THEME_PROMPT_SET_R1.md

music/THEME_PROMPT_SET_R1.md

THEME PROMPT SET R1 — the Flores Highlands slate, twelve crafted prompts under Josh's arc grammar

CANON SUBORDINATION — this document is PROPOSAL-TIER: it serves canon and never outranks it.

Canon served: Josh's 2026-08-09 rulings (docs/spine/DECISIONS_PENDING_JOSH.md) — AIVA Pro

enters the stack, composed-never-prompted is DISSOLVED, the four rounds' composed themes are

RETIRED as source material, the prompt-only engines become THEME GENERATORS; **the ARC GRAMMAR he

dictated live the same day, which is now law for every prompt here**; the melody-first / hummable

thirty-year bar; the ENERGY FLOOR; CVD §17.1 care and the §17 floor; the Flores region page's own

Section 10 and its named prohibition.

If this document disagrees with canon, CANON WINS and this document is the defect.

TIER: PROMPT SET — PROPOSAL. ENGINE, not CANON. It sets no canon, names no new region content,

changes no spine or registry row, and buys nothing. It is HOW, never WHAT.

Written 2026-08-09; revised the same sitting after Josh graded an ElevenLabs v2 round live.

Consumer: the AIVA + ElevenLabs browser session the director drives in Josh's Chrome. Scoring

consumer: harness/music_gen/benchmark_judge.py → build/audio/benchmark/SLATE_R1.json.

---

0. THE ARC GRAMMAR — the law every prompt below is built around

Josh graded an ElevenLabs v2 round with the engine in front of him. **He loved it, and he named

exactly how it overcorrected.** The correction is not a preference note; it is the shape of every

prompt in this document.

0.1 What v2 got right, and must keep

• The hook lands inside three seconds. No slow-burn opening.
• No silent intro and no silent outro. The file starts sounding and ends sounding.
• A real B-section — a second idea, not one idea with new paint.
• Drop-and-surge as a device.

0.2 What v2 got wrong, in his words

**Too heavy throughout. No valleys. The melody never settles and never passes between instruments.

Bored by 0:60.** Wall-to-wall intensity is the failure. A drop only reads as a drop if it lands

inside a landscape that has somewhere lower to go.

0.3 THE FOUR MOVEMENTS — law, in order, in every cue class

1. Introduce the melody ALONE first. One instrument, the tune, immediately — inside three

seconds, no pad-and-swell intro.

2. Then introduce the beat. The rhythm arrives under a melody the listener already knows.

3. Then add, layer and change — and this is the movement v2 skipped. The melody is **passed

BETWEEN instruments, and on each return it is varied**: up or down in register, soft or

strong in weight. Between the returns there are **real quiet valleys where one instrument holds

the tune alone**. Never heavy for long stretches.

4. Settle back to the melody at the end, in a calm resolved phrase that **hands off naturally

to the next track**.

0.4 How the grammar BENDS per class

The grammar is the same four movements everywhere. What changes is the depth of the valleys, the

hardness of the surges and the length of the settle.

Read the movement-1 row as a DURATION, not a start time. The tune always begins inside three

seconds — that is the v2 win, and it is non-negotiable. The number below is how long it stays alone

before the beat arrives.

One clause governs the whole table: *never heavy for long stretches.* It is the sentence that

turned a loved round into a rejected one, and it applies to the serene class as much as the boss

class — heaviness there is a different thing (density, not volume) and it bores just as fast.

0.5 The honest note this section owes

Two of the four movements have no instrument. instr_dynamic_arc reads the build, the holds and

the valleys, so movements 1–3's *dynamics* are measured. But **nothing in the battery reads WHICH

instrument is carrying the tune** — the melody extractor returns one predominant line and folds

octave doublings — and nothing reads whether the final phrase resolves. So **carrier migration

and the settle are graded by ear only**, on every row, and the judge prints that sentence on every

card rather than letting a green wall verdict imply otherwise

(benchmark_judge.UNMEASURED_BY_ANY_INSTRUMENT).

---

1. Where this sits on the ladder, and the compliance statement it owes

Josh's ladder, verbatim from the ruling: **prompt-engines generate candidate themes → the judge wall

and Josh's ear select the hummable winners → the winner moves into AIVA Step by Step for development

into full cues → pre/post iteration to the review page.** This document is the FIRST rung only.

