/Oct from C4
    pub chord: usize,   // additional gate-synced voices
    pub on_start: usize,
    pub on_end: usize,
}

There is no velocity field, anywhere in the render path. Every audition note is struck identically.

Three feet away in the same repository, the live instrument takes MIDI velocity, pitch bend and sustain, and the guide documents playing it that way. So the instrument responds to a dimension the measurement holds constant.

The consequence is not a missing coordinate; it is a category error about what a candidate is. A patch is a function from performance to sound. $\varphi$ evaluates that function at a single point and hands the result to a model that then speaks about patches. Everything downstream — $\theta$, the style lenses, the map, the calibration diagram — is a faithful model of preference over point samples of instruments, presented as a model of preference over instruments.

What is invisible as a result, in rough order of how much a player would care:

• Velocity response. The single largest determinant of whether a patch is playable at all.
• Note-length behaviour. Whether it blooms, or just holds.
• Retrigger under fast playing. Whether the envelope survives a run.
• Key tracking, beyond the one octave-up check high_ratio performs.
• The arpeggiator. It ships in the instrument and appears in no audition.

This is the one thing on this page that cannot be fixed inside $\varphi$. Every other entry is a coordinate or a surface; this is a change to what is being measured. The in-idiom version is exactly the move the v2 phrase already made: a second render at a different velocity and articulation, contributing difference coordinates — $\Delta$brightness per velocity, $\Delta$level per octave — rather than more absolute ones. Two renders instead of one is the same cost-for-coverage trade the v2 phrase paid knowingly, and the segment-local coordinates are the precedent for measuring a contrast rather than a level.

2. Modulation is the distinctive claim, and its rate is not measured

The instrument's distinguishing feature is that modulation is a whole chain: an s&h rand → quantize → slew can reach a cutoff, the node bank exists to explain where a modulator may legally go, and nearly every module carries a mod slot.

Here is everything $\varphi$ records about that:

Rate appears nowhere. Shape appears nowhere. A 0.3 Hz filter sweep and a 7 Hz tremolo at equal depth land on near-identical coordinates. A random sample-and-hold and a sine LFO are indistinguishable to every one of them.

So "I like slow evolving movement and dislike fast wobble" is not a preference that is hard to learn. It is inexpressible — in exactly the sense the log-axis argument uses about brightness on a linear-Hz axis, and for the same reason: the model is linear in $\varphi$, so a distinction absent from the coordinates is absent from the hypothesis space.

The second-order consequence is worse than the first. Proposal tilts read the structural coefficients, and mod_density is one of them — so the search can learn more modulation and can never learn slower modulation. The instrument's most distinctive capability is the one the search cannot be rewarded for using well.

The measurement already has its window open. The held note exists, in the words of its own page, to reveal "sub-Hz modulation over a register-constant sustain". The segment was built for this and then not measured for it. An autocorrelation of the centroid or RMS envelope across that span yields two coordinates:

• dominant modulation rate, on the shared octave axis every other frequency in $\varphi$ already uses, and
• periodicity strength — how much of the movement is periodic at all, which is what separates an LFO from a random walk.

If two numbers could be added to $\varphi$, these are the two.

Status — rate is measured now; periodicity was tried and cannot be, in this window. Rate arrived as three motion bands rather than one dominant-rate number: a single "rate" coordinate is undefined for a static patch and meaningless for a random walk, while band energies are defined for everything and let a linear model hold slow yes, fast no as two coefficients of opposite sign. On the probe ladder the band that reads highest follows the LFO rate from 0.55 Hz to 13 Hz. Periodicity failed its own test: both an autocorrelation peak and the harmonic share of the modulation spectrum separate an LFO from a random walk at 2.7 Hz and above and not at all below 1.5 Hz, because the held span holds fewer than three slow cycles. That is the range evolving textures live in, so the coordinate was not shipped; it is now a stimulus-length question, and belongs with §1.

3. Loudness is normalized away, and the raw level is already computed

Normalizing every render to −18 LUFS is correct, and the reason given is the right one: loudness bias would poison the preference data.

The other half of that trade is not stated anywhere. Loudness is part of a sound's identity. Hits hard, sits back, has weight are ordinary, strongly-held preferences, and they have been normalized out of existence by design. crest and rms_std recover dynamics within a patch; absolute level is gone.

What makes this an easy win rather than a lament is that the number already exists. VetReport measures peak and rms on the raw, pre-normalization render — deliberately, because the gate's thresholds are about real output level — and both are then discarded. The quantity how loud is this patch natively is computed on every candidate this project has ever rendered.

Adding $\log(\text{raw rms})$ as a coordinate reintroduces no playback bias, because playback stays normalized; the confound the loudness decision guards against is between what the listener hears and what they are comparing, and a feature the model reads is neither. It simply lets the model see an axis that is currently invisible.

The honest caveat, which decides it: raw level is partly an artifact of gain staging inside the grammar rather than a property anyone hears. That makes it possibly noise — and it makes the question a single sweep on existing data rather than an argument.

4. Everything is judged in a silent room

No mix context, no drums, no key, no other material. Every duel asks which of these two do you prefer, in silence.

Musicians almost never choose sounds absolutely. The bass that wins in isolation is routinely the one that disappears under a kick; the pad that wins alone is the one that eats the vocal. $\varphi$ carries bass_fraction and three brightness measures, but nothing about spectral occupancy relative to other material, because there is no other material for it to be relative to.

