COMPOSED_TRACKS_ARM_A.md

music/complete_tracks/COMPOSED_TRACKS_ARM_A.md

COMPOSED TRACKS — ARM A, the ten odd-indexed cards played

CANON SUBORDINATION — this document is PROPOSAL-TIER: it serves canon and never outranks it.
The canon: the CVD · the T1 foundation docs · the T0 registries · the spine · the region pages.
Authority order: docs/DOC_MAP.md § 0. **If this document disagrees with canon, CANON WINS and
this document is the defect.**

Emitted by harness/music_gen/complete_track.py. Measured against

build/audio/complete_tracks/SAMENESS_BASELINE.md, which is the arithmetic of Josh's ruling.

---

0. What was owed

RECOMPOSE_RUNG_TWO_FINDINGS.md § 5.3 named the one thing the identity cards did not deliver:

"F8 is NOT discharged. A specification is not a sound." The cards specified twenty tracks with

zero variety-matrix collisions. Nothing had been composed or played.

This arm composes and plays the ten ODD-indexed cards, end to end: score, orchestration, render,

records, teeth.

---

1. The ruling answered on its own axes

The baseline measured the shipped cohort on the axes a listener hears in four seconds. The same

axes, measured on this arm's ten renders:

AxisShipped cohort (15 tracks)Arm A (10 tracks)
Composed sections per track1 distinct value, 15/15 — no B section anywhere4 distinct, 5–9, mean 7.4
Durationmean 29.3 s, 15/15 under a minute, 0/15 over 150 smean 251.5 s, 10/10 in the 150–300 s band
Instrument signature4 distinct, modal share 12/1510 distinct over 10 tracks
Voice / part count2 distinct, modal 4 at 14/155 distinct, 7–19 stems per track
Time signature14/15 in 4/44 distinct — 4/4, 12/8, 7/8, 6/8
Dynamic range (p95−p10)mean 9.1 dBmean 23.5 dB, range 14.6–33.9
Crest factormean 14.2 dBmean 21.6 dB
Percussionnone anywhere in the cohortpresent on 8 of 10
Key10 distinct over 10

Duration is 8.6× the cohort's. Dynamic range is 2.6×. Every track has a real form.

1.1 The ten, as rendered

CardDurationSectionsMetreBPMStemsDynLongest verbatim run
HF_MT_01_TITLE221.5 s74/41041824.2 dB0 bars
HF_MT_03_BATTLE_ORDINARY157.9 s74/41521614.6 dB0 bars
HF_MT_05_BOSS_APEX240.0 s94/41761933.9 dB6 bars
HF_MT_07_TRADE_AND_FESTIVAL261.8 s812/81321331.2 dB3 bars
HF_MT_09_WONDER_DISCOVERY240.0 s54/472721.4 dB2 bars
HF_MT_11_REGION_FLORES243.8 s84/41261319.3 dB1 bar
HF_MT_13_REGION_SUMATRA_JAVA282.4 s74/468818.0 dB1 bar
HF_MT_15_REGION_SOUTH_INDIA282.7 s87/81041322.3 dB1 bar
HF_MT_17_REGION_ETHIOPIA293.7 s76/8761124.3 dB2 bars
HF_MT_19_REGION_WEST_AFRICA291.6 s812/81581525.4 dB0 bars

---

2. The stack gap, closed — because half the mechanism was not compositional

§ 5.1 said it plainly: the landed 23-instrument stack held no percussion, no choir, no trumpet, no

tuba and no period instrument, so part of "one type of sound" was what the renderer could

physically play.

harness/music_gen/extend_stack_wave2.py added 71 instruments, 1928 files, 1.77 GB, at the SAME

pinned CC0 commits as rung one, every byte hashed into

build/audio/stack/sfz_stack_extension_wave2.json:

rope-tension snare, toms, clash and suspended cymbals, gong, tambourine, triangle, woodblock,

claves, shakers, cabasa, agogo, cowbells, finger cymbals, anvil, brake drum, claps, bell tree,

slit drum.

vibrato.

plus the FULL-dynamic section sustains. The cohort's modal signature was three QUIET sustains

plus a carrier: one articulation.

upright piano, piccolo, staccato winds.

strumstick, plucked psaltery.