THE RETIRED-THEMES COMPLIANCE STATEMENT. Josh: *"Remember our melodies and crap you've already

created sucks."* Not one prompt below carries any of the four rounds' tunes, head cells, motifs,

section maps, hook plans or cue names, and no prompt uploads a round-1..4 render as an audio

influence. What the prompts DO carry from our own tree is three things, none of which is a retired

melody: the region's documented dialect (the gong-waning and its interlocking-cycle structure,

from the region page's own Section 10), the corpus's measured bands (percentiles of the thirty

carded tracks Josh named), and Josh's graded axes turned into prompt language.

A PROMPT IS NOT A CUE. Every candidate here is 60–240 seconds of finished stereo audio with no

stems, no score, no layer manifest and no adaptive states. The cue architecture — swap layers,

colotomic clock, novelty schedule, per-class room — is untouched and is where a WINNER goes next.

Judge these as themes, on the tune.

---

2. THE SHARED STEERING BLOCK

2.1 The two identities every single prompt must state

Josh's v2 round lost both of these, and the loss is named as a defect rather than a nuance:

• FANTASY ACTION-ADVENTURE RPG. Not "cinematic", not "world music", not "trailer". The frame is

a big-budget fantasy action-adventure role-playing game, and it belongs in the first sentence.

• FLORES ISLAND. The place is named — the volcanic highlands of Flores in eastern Indonesia —

in every variant including the score-first arms, where it is named as the SETTING even when the

palette is orchestral. A prompt that drops the island produces a theme for nowhere.

Both appear in all twelve prompts below. If a variant is re-rolled or reworded in the browser, these

two are the clauses that may not be dropped to make room.

2.2 The region, and what it is NOT

• The ensemble is the Flores gong-waning — the tuned gong-and-drum ensemble of the Manggarai

highlands, central to the caci ceremony and canonised as the region's Thread-21 vril-active

instrument (_source/02_Tier_2_Region_Pages/flores_island.md §10). The theory named there is the

interlocking-cycle structure and the ceremonial drumming that scores the caci rounds.

• It is NOT Balinese gamelan. Bali is Ch 4 with its own card and its own kotekan architecture.

Writing "Balinese gamelan" into a Flores prompt is wrong about the island and collapses two

distinct traditions into one exotic wash — the generic world-music wash the care doctrine names as

a failure. The prompts say tuned bronze gongs and interlocking hand drums of Flores, and never

name Bali or Java.

• The page flags its own substrate as thin on the caci contest detail and gong-waning theory. So

the prompts describe acoustic behaviour — a repeating cycle marked by the largest gong, two

drum parts that interlock rather than double, a bamboo flute over the top — and never claim a

repertoire, a tuning system or a ceremonial form the page does not attest.

2.3 The care line, expressed as things a prompt may never contain

• No consecrated performance as a target. No prompt asks for a recording, a sample, a rite, a

liturgy or a named sacred repertoire. Reference-composition only — the page's own prohibition,

binding a prompt exactly as it binds a composer.

• No living named performer, ensemble or composer as a style target. The exemplar references in

§3 name works, as spirit references, never people.

• The BATTLE and BOSS cues are CONTESTS, not menaces. Caci is a ceremonial duel with a ring, a

crowd and a kept count. The region's own register may never be the threat signal — the darkening

operations belong to the antagonist material and travel with it, never with the place (CVD §17.1

collective protection; T0_Theme_Registry :: REGION_FLORES_ISLAND). So these prompts read as

high-stakes ceremonial contest with real danger, never as horror, dread or evil.

• No reveal leakage. Ch 2–3 is fifteen chapters before the antagonist pattern may be perceived.

No conspiracy, no hidden power, no force behind events.

• Instrumental only, always. A vocal line here would be an invented ceremonial text, squarely

inside the prohibition. Every prompt says so.

2.4 The cue classes, with the corpus's own numbers — verified at HEAD

**Read the BATTLE column carefully — it inverts the intuition and it is the most useful number

here.** The five battle tracks Josh named carry a *sparse* extracted lead — 0.42 notes a second at

the median, a fifth of the exploration rate — with *long* phrases (7.5 s) and the *highest* share of

the lead doubled in parallel thirds or sixths. That is a riff stated in harmony, not a shredding

line. A battle prompt asking for "fast intricate melody" aims away from his own reference set; one

asking for a hard, simple, repeated riff doubled in thirds over a driving pulse aims at it. It

is also exactly the shape the arc grammar wants, because a riff is what you can hand between

instruments.