This is probably the deepest reason isolated-audition tools feel wrong to practitioners in a way that is hard to articulate, and it bites harder here than elsewhere: the premise of this project is that the instrument is learning what you like, and what it is learning is what you like in an empty room.

The scope that would test it is not "build a DAW". One user-supplied backing loop, played underneath the audition, changes what every vote means — and it interacts directly with §3, because level against a bed is exactly the judgement normalization removes.

5. There is no tempo, and the presets are working around it by hand

An LFO's rate is sampled as u01() and mapped to Hz. There is no global tempo, no host sync and no note divisions. A Clock tempo exists for the euclidean sequencer's own bpm port, and it is not a session-level musical clock.

The evidence that this costs something is in this repository's own preset source, in comments:

rate: 0.55, // ≈83 bpm
rate: 0.62, // ≈107 bpm
p0: 0.55,   // ≈83 bpm — about one jump per bar

Preset authors are hand-computing, in comments, the thing a sync control would compute. An unsynced modulator is among the most common reasons a good patch is unusable in a track: it can be set close by ear and it drifts over sixteen bars.

For an instrument whose output is meant to end up in music, tempo-relative rates are not a convenience. They are the difference between a sound and a part — and they would also give §2 a natural axis to express rate on, since a listener's preference about movement is far more plausibly "a cycle per bar" than "0.34 Hz".

6. The loop is selection; sound design is pursuit

A duel asks which do you prefer. A player with a sound in their head wants get closer to this.

Locks plus ⚡ evolve from this is a genuine pursuit affordance and a good one. But the primary loop is judging what you are handed, and the target-directed entry in the other register is better understood from here: it is not a feature, it is the missing half of a workflow.

There is a sharper form of this, and it is about what the model converges to. A preference fitted over everything a person has ever voted on approaches the average of their taste. Max-of-experts answers part of it — islands rather than a centroid, which is the whole reason for the max — but the islands are discovered from $\varphi$, not selected by intent. So on any given evening the instrument proposes toward the union of everything the listener has ever liked, when what they need is the one corner they are working in tonight.

That is the musical argument for declared context, and it is a stronger one than the modelling argument on that page. It is the difference between a tool that knows you and a tool that is useful today.

7. Comparability constrains the measurement, not the listener

The decisions log records the audition as "Standard 5.05 s phrase + free-play", with the rationale "feature comparability requires fixed stimulus".

The rationale is true of $\varphi$. It is not true of the person.

$\varphi$ comes from a deterministic offline render. What a human hears while deciding is an independent choice, and could be anything at all — including their own playing, on both patches, with the same lick. The coordinates would remain exactly as comparable, because nothing about them depends on what came out of the speakers during the vote.

So duel in your own hands costs $\varphi$ nothing. It is a far more musical elicitation than voting on four notes chosen by the instrument, and it is the direct remedy for §1: a player testing velocity response themselves is measuring the dimension the phrase holds constant, even if the coordinates still cannot see it.

The cost, stated rather than hidden. The vote becomes a judgement about an experience the coordinates only partly describe. That adds noise, and adds bias if a player's own gestures systematically emphasize something the phrase does not. But it is noise in the measurement of something a listener cares about, against precision about something they do not — and the project already owns the instrument that would detect it, since calibration by provenance exists precisely to score differently-collected answers against each other rather than assume they are equivalent. A PlayedDuel provenance would make this measurable on arrival.

8. Forty addresses and no macros, when the macro axis is already fitted

The rack exposes every trace address, and the word macro does not appear in the web app. A player performing wants two to four hands-on controls, not one per address in the term.

The interesting part is that the fitted model already contains the right axes. A style lens $\theta_k$ is a direction in feature space whose meaning is what this listener cares about. A macro that moves a patch along the model's top learned direction is a control that is personal by construction — bite for one user, movement for another — and nothing else can offer it, because nothing else has fitted the model.

It is also the continuous form of the counterfactual explainer: same posterior, same trace addresses, same predicted-gain arithmetic. One reports a discrete edit; the other puts the direction under a finger.

If these were ordered by value per unit of work

1. Two coordinates for modulation rate and periodicity (§2). Rate done, as three motion bands; periodicity blocked on stimulus length. Cheapest and highest value: it lets the search be rewarded for the instrument's best feature, and the segment it needs already exists for this exact purpose.
2. log(raw rms) as a coordinate (§3). Already measured on every candidate, currently discarded; recovers an axis of ordinary preference at zero audition cost.
3. Velocity in the phrase, as difference coordinates (§1). The expensive one, and the only one that moves the modelled object from a sound toward an instrument.
4. Tempo-relative modulation rates (§5). Turns output into something usable in a track, and gives §2 its natural axis.
5. Declared context (§6), promoted from the other register on musical grounds.
6. Duels in the player's own hands (§7). Nearly free, conceptually the largest unlock, and measurable via provenance from the day it ships.

The summary worth keeping

This is an unusually rigorous instrument for measuring timbre preference under a fixed gesture, and much of the design reads as though that were the same thing as musical taste. It is not, and the gap between them is where every entry above lives.

The tool for closing it is already in the book. The grammar can express what the audition cannot reveal, and no amount of model improvement fixes a measurement problem — applied once, to four holes, and it deserves to be applied to the rest.