music/MUSIC_CRAFT_DOCTORATE.md
CANON SUBORDINATION — this document is PROPOSAL-TIER: it serves canon and never outranks it.
Canon served: the composition floor ruled by Josh on 2026-08-07 -- docs/spine/DECISIONS_PENDING_JOSH.md,
MUSIC ROUND 1 GRADED + THE COMPOSITION FLOOR (commit da627060); the theme rows of
registries/T0_Theme_Registry [DRAFT v0.1].If this document disagrees with canon, CANON WINS and this document is the defect.
Tier: RESEARCH SYNTHESIS, proposal-tier. This document contains no canon and rules on nothing.
It is the craft layer underneath docs/proposals/MUSIC_COMPOSITION_DOCTRINE.md — the reasons its
rules are the rules, the places the reasons do not support them, and the single prioritised list of
what to build next.
Chartered by: Josh, 2026-08-07, at the tail of docs/spine/DECISIONS_PENDING_JOSH.md
(commit da627060), grading round 1 of the generated music and turning all eighteen candidates
down: *"Learn real music theory and what someone who studied music and got their phd would focus
on and improve."* Under the standing law of 2026-08-01 that ruling is a FLOOR TO ENHANCE, not a
specification to meet.
What it is not. It is not a summary of the six lane documents and it does not restate them.
Sixty-two thousand cited words sit under docs/proposals/music/craft_research/ and each lane is
authoritative for its own domain. This document carries only what no single lane could produce:
where six independent lanes CONVERGED, where they CONTRADICT what this program already believes,
the thirteen principles that actually change what we build, and the measurement debt in the order
it should be paid.
Every number about our own corpus in this document was re-derived at HEAD through
harness/music_gen/derive_patterns.py, not quoted from a prior artifact. Where a lane's figure anda re-derivation disagree, both are printed with their denominators.
craft_research/COUNTERPOINT_AND_VOICE_LEADING.md — species counterpoint and what survives it,voice-leading derived from perception rather than authority, an ERB-computable low-interval-limit
floor, dissonance treatment and the deliberate clash, and non-Western polyphony read as
counterpoint with its own rules. Authoritative for: why many layers do or do not become a wall.
craft_research/ORCHESTRATION_TEXTURE_AND_DENSITY.md — the orchestration canon, real layer countsin four different units, texture types, the staged entrance schedule, density management, and the
non-orchestral layers Josh named. Authoritative for: how many layers, and when each one arrives.
craft_research/LEITMOTIF_TECHNIQUE.md — Wagner through Williams through the JRPG masters, thearchitecture question, deployment across a long game, and the measurable signature of a motif.
Authoritative for: what a theme is and how it is transformed.
craft_research/REPETITION_AND_VARIATION_FORM.md — the variation canon, ostinato development,minimalism's written-down rules for literal repetition, cyclic traditions as engineering, and game
loop form. Authoritative for: the 3-to-5-cycle law and the novelty schedule.
craft_research/HOOK_AND_TENSION_ARCHITECTURE.md — hook construction theory, the empiricalmemorability literature and its limits, the tension parameters one by one, arch form and climax
placement, and the drop. Authoritative for: what makes a cue land and what makes it exhausting.
craft_research/ORGANOLOGY_SLICE_CULTURES.md — the instrument roster, tuning system, ensemblepractice, voice layer, environmental layer and care posture for all thirteen slice nodes.
Authoritative for: what may sound in which chapter, and what must never.
Josh's floor contains what looks like a contradiction with our own measured derivation, and it is
not one. build/audio/exemplars/PATTERN_FINDINGS_V2.md found that the tracks Josh named as lifelong
favourites sit BELOW their own albums on structural novelty and ABOVE them on self-similarity — they
repeat more and change less than the filler beside them. Josh's floor says change every three to five
cycles. The repetition lane resolved it by measurement rather than argument, and the resolution is
the governing idea of the whole pass 2.