Read SERENE and BOSS as declared coverage holes. SERENE has no corpus rows at all: its melodic

axes resolve to the whole corpus (battle cues included) and its loudness has two defensible borrows

that disagree by 2.4 LU. BOSS has two carded rows, under the melodic instrument's own n≥5 floor, so

its melodic axes come from a pool of five small classes while its loudness comes from two tracks.

The judge prints both notes on every row. Closing these is a corpus-acquisition job, not a prompt

job — and it is the strongest argument in this document for buying serene and boss exemplars next.

2.5 The grades, turned into prompt language

---

3. THE EXEMPLAR INFLUENCE VOCABULARY — the 114-track roster as style references

build/audio/exemplars/CORPUS.json carries Josh's roster: **114 original rows across seven theme

buckets, plus a 68-row expansion (182 total). The 114 originals** are the vocabulary used below,

because they are the list he wrote.

HOW A TITLE IS USED, AND THE GUARDRAIL. A title is named as a spirit reference — the feel,

the energy, the idiom — never as a copy instruction, never as "recreate", and never attached to a

composer's name. Two things make that safe rather than hopeful: the prompt language is always *"in

the spirit of"* plus a description of the QUALITY being borrowed, and every downloaded track is run

through the judge's quotation screen, which compares the extracted line's interval sequence

against corpus hook signatures and REFUSES a contiguous match. The screen's zero is controlled — a

partial seed set, so CLEAR is not a clean-room certificate — but a regurgitated hook has a real

chance of being caught before Josh ever hears it, which is more than a prompt-only workflow usually

has.

3.1 The buckets, and what each lends

The seven buckets of the 114, as the roster itself carries them: EXPLORATION 30 · TITLE_CHARACTER 14

· BATTLE 17 · BOSS 15 · MELANCHOLIC 14 · TENSION 13 · CREDITS_TRIUMPH 11.

3.2 The picks, per class

• EXPLORATION: *Gerudo Valley* (EX_023) — the propulsive folk-modal tune over a driven pulse,

the single closest structural cousin to a highland road cue. *Wind Scene* (EX_022) — the wide lift

and the handoff between solo voices. *The Fields of Ard Skellig* (EX_015) — folk instruments

carrying a big tune without becoming wallpaper. *Terra's Theme* (EX_032) and *Corridors of Time*

(EX_033) as alternates when a variant needs more melancholy warmth or more cyclic hypnosis.

• SERENE: *Kokiri Forest* (EX_024) — calm with a constant light pulse, the exact "peaceful and

still uplifting" target. *Aquatic Ambiance* (EX_028) — slow tune over a shimmering moving surface.

*Goddess Pond* (EX_044) — small, warm, unhurried, and never sedative. *Song of Healing* (EX_090)

only where a variant wants tenderness, and never as the lead reference, because the class is not

a lament.

• BATTLE: *Battle on the Big Bridge* (EX_051) — the riff that everyone can hum, handed around

the ensemble. *Those Who Fight* (EX_060) — drive with a real tune on top. *Rivers in the Desert*

(EX_048) — modern surge and release. *Silver for Monsters…* (EX_052) as the folk-instrument

alternate when a variant needs the region palette to carry the fight.

• BOSS: *Hopes and Dreams* (EX_072) — the theme returned at maximum weight, which is the whole

boss-class idea. *Sealed Vessel* (EX_073) — phased escalation with air in it. *Dancing Mad*

(EX_075) — movements that restate rather than merely intensify. *Soul of Cinder* (EX_066) as the

heavier alternate.

3.3 The stronger channel, and it is ROUND 2

ElevenLabs Music takes an Audio Reference upload of roughly 30 seconds to guide style and sound

— a far stronger influence channel than any adjective. It is deliberately NOT used in round 1,

for two reasons worth stating rather than assuming: uploading a commercial recording makes the

output a derivative of that recording and puts a provenance question on a candidate that does not

need one, and mixing a prose arm with a reference arm in the same round would confound the prompt

comparison this document is built to run.

Round 2 is where it earns its place, and the design is already clear: take the prompt that

grades best in round 1, re-run it **with a 30-second reference clipped from the matching corpus

track**, and compare. That is a controlled single-variable test of the reference channel itself. The

provenance question is answered the same way it is answered for every judged-only artifact — the

output is evaluation material, it is scored, and **nothing enters a build pack until the engine's

licence posture clears**, which today it does not for any of them (§6). The vendor's own warning

that Output "may be similar or identical to Output returned to other users" is a second reason the

quotation screen runs on every reference-conditioned track.