2.1 The care line was NARROWED, not loosened

sfz_palette.py excluded VCSL's world-instrument half on a reason that was correct then: the region

cells that would carry those cards "are explicitly NOT COMPOSED yet". Those cells now exist as

cards — so the rule adopted here is narrower than "the cards exist, load everything":

A VCSL instrument carrying a NAMED LIVING TRADITION in its own file name (Kalimba Kenya, Mbira
dzaVadzimu Zimbabwe, Nyunga Nyunga Mozambique, Balafon, Darbuka, Dan Tranh, Didgeridoo, Nepalese
hand bells) is loaded ONLY for a card whose region page attests that instrument class FOR THAT
REGION.

None of the five region cards in this arm attests one, so none was downloaded. Using a

Zimbabwean mbira to score Ethiopia, or a Middle-Eastern goblet drum to score Tamil Nadu, is a worse

care outcome than absence: it puts a real tradition's instrument in another tradition's mouth and

calls it research. The region cards are served by the generic organological class the card itself

names — "a graded drum family", "struck iron bell", "a large floor harp", "double-reed lead" — and

every substitution is written onto that track's generation record with its entry condition.

That is the cards' own instruction; Ethiopia's card says the instruments are named "in their generic

organological form so a builder sources or models them rather than lifting a recording."

2.2 The choir gap is declared, never faked

Neither sample set contains a human voice. Verified against both pinned trees. Two arm-A cards

name a choir or a solo voice. Both compose those parts as written and render them on a NAMED

stand-in, recorded per track, so the gap stays visible until a CC0 voice set is sourced. Six of ten

tracks carry declared substitutions.

---

3. How the composition works — the card's prose is the instruction sheet

A card that says "brass silent", "four bars of silence written into the score", "the flute drops out

entirely", "the timeline displaces by one stroke", "the cycle shortens by two beats each repetition"

has already composed the architecture in prose. A generator that ignored those strings and applied

one template to twenty cards would reproduce the exact defect Josh named, with better paperwork.

So section_orders() parses them into 23 recognised orders — SILENCE, RECEDE, SOLO, TUTTI,

FRAGMENT, AUGMENT, INVERT, OCTAVE, OSTINATO, COUNTER, PERCUSSION, RUBATO, DISPLACE, MODE_SHIFT,

REHARMONISE, SEAM, ESCALATE, LIQUIDATE, DRONE, THIN, RESCORE, ENTER, BEND, SPOKEN — and **every

order read is written onto the artifact**, per section, so a reader can check the score against the

card without trusting the code.

Above that: Caplin sentences rather than tiles (basic idea — varied repetition — fragmentation —

cadential descent); harmony per bar with modal mixture as the default colour source; a weakened

cadence at every loop point (PHRASE-05, the anti-fatigue law); development by inversion,

fragmentation, augmentation, liquidation, metric displacement and mode migration; and a motif set

whose four-motif floor is measured (worst-pair unlikeness ≥ 0.55 on a combined interval and

rhythm edit distance), not counted.

3.1 Dynamics are authored in two layers, and both are in the score

Layer one is per-note velocity — metric weight, phrase arch, section crescendo, and a small human

inconsistency. Layer two is a composed section-by-section envelope, interpolated across bar lines

and written into the MIDI as CC11 expression so the notation carries it.

The envelope carries 62% of the card's target and density plus velocity carry the rest, because

most of a real orchestra's dynamic range comes from how many players are playing and how hard. The

first title render put its solo horn intro 36 dB under the tutti — wider than any orchestra

plays — because three reductions were compounding. That is a fix with a number, not a taste.

---

4. The teeth, and the five defects they caught

--selftest is 35/35, and the anti-tiling tooth is mutation-proved in both directions: forced

to FIRE on a deliberately tiled score and to read CLEAR on a genuinely varied one.