The loved tracks repeat their MATERIAL and change their TREATMENT — continuously, in small
increments, under a sparse ridge of about three genuine structural events. Measured at HEAD on tracks
of forty-five seconds or more: a corpus row carries about half again as many structural boundaries as
its own album's filler (16 against 11) and makes each one SMALLER — its share of boundaries at or
above 0.7 of its own maximum is 17.9% against the siblings' 36.4%, and its strong boundaries arrive
at 0.88 per minute against 1.89. It brings a new voice in more often and raises the level less when
it does (4.55 lane-adding raises per minute at 7.49 dB, against 3.10 at 10.09 dB). And the ratio of
its largest event to its typical one is steeper, 1.91 against 1.65, which is what a hierarchy looks
like when it is real rather than flat.
So the law and the measurement are the same fact read at two scales. Josh's floor governs the dense
tier: something must change every few cycles, and in the corpus something changes every 5.79 bars.
The measurement governs the sparse tier: most of those changes must be SMALL, because a track that
makes every change a big one has spent its hierarchy and has nothing left to arrive at.
That single sentence reframes what round 1 got wrong. Round 1 was not short of events. It had far too
many of the wrong size — a median of 14.08 drops per minute against the corpus median of 5.57
(2.53x, re-derived at HEAD), which is a track shouting a structural event two and a half times more
often than any music Josh loves, while changing its treatment almost never in between.
Two corroborations are worth recording beside it. First, Josh's number was arrived at independently
by the game-audio literature: a 2012 Game Developer treatment of loop fatigue gives one to two
repeats as tolerable, THREE TO FIVE as problematic and more than five as significant fatigue — the
same band, neither source citing the other. His floor is not a preference we are accommodating; it is
a threshold that has been found twice. Second, a corpus study across 21,391 pieces puts bar-level
repetition at 62.79% and finds it flat from the NES era through the PlayStation era, across roughly a
tenfold jump in storage. Looping is an aesthetic choice, not a memory constraint, and it has never
stopped being one.
Convergence across lanes that never read each other is the strongest evidence this pass produced.
Six items, each reached by at least two lanes independently.
counterpoint lane names accidental surface-texture integration as the mechanism and declared
stratification as the fix. The orchestration lane reaches it from Huron's finding that listener
counting accuracy collapses past three or four concurrent voices, and concludes that dozens of
parts must group into three or four perceptual STRATA. The organology lane finds the same shape in
the traditions themselves. The reconciliation of Josh's two sentences is therefore structural: many
PARTS, few STRATA. Enforce the strata count, not the part count.
audio-side methods and found none defensible for a part count; the counterpoint lane reached the
same conclusion from the card's own method note; the repetition lane found the proof, which is that
max_lanes is pinned at 9 for hits and siblings alike — a ceiling in the ten-band proxy method
rather than a fact about any music. The layer count must be a FIELD we wrote, with every audio
figure as corroboration only.
colour_change field drawn from an enumerated eleven-device list; the repetition lane derives the
same rule from the variation canon and from irama; the leitmotif lane's transformation budget is the
same discipline applied to a theme. This is the 3-to-5-cycle law made mechanical, and it is
checkable before a sample is rendered.
separable density levels — is named by the organology lane as the cross-cutting device and by the
repetition lane as the mechanism this project needs most. Two of our slice chapters are literally
gamelan cultures, and our own layer_mode: swap already means this without knowing it did.
ambience in his floor. The orchestration lane treats them as parts with register slots and entry
schedules; the organology lane independently attests body percussion in six of twelve slice nodes.
This is not decoration and it is not a texture layer; it is scored material with a frequency slot.
our own rejection filter all land on the same axis, and it is the one axis on which round 1 failed
unanimously: every one of the fourteen round-1 rejections fell out of the envelope on composed rest
and nothing else.