---

4. THE EXPERIMENTAL DESIGN — twelve, and what each is FOR

Twelve prompts, not twelve guesses. The arc grammar is now constant across all of them (it is

law, not an arm), so the lattice runs over the axes that remain open, and every arm has an isolation

partner that differs on exactly one.

The decisive comparison is B, because of one sentence in the round-4 grade: *"I hate the

instruments you are using."* It has two readings — our CC0 sampler sounds bad, or the region palette

is not what he wants under an ARPG — and they lead to opposite programs. TP-E1 (region-first), TP-E3

(score-first) and TP-E4 (fusion) are the same intent, same tempo, same arc grammar, three palettes.

Three grades on those three rows answer what four rounds of building could not.

---

5. THE ENGINE REGISTERS

5.1 ElevenLabs Music v2 — a natural-language prompt engine

VERIFIED from https://elevenlabs.io/docs/overview/capabilities/music and its best-practices

page, fetched 2026-08-09:

• Inputs: a natural-language text prompt, an Audio Reference of roughly 30 seconds, and

section-by-section composition specifications (composition plans).

• Vendor prompting guidance: include genre/mood, tempo as "130 BPM", key as "in A minor",

length, and expressive descriptors (raw, live, breathy, aggressive). Add **"instrumental

only"** for no vocals. The model "accurately follows BPM and often captures the intended musical

key". It warns that **prompt length and detail do not always correlate with better quality

outputs** — which is why axis A exists rather than being assumed.

• Output MP3 (44.1 kHz) or WAV, 3 seconds to 5 minutes.
• NOT VERIFIED and therefore not used: the composition-plan JSON field names (the plans document

404s at its documented path). Where a prompt needs a section plan it is written as **timed prose

inside the prompt**, which is the register the best-practices page actually documents.

5.2 AIVA — a style-and-settings engine, not a prose-prompt engine

VERIFIED from the official user manual, fetched 2026-08-09: four modes — From a Style

(presets or a custom Style Designer creation), From Chord Progression (style, key, time

signature, tempo, duration, number of compositions), Step by Step (*"turn a specific melody, an

accompaniment, percussion rhythm or even an instrument into a full composition"*), and legacy

Presets. MIDI influences require 8 bars minimum; audio influences 60 seconds minimum. The

editor carries per-track mute/solo/octave, filters, reverb, delay, Auto Staccato, section

insert/delete with measure counts, a BPM edit, a percussion pattern editor, and **dynamics

automation curves**.

THE CONFOUND, STATED BEFORE ANY RESULT. AIVA has no free-text style prompt. Its style control is

a preset or a Style Designer creation plus a settings row, so **engine and prompt-wording are

confounded by construction. The mitigation is the only honest one: the same INTENT** runs through

both registers, the AIVA column is a settings row, and a grade attributes to the INTENT first and the

engine second. Nobody may read "AIVA beat ElevenLabs on TP-E1" as a prompt-wording result.

AIVA's section editor is where the arc grammar is enforceable rather than requested. ElevenLabs

is asked for the four movements in prose and may or may not comply; AIVA's editor lets an operator

build them. That asymmetry is itself worth grading, and it is the reason the movement-sensitive

classes (BOSS especially) are run on both.

**Round 1 asks AIVA the question every engine gets. Its unique capability — Step by Step, the only

mode in the commercial field that takes a supplied melody — is the DEVELOPMENT rung, used on the

winner, not here.**

5.3 The other engines

Lyria 3, Mureka and Suno are judging-only comparators with unresolved or disqualifying terms

(CURRENCY_SWEEP_2026-08.md §5.1). The ElevenLabs prose block is the right register for them, and

the judge already carries their licence posture. Nothing from them may reach a build pack.

---

6. THE TWELVE PROMPTS

Each block carries an INTENT (one line, so a grade attributes), its axis, the **ElevenLabs

prose prompt as it should be pasted, and the AIVA settings row. Run three variations per

prompt per engine**, keep the best one or two, download to

D:/audio/benchmark/<engine>/<PROMPT-ID><variant>.<ext>, and judge.

Where a prompt names a key it names D major with a pentatonic instruction in prose: the plans'

pitch set is D anhemitonic pentatonic (D E F♯ A B), and neither engine takes a scale as a setting.

---

EXPLORATION — the road that keeps count

TP-E1 — region-first, tight

• INTENT: the fully specified region prompt under the arc grammar, at our own plan's tempo. If a

crafted, fully constrained region prompt cannot produce a hummable tune, constraint is the problem.

• Axis: A=TIGHT · B=REGION-FIRST · E=acoustic.