T1 anti-tiling. Reduces every bar to a signature over (part role, pitch, onset-in-bar,

duration). **Velocity is deliberately excluded, so "the same bar played louder" cannot launder a

repeat.** Refuses any contiguous run of 8+ bars occurring twice — the exact defect

RECOMPOSE_RUNG_TWO_FINDINGS § 1.3 hand-verified ("bars 9-16 are a verbatim repeat of bars 1-8").

T2 duration in the 150–300 s band and within 3 s of the card's derived target. T3

instrumentation family resolves. T4 four-motif floor with unlikeness measured. T5 section

count matches the card. T6 ≥ 5 orchestrated roles. T7 measured dynamic range, on the

rendered audio, against a card-relative floor.

4.1 Five defects found by reading the output rather than the metrics

1. The melody left every second bar EMPTY — a 2.5-beat head cell inside an 8-beat Caplin idea

wrote 2.5 beats and stopped. Bars 9, 11, 13 and 15 of every sixteen were silent in the lead.

RHYTHM-04 says rests are structural; five and a half beats of nothing is a hole. Fixed by

_fill_idea, which occupies the idea with an ECON-06 answer and leaves one pulse of rest.

2. The three inner voices moved across THREE SEMITONES for an entire piece. Nearest-tone voice

leading with no register band collapses the frame into a pad. Fixed with a band per voice.

3. A trumpet doubled the horn for all 96 bars — including the eight the card scores as "horn

call alone, no accompaniment". A doubling is a section decision; it is now gated to the sections

whose orders ask for weight.

4. A returning section re-STATED its motif instead of re-scoring it — sections 0 and 3 of the

wonder cue drew the same motif under the same SOLO order and came out bar-for-bar identical, a

10-bar verbatim run. Caught by T1. Fixed by the cards' own doctrine ("A_return re-scores A

rather than restating it"): a reused motif returns transposed, which LEIT-03 licenses because

the interval spine is the invariant.

5. A section's second 8-bar sentence restated its first — the cohort's defect reappearing one

level down, and § 1.3's finding exactly ("L1's own citation reads the exemplar's A section as TWO

8-bar SENTENCES, not one 8-bar unit stated twice"). Caught by T1 on the South India cue. Each

successive sentence now takes its own pitch level and continuation treatment.

Two further calibrations were made honestly rather than by moving a pass line: the dynamics floor is

card-relativemax(11, min(15, card_target − 2)) — because a flat 15 dB floor failed the

ordinary-battle cue at 14.7 against a card that ASKS for 14, and a driving-combat cue is

deliberately compressed. And SILENCE was split from RECEDE, because conflating "four bars of

silence written into the score" with "the working sounds take over; the score gives way" had put 32

seconds of dead air in the festival cue's break. On that card's own instrumentation the working

sounds are the anvil and the drums — so the section is scored for them.

---

5. Records, licence, provenance

Records paired 10/10; each record's sha256 matches its WAV byte-for-byte (verified). Reveal

firewall: 13 held tokens scanned across all 20 record files, 0 hits — the pattern fragment stays

unnamed and no record names the House parent theme.

lane redistributes rendered audio only.

other generator.**

played by sfizz from notes composed in complete_track.py.

recorded orchestra and no line of any record claims it is.

under the same contract as every other audio lane, and the audio re-renders from the tracked

notes.

Gates: 9/9 music gates PASS. The three suite failures (boards, ledger, line_citations) are

pre-existing and read nothing under build/audio or docs/proposals/music.

---

6. What is still owed

artifacts, not a listening verdict. A specification was not a sound; a measurement is not one

either. The listening pass in front of Josh remains the gate.

discipline, would retire six of the declared substitutions.

stand-ins. The honest resolutions are sourcing or modelling those instruments, or a performance

licensed by the community that holds it — not reaching for a differently-attributed sample.

reasons and entry conditions are on the record in cards/INDEX.json.

RECOMPOSE_RUNG_TWO_FINDINGS § 2 states them.

extend_stack_wave2.py; they are not written here because this lane's declared write scope is

build/audio/ and docs/proposals/music/.

Generated by harness/site/structure_site.py — the URL path is the repo path. review root