These are findings for the program to rule on. Nothing here was applied.
complexity AT structural boundaries and hold it flat inside sections. Measured at HEAD, corpus rows
put a SMALLER share of their complexity steps at boundaries than their siblings do — 0.358 against
0.500. Change is not carried by the form; it happens inside the sections, which is exactly what
continuous re-treatment over a returning cycle looks like from outside. §11.6's own instruction is
"if the corpus says otherwise, the corpus wins." The corpus says otherwise.
0.575 and a flow-peak median of 0.606, which looks like a confirmation until the spread is printed:
p10 to p90 runs 0.16 to 0.87 and only seven of twenty-eight rows land inside 0.55 to 0.70. Pinning
every cue's peak at 0.618 would make our output MORE uniform than the music we are trying to match.
C3 head-cell range band of 7 to 17 semitones and the card's the_hook.range_semitones, whose corpus median is 22, are not the
same object, and a cue can pass one while failing the other for no musical reason.
found this independently. motif_economy and repeat_map run on raw chroma and MFCC
self-similarity, so a theme returning a fourth higher scores LOWER than one returning in the same
key. For a leitmotif score that is exactly backwards: the transposed return is the craft, and our
instrument reads it as new material.
PATTERN_FINDINGS_V2 documents 28 hits, 851 controls and 1181 cards with EX_014 quarantined. The live pool at HEAD is 30 / 927 / 1180 —
EX_014 and EX_094 had their quarantines retired by inspection on 2026-08-07, both legitimately.
Nothing is wrong; the derivation is simply two positives stale, and the drift is material (corpus
median dynamic range moves from 11.7 to 12.23 dB). Any lane quoting V2's tables must say which
denominator it ran at, and a V3 re-run is owed.
Ordered by how much they change the next track, not by how interesting they are.
1. COMPOSE THE STRUCTURE BEFORE THE SOUND. A track is a section map, a cycle grid, an entrance
schedule, a novelty schedule and a hook plan before it is audio. Everything else here is
downstream of this one. PIPELINE: the pass-2 track plan; single-shot generation cannot satisfy any
clause of the floor because it places nothing.
2. MANY PARTS, FEW STRATA. Twenty-four to forty parts for a full cue, three to eight for the sparse
nature class, grouped into three or four perceptual strata. Our ten authored tracks currently
average 9.3 parts, range 4 to 19 — about three times short on the full-cue class, and the gap is
known exactly in the right unit before anything renders. PIPELINE: stratum and
part_count_floor on the manifest; the battery refuses out-of-band.
3. THE MANIFEST IS A SCHEDULE, NOT A ROSTER. Every part carries entry, exit and rest spans. Three
predicates then become checkable before rendering: no part runs the whole cue except a declared
ostinato, at least one real subtraction occurs after the midpoint, and the peak part count lands in
the last third. PIPELINE: entry_bar / exit_bar / rest_spans.
4. CHANGE THE TREATMENT CONTINUOUSLY, THE FORM RARELY — AND ON TWO TIERS. A TIDE of small treatment
changes every four to eight bars (the corpus does it every 5.79), under about three LANDMARKS per
track arriving at roughly 0.88 per minute where the filler beside them fires 1.89. Josh's law
governs the tide; the measurement governs the landmarks; the hierarchy between them is the whole
deliverable, and it has a number — largest event over typical event is 1.91 for corpus rows against
1.65 for filler. PIPELINE: the novelty schedule carries a MAGNITUDE and a TIER per event, not just
a time.
5. THE LAW IS NON-DECAY, NOT ESCALATION. This corrects an assumption the program was carrying.
Corpus strong events are close to uniformly spread — median position 0.477, with 47.2% in the
second half — so a cue that escalates monotonically toward its loop point collapses at the wrap.
What separates the loved tracks is that they do not DECAY: complexity in the last third minus the
first third is -0.020 for corpus rows against -0.094 for siblings. Gate the non-decay, never the
escalation.