Instrumental only, no vocals. A 116 BPM exploration theme for a big-budget fantasy action-adventure
RPG, set in the volcanic highlands of Flores island in eastern Indonesia. In the spirit of Gerudo
Valley and Wind Scene: a propulsive folk-modal tune with room to breathe, never wallpaper.
Structure it exactly like this. Start with the melody ALONE in the first three seconds - a solo
bamboo flute, no drums, no pad - playing a clear, singable, memorable tune in a major pentatonic
scale on D, in phrases of about two to three seconds. Then bring in the beat: a tuned bronze gong
marking a repeating four-bar cycle and two interlocking hand drums that never land on the same beat.
Then layer and change - hand the SAME melody between instruments, flute to strings to plucked
bronze, varying it each time, higher then lower, soft then strong, with real quiet valleys of
fifteen to twenty-five seconds where one instrument holds the tune alone before the ensemble comes
back. Include one contrasting second section with its own idea. Give it one big return where the
melody is doubled in thirds by the full ensemble. Then settle back to the melody at the end, calm
and resolved, so it hands off naturally to the next track. Real chord changes you can hear
throughout. Warm, bright, forward-moving, hopeful, in motion. Never heavy for long stretches. Live
acoustic players in a real room. No silence at the start or the end.

• AIVA: From a Style → Style Designer, influence description "acoustic world-fusion adventure,

tuned gongs and hand percussion with orchestra". Key D major. 4/4. Tempo 116. Duration

3:30. Compositions 3. Then use the section editor to enforce the four movements.

TP-E2 — region-first, loose

• INTENT: the isolation partner for TP-E1 on axis A alone. The vendor's own guidance says detail

does not always help; this tests it on our material.

• Axis: A=LOOSE · B=REGION-FIRST · E=acoustic.

Instrumental only, no vocals. The main exploration theme of a big-budget fantasy action-adventure
RPG, set in the volcanic highlands of Flores island - the moment you crest a ridge and the whole
valley opens up. In the spirit of Gerudo Valley and The Fields of Ard Skellig. Tuned bronze gongs
and interlocking hand drums under a soaring bamboo flute melody, memorable enough to hum after one
listen. Start with the melody alone, then the beat, then pass the melody between instruments with
real quiet stretches in between, then settle at the end. Live players, not synths. No silence at the
start or the end.

• AIVA: From a Style → the nearest cinematic-adventure preset. Key D major. Tempo default.

Duration 3:30. Compositions 3. Change nothing else — this is the loose arm on both sides.

TP-E3 — SCORE-FIRST, region palette removed

• INTENT: the decisive isolation arm for *"I hate the instruments you are using"*. Same scene,

same island, same arc grammar, region palette removed. If this grades up and TP-E1 grades

down, the problem was never our sampler.

• Axis: A=TIGHT · B=SCORE-FIRST · E=acoustic.

Instrumental only, no vocals. A sweeping orchestral overworld theme for a big-budget fantasy
action-adventure RPG, set in the volcanic highlands of Flores island - in the spirit of Wind Scene
and Terra's Theme. 112 BPM, 4/4, D major, hopeful and in motion. Structure it exactly like this.
Open with the melody ALONE inside the first three seconds on a solo woodwind - one hummable tune,
no intro swell. Then bring in the rhythm: light tuned percussion and a walking string pulse. Then
layer and change - hand the same melody between instruments, woodwind to strings to French horn,
varying it each time, higher then lower, soft then strong, with real quiet valleys of fifteen to
twenty-five seconds where a single instrument holds the tune. Add one contrasting second section.
Give it one full-orchestra return with the melody doubled in thirds and a horn counter-line
answering in the gaps. Then settle back to the melody at the end, calm and resolved, handing off to
the next track. Clear functional harmony with a real chord progression underneath throughout. Never
heavy for long stretches. Live orchestra in a concert hall. No silence at the start or the end.

• AIVA: From a Style → "Epic / Cinematic orchestral" family. Key D major. 4/4. Tempo

112. Duration 3:30. Compositions 3. Section editor for the movements.

TP-E4 — FUSION: orchestral tune, region rhythm layer

• INTENT: the hypothesis the whole slice rests on — the region lives in the RHYTHM layer and the

tune is orchestral. Stated as a hypothesis so a down-grade is informative.

• Axis: A=TIGHT · B=FUSION · E=acoustic.