6. A NOVELTY EVENT ONLY COUNTS IF IT ADDS OR REMOVES A FOLLOWABLE VOICE, or changes the rule the
music is running under. A filter sweep, a reverb change, a volume ramp and a pan move are CHEAP and
satisfy nothing. The warrant is measured: corpus raises average 7.49 dB against the siblings' 10.09
while firing lane-ADDING raises at 4.55 per minute against 3.10 — novelty comes from arrivals, not
from loudness. PIPELINE: classify on lanes_added and lanes_removed; anything rounding to zero
is cheap and cannot close a cycle.
7. TAKE THE DENSITY-LEVEL SHIFT AS THE DEFAULT LANDMARK. Javanese irama re-tiles a fixed cycle at
separable density levels and Carnatic laya reached the identical ladder independently; it is a real
structural event that needs NO new material, and it leaves a measurable signature — a step in note
rate and per-bar complexity while repeat similarity STAYS high. Two of our slice chapters are
gamelan cultures and our own layer_mode: swap already means this.
8. REPETITION IS LICENSED ONLY BY A NAMED COLOUR CHANGE. Carrier, family, register, solo or tutti,
mute, attack mode, articulation, added counter-line, reharmonisation, composite timbre, timbral
echo. A verbatim repeat with no named device is a battery failure. PIPELINE: colour_change
required on every licensed repeat.
9. SUBTRACTION IS THE STRONGEST EVENT AVAILABLE. Removal creates more energy than addition, and the
loudest moment is followed by the least. Round 1 dropped 2.53 times more often than the corpus and
was rejected unanimously on that axis, which is what happens when the strongest device is spent as
small change. PIPELINE: a required removal event of declared depth after the flow peak; composed
rest held inside the corpus band of 2.68 to 17.43.
10. STATE THE THEME PLAINLY ONCE, THEN NEVER NEUTRALLY AGAIN. The plant is unhurried and in the home
carrier; every later appearance is fragment, transposition, augmentation, reharmonisation, carrier
migration or mode change. If it needs its orchestration to be recognisable it is not a motif — test
it through three carriers before spending anything else on it. PIPELINE: plant_node, the
transformation budget, and motif_compare.py, which exists today and is not wired.
11. COMMON CONTOUR, ONE UNCOMMON LEAP. The single melodic prescription the empirical work actually
supports. Everything else about melodic memorability in the literature is weaker than it is usually
quoted as being, and this document says so rather than dressing craft consensus as evidence.
PIPELINE: the_hook.pitch_contour plus one licensed distinctive_deviation.
12. AUTHENTICITY IS A CONSTRAINT ON MECHANISM, NOT ON TIMBRE. The traps the organology lane found are
not wrong-sounding instruments; they are right-sounding instruments from the wrong century — gong
kebyar is 700 years late for Bali, the djembe is a twentieth-century global export, the likembe
arrived with colonial labour migration, and a bone flute at Olduvai is 1.76 million years early.
Where the record is genuinely absent, invention is allowed and required; it just has to be declared
as invention rather than dressed as evidence.
13. MENACE NEVER RIDES THE PLACE'S OWN REGISTER. Already canon in every theme row's
deny_register_list; the exoticism literature independently supports it as craft. The antagonist
material carries its own darkening operations and travels with him. PIPELINE: a per-cue check that
a region's instrumentation set never co-occurs with the antagonist's darkening operations — which
is checkable from the cue table as soon as cues carry both fields, and is not checked today.
The four instruments this sitting built cover the first four. Items five through seven are named and
NOT built, which is the honest state.
from an STFT with no transcription. Two lanes independently named it the cheapest high-value
addition available, and it is the only thing that would let us grade "harmonies and clashes" on a
render at all. Built this sitting inside the noise-versus-structure instrument.
twelve circular rotations, with the winning rotation recorded so the INTERVAL of the return is
itself reported. This turns "the theme comes back in a new key" from invisible into a measured event
with a number attached. Built this sitting inside the hook-presence instrument.
from audio is corroboration whose error must be published beside every number it produces. Built
this sitting across the layer-census instrument and the pass-2 manifest.
needs anything else, and the period must be cross-checked between methods rather than asserted from
one. Built this sitting inside the cycle-law instrument.
2's resolution: take every pair of sections whose repeat similarity is 0.95 or higher and ask
whether the set of sounding instruments differs. If corpus rows re-orchestrate across identical
music more than their siblings do, the reconciliation is confirmed rather than inferred. It needs a
real instrument roster — source separation or a trained recogniser — and it is the highest-value
measurement this pass could not build.
mode and raga and qenet identification, and expanding the anti-plagiarism signature set past its
fourteen declared gaps. Polyphonic transcription of a forty-layer orchestral render will not become
reliable soon; separate-then-transcribe-per-stem is the honest partial path, with the caveat that
the available separators are trained on popular-music stems and orchestral material is out of their
distribution.
that restores what it took. Our card carries every component — raises[].end_s, drops[].at_s,
bars_held_s, re_enters_at_s — and nothing computes the PAIRING. It is the highest-value derived
feature available from data we already have.
declares its cycle period, its novelty schedule, its layer manifest with entries and exits, and its
hook statement and return times; the instruments measure the render; the two are compared. A
prototype that produces a good mixdown by abandoning its plan is a FAILURE of this architecture even
if it sounds better, because the thing being proved is that the structure survives realisation.
allocated so no two parts in a stratum fight for the same band.
and about three of them large.
its care line quoted, and the menace, where a cue has any, carried by material that is not of the
place.
structure read. The honest tier at this rung is STRUCTURE. A CC0 sampler orchestra is not a live
session and must never be labelled as one.
Josh, 2026-08-07 night, in the round-2 floor addendum: "I like more upbeat." His taste runs
energetic and adventurous — the classic JRPG overworld optimism. Under the standing law that is a
floor to enhance rather than a preference to accommodate, so it lands here as a REGISTER DEFAULT
per purpose class.
It was measured before it was applied, and the measurement stopped a naive over-correction that
would have made the score worse. Across the thirty corpus rows — the tracks Josh named as loved for
decades — the register reads:
| purpose class | n | major-mode share | median bpm | median spectral centroid Hz |
|---|---|---|---|---|
| EXPLORATION | 8 | 4 of 8 | 114.9 | 1845 |
| BATTLE | 5 | 0 of 5 | 117.5 | 2575 |
| TITLE_CHARACTER | 4 | 0 of 4 | 117.7 | 2277 |
| MELANCHOLIC | 4 | 2 of 4 | 136.4 | 1665 |
| CREDITS_TRIUMPH | 4 | 4 of 4 | 99.4 | 1977 |
| TENSION | 3 | 1 of 3 | 117.5 | 1843 |
| BOSS | 2 | 2 of 2 | 126.1 | 1551 |
| ALL | 30 | 13 of 30 (43%) | — | — |
THE FINDING, and it is the reason this section exists rather than a one-line instruction to write
in major: the music Josh loves is NOT predominantly major. It is 43% major overall, BATTLE is zero
of five, and TITLE_CHARACTER is zero of four. A pass that answered "more upbeat" by defaulting to
major keys would have moved AWAY from the corpus on two of its most energetic classes, and it would
have been done in Josh's name.
So UPBEAT IS NOT A MODE. In the corpus it is carried by three other things, and BATTLE is the proof:
entirely minor, and the BRIGHTEST class in the whole corpus at a median centroid of 2575 Hz against
an all-class spread of 1551 to 2575, with driving tempo and continuous forward motion. That is what
energetic sounds like in the music that stayed loved. The register is ENERGY, BRIGHTNESS and FORWARD
MOTION, and mode is nearly orthogonal to it.
actually names, it is the class most of our region cues belong to, and it is the one where the
corpus most supports him: half of it is already major and its tempo sits mid-band. Default to a
major or bright-modal centre, forward motion in the accompaniment, and a melody that climbs. This
is the strongest single application of his note.
it; the corpus is unanimously minor here and unmistakably upbeat anyway.
scale and ambition rather than by brightness of key.