Instrumental only, no vocals. A 116 BPM orchestral adventure theme for a big-budget fantasy
action-adventure RPG, set in the volcanic highlands of Flores island, whose entire rhythm section is
a tuned bronze gong-and-drum ensemble from that island. In the spirit of Gerudo Valley and Wind
Scene. Structure it exactly like this. Open with the melody ALONE in the first three seconds on a
solo oboe - one clear, singable, memorable tune, D major with a pentatonic flavour. Then bring in
the beat: a repeating gong cycle and two interlocking hand drums that never land on the same beat.
Then layer and change - pass the same melody between oboe, strings and plucked bronze, varying it
each time, higher then lower, soft then strong, with real quiet valleys of fifteen to twenty-five
seconds where one instrument holds it alone. Add a contrasting second section. One big return with
the melody doubled in thirds and a horn counter-melody answering in the gaps. Then settle back to
the melody at the end, calm and resolved, handing off to the next track. Real chord changes
underneath throughout. Bright, warm, forward-moving, never sleepy, never heavy for long stretches.
Live players in a real room. No silence at the start or the end.

• AIVA: From a Style → Style Designer, "orchestral adventure over interlocking tuned-percussion

groove". Key D major. 4/4. Tempo 116. Duration 3:30. Compositions 3.

---

SERENE — the light on the water

Under Josh's serenity ruling: *"peaceful and still uplifting… an ARPG valley still moves."* The

class bend is long settles, valleys that are most of the piece, a lift instead of a surge.

TP-S1 — region-first, lifted

• INTENT: does the "calm but never sleepy" language plus the long-settle bend actually produce a

serene cue that still carries forward energy?

• Axis: A=TIGHT · B=REGION-FIRST · E=acoustic.

Instrumental only, no vocals. A calm, luminous sanctuary theme for a big-budget fantasy
action-adventure RPG - a waterfall in the cloud forest of Flores island, where the player rests and
restores. 76 BPM, D major, pentatonic. In the spirit of Kokiri Forest and Goddess Pond: peaceful and
still uplifting, never ambient and never sleepy. Structure it exactly like this. Open with the
melody ALONE in the first three seconds on a solo bamboo flute - a clear, singable, memorable tune
in long phrases. Then bring in a soft pulse only: a slowly ringing bronze gong and a quiet frame
drum, never a drum kit. Then let it grow gently - hand the melody between the flute, a plucked harp
and a solo cello, varying it each time, higher then lower, softer then a little stronger, with long
quiet stretches where one instrument holds the tune alone. One single lift where a second voice
harmonises the melody in thirds - a lift, not a surge. Then settle back to the melody at the end,
the longest and calmest phrase in the piece, fully resolved, handing off naturally to the next
track. Warm chord changes underneath throughout, and the pulse is always present. Live acoustic
players in a real room, water in the air. No silence at the start or the end.

• AIVA: From a Style → Style Designer, "gentle acoustic world-fusion, tuned percussion and

flute". Key D major. 4/4. Tempo 76. Duration 3:00. Compositions 3.

TP-S2 — SCORE-FIRST serene

• INTENT: the palette isolation arm in the class where a sampler's weaknesses — no breath, no

bow, one dynamic layer a note — are most exposed.

• Axis: A=TIGHT · B=SCORE-FIRST · E=acoustic. Isolation partner: TP-S1.

Instrumental only, no vocals. A calm, luminous sanctuary theme for a big-budget fantasy
action-adventure RPG - the safe, beautiful place in the cloud forest of Flores island that the
player comes back to. 76 BPM, D major. In the spirit of Kokiri Forest and Aquatic Ambiance:
peaceful and still uplifting, never sleepy. Open with the melody ALONE in the first three seconds on
a solo oboe - one clear, singable, memorable tune in long phrases. Then bring in a soft pulse:
harp arpeggios and a quiet string bed, no drums. Then let it grow gently - hand the melody from oboe
to solo cello to celesta, varying it each time, with long quiet stretches where one instrument holds
it alone. One single lift where strings harmonise the melody in thirds. Then settle back to the
melody at the end, the longest and calmest phrase in the piece, fully resolved, handing off to the
next track. Real chord changes you can hear throughout. Live orchestra, real hall. No silence at the
start or the end.

• AIVA: From a Style → "Cinematic / Emotional" family. Key D major. Tempo 76. Duration

3:00. Compositions 3.

TP-S3 — fast surface under a slow tune

• INTENT: the WATERFALL successor design named in DECISIONS_PENDING_JOSH.md and never built —

a present pulse, a climbing contour, interlocking surface motion under a slow tune, and a

counter-melody answering in the gaps, all without raising the volume.

• Axis: A=TIGHT · B=FUSION · E=acoustic.

Instrumental only, no vocals. A calm sanctuary theme for a big-budget fantasy action-adventure RPG,
set at a waterfall in the cloud forest of Flores island, where the surface is busy and the tune is
slow. 76 BPM, D major, pentatonic. In the spirit of Aquatic Ambiance and Stickerbush Symphony. Open
with the melody ALONE in the first three seconds on a solo bamboo flute, slow, in long notes. Then
bring in the surface: fast, quiet, interlocking figuration on tuned bronze and plucked harp, two
parts weaving and never on the same beat, all of it soft. Then hand the slow melody between the
flute, a solo cello and a soft mallet instrument, varying it each time, with long quiet stretches
where one of them holds it alone and the figuration thins out. A cello answers the flute in the gaps
between phrases. One single lift, no surge. Then settle back to the melody at the end, long and
resolved, handing off to the next track. Warm chord changes underneath. Quiet throughout, but never
still and never sleepy - the contour keeps climbing. Live acoustic players in a real room. No
silence at the start or the end.

• AIVA: From a Style → Style Designer, "minimalist interlocking tuned percussion with a slow

lyrical lead". Key D major. Tempo 76. Duration 3:00. Compositions 3.

---

BATTLE — the caci rounds

Built on §2.4's inverted reading: the corpus's battle lead is **sparse, long-phrased and doubled in

harmony. These ask for a riff, not for shredding. All three are contest** cues. The class bend is

short valleys (8–12 s), harder surges, fast handoffs, a short cool-down rather than a full calm.

TP-B1 — region-first contest

• INTENT: the ceremonial-duel register at our plan's tempo, with the care line as prompt

language — stakes and heat, never dread.

• Axis: A=TIGHT · B=REGION-FIRST · E=acoustic.

Instrumental only, no vocals. A driving 148 BPM battle theme for a big-budget fantasy
action-adventure RPG: a ceremonial duel in a ring on Flores island, with a crowd around it keeping
the count. In the spirit of Battle on the Big Bridge and Silver for Monsters - a riff you can hum,
not a wall of noise. Structure it exactly like this. Open with the RIFF ALONE in the first three
seconds - hard, short, few notes, long phrases, on a low reed, no drums yet. Then the beat hits: a
tuned bronze gong marking a hard four-bar cycle, two interlocking hand drums, a slit drum, and
clapping and stamping feet on the off-beats. Then layer and change - hand the same riff between low
reed, brass and plucked bronze, varying it each time, higher then lower, and doubled in thirds when
the full ensemble takes it. Keep real quiet valleys of eight to twelve seconds where one instrument
holds the riff alone and the drums drop to a single pulse, then surge back harder. At least two
full-weight surges. One contrasting second section. Real chord changes underneath with a strong
forward-leaning cadence. Then settle at the end into a short cool-down that returns to the riff on
the first instrument and resolves, handing off to the next track. Heroic, hot-blooded, exciting -
a contest with everything on it, not horror and not evil. Never heavy for long stretches. Live
acoustic players, big room. No silence at the start or the end.

• AIVA: From a Style → Style Designer, "percussive world-fusion battle, gongs and drums with

brass". Key D major. 4/4. Tempo 148. Duration 3:00. Compositions 3.

TP-B2 — SCORE-FIRST battle

• INTENT: the palette isolation arm in the class where raw energy is most likely to carry a

grade on its own.

• Axis: A=TIGHT · B=SCORE-FIRST · E=acoustic. Isolation partner: TP-B1.

Instrumental only, no vocals. A driving 148 BPM orchestral battle theme for a big-budget fantasy
action-adventure RPG - a ceremonial duel in a ring on Flores island, a crowd keeping the count. In
the spirit of Those Who Fight and Battle on the Big Bridge. Open with the RIFF ALONE in the first
three seconds on brass - hard, short, few notes, long phrases. Then the beat hits: fast string
ostinato, timpani and heavy low percussion. Then layer and change - hand the same riff between
brass, low strings and woodwind, varying it each time, doubled in thirds at the biggest statement,
with real quiet valleys of eight to twelve seconds where one instrument holds it and the percussion
drops to a pulse, then surges back harder. At least two full-weight surges. One contrasting second
section. Real chord changes with a strong forward-leaning cadence, D minor turning to D major at the
peak. Then settle at the end into a short cool-down that returns the riff to the first instrument
and resolves, handing off to the next track.
Heroic, urgent, exciting, high-stakes. Never heavy for long stretches. Live orchestra, big room. No
silence at the start or the end.