Slowest median tempo of any class at 99.4 bpm, which says triumph is BROAD rather than fast.
upbeat is a lament destroyed. Josh's note is a default, not a floor to apply everywhere, and the
class whose entire purpose is sorrow does not receive it. Our own Home and Loss cue is in this
class.
and non-Western material — several of our slice cultures are not in a major/minor system at all,
so "major share" for a gong-chime or pentatonic cue is a category error rather than a measurement.
Read the mode column as a weak signal and the centroid and tempo columns as the strong ones.
no per-class number here has a calibrated multiplicity null.
the ABSOLUTE centroid figures are not directly comparable to ours. The ORDERING across classes is
the usable part.
DIRECTOR RULING, 2026-08-07, applied rather than flagged (methodology is creative-within-vision;
the path is shown here so it can be argued with). It supersedes the working assumption that the
formula card is where music gets graded.
Five measurement instruments now exist against Josh's floor, every one of them built with real
controls and mutation arms, and NOT ONE separates the music he loves from the music he rejects.
AUC 0.501. Eight threshold settings swept; none separates.
it cannot adjudicate its own clause at all.
two independent research lanes as the highest-value measurement available to this program, came
back at AUC 0.4963. Its one apparently-strong separator collapsed to 0.5208 when held to the
lossless stratum and was withdrawn as codec provenance.
the WRONG side of chance — those tracks carry more restatement and better textbook hook geometry
than the corpus.
counter-melody presence lands at AUC 0.195, meaning the rejected round scores HIGHER.
Against that, the SCORE-side census separated round 2 instantly, trivially, and correctly: the
entire melodic content of every cue is one six-note two-bar cell, three genuine development
operations across four minutes, no phrase longer than two bars, no cadence anywhere. That census
took one pass over a JSON file. It is exactly what Josh heard, and it needed no DSP at all.
authored symbolically before it is played, so for our own output the note list is available,
exact, and free. Melodic complexity, development operations and their chain depth, phrase
grammar and cadences, counter-melody independence, voice-leading, range and playability, the
ornament layer, and the layer count are ALL exactly computable from the score. Measuring them by
extracting a monophonic line from our own finished dense mix — and then reporting the extraction
error as the headline caveat — is measuring a thing we already know, worse.
and become the RENDER-FIDELITY CHECK: does the mixdown express the structure the score declares?
That is the question they are actually good at, it is the question the pass-2 plan-versus-render
contract already asks, and it is the one place their measured error is tolerable because both
sides of the comparison are ours.
audio is the only way to measure them and every corpus number stays audio-derived. What changes
is the inference direction: the corpus sets BANDS and shows what the masters' structures look
like; it does not certify our cues. A corpus band is containment, never a target — a rule this
program already learned once and now applies to a second class of number.
hook presence's negative axes are reported as measurements and are not permitted to grade round 3.
Tuning an inverted axis until it points the right way is fitting to three data points.
listener knows whether it works, and nobody in this program has heard any of this music. Josh
remains the first ear and the honest tier stays STRUCTURE.
and they bound where the defect is NOT, which is why round 3's defect had to be melodic: the
structural instruments had already cleared that ground.
corpus. Their calibration comes from the craft literature and from our own before-and-after, and
that weaker warrant must be stated wherever they are quoted.
checked against a human ear, and the program's own record is that the ear is the thing it keeps
failing. Josh is the first ear and remains so.
found craft consensus rather than evidence it labelled it, and those labels were preserved here.
regions, and the reasons differ by region — Bali fails on tuning, West Africa on mechanism (the
dundun's defining property is a continuous tension glide no fixed-pitch sample produces), Flores on
simple absence. The single highest-leverage acquisition named is a CC0 voice set, whose absence
alone blocks four nodes.
arithmetic on fields we already have. None of them measures whether the music is beautiful. That
remains outside every instrument in this program, and every artifact it produces says so.