• AIVA: From a Style → "Epic / Battle" family. Key D minor. 4/4. Tempo 148. Duration

3:00. Compositions 3.

TP-B3 — HYBRID MODERN

• INTENT: Josh's genre span runs ancient through modern to techno. Does a modern low end and

processed percussion under the gong cycle read as more *shippable* to his ear than the acoustic

arms? The only prompt in the set that leaves the acoustic register, deliberately one and not four.

• Axis: A=TIGHT · B=FUSION · E=HYBRID MODERN.

Instrumental only, no vocals. A 148 BPM hybrid orchestral-electronic battle theme for a big-budget
fantasy action-adventure RPG - a ceremonial duel in a ring on Flores island. In the spirit of Rivers
in the Desert and Battle on the Big Bridge. Open with the RIFF ALONE in the first three seconds on
brass - short, hard, memorable. Then the beat hits: live tuned bronze gongs and interlocking hand
drums from the island sitting on top of a modern cinematic low end, sub bass and tight punchy
processed kit. Then layer and change - the riff comes back every eight bars handed to a different
voice, bigger each time, doubled in parallel thirds at the peaks, with real quiet valleys of eight
to twelve seconds where the low end cuts out and one instrument holds the riff, then it surges back
harder. At least two full-weight surges. One contrasting second section. Real chord changes with a
strong forward-leaning cadence, in D. Then settle at the end into a short cool-down that returns the
riff to the first instrument and resolves, handing off to the next track. Heroic, urgent, modern,
high-stakes - a contest with everything on
it. Never heavy for long stretches. Big room, real players over the electronics. No silence at the
start or the end.

• AIVA: From a Style → Style Designer, "hybrid orchestral-electronic action". Key D major.

Tempo 148. Duration 3:00. Compositions 3.

---

BOSS — the crater-lake encounter at full weight

The class bend is PHASED ESCALATION: the grammar runs twice at rising intensity, a breath valley

between the phases, and the settle is a resolved full statement rather than a cool-down. This is the

class the corpus is thinnest on (two carded rows) and the class the arc grammar most obviously

serves, so it gets two arms rather than four.

TP-BS1 — region-first, phased

• INTENT: does an engine execute a two-phase escalation from prose, and does the region palette

carry weight at boss scale?

• Axis: A=TIGHT · B=REGION-FIRST · E=acoustic. Class bend: phased.

Instrumental only, no vocals. A boss battle theme for a big-budget fantasy action-adventure RPG - a
vast creature at the crater lakes of Flores island. 132 BPM, D. In the spirit of Hopes and Dreams
and Sealed Vessel: a theme that returns bigger, not a wall of noise. Structure it in two phases.
PHASE ONE: open with the melody ALONE in the first three seconds, bare and slow, on a low bamboo
flute. Then the beat enters underneath it - a single deep gong on the cycle, then interlocking hand
drums building. Then hand the melody between low flute, brass and bowed strings, varying it each
time, with quiet valleys where one instrument holds it, building to a first full statement doubled
in thirds. THE BREATH: everything drops to almost nothing for about ten seconds, one instrument
holding the tune. PHASE TWO: the same melody restated heavier from the bottom of the ensemble, with
a NEW counter-line above it, the drums doubled, driving to a bigger peak than phase one. Then settle
at the end - the melody returned to the low flute it started on, resolved, handing off to the next
track. Real chord changes throughout with a strong cadence. Enormous, high-stakes, awe and danger -
never horror, never evil. Never heavy for long unbroken stretches. Live acoustic players, huge room.
No silence at the start or the end.

• AIVA: From a Style → Style Designer, "monumental acoustic world-fusion, deep gongs and massed

drums with brass". Key D minor. 4/4. Tempo 132. Duration 3:30. Compositions 3.

Build the two phases and the breath in the section editor — this is the arm where AIVA's

editor rather than its generator is under test.

TP-BS2 — FUSION, phased, with the modern low end

• INTENT: the isolation partner for TP-BS1 on palette and register at once — the only place in

the set where two axes move together, because the boss class has two arms and the modern question

matters most where the weight is greatest. **A grade here attributes to the PACKAGE, not to one

axis**, and that is stated so nobody over-reads it.

• Axis: A=TIGHT · B=FUSION · E=HYBRID MODERN. Class bend: phased.

Instrumental only, no vocals. A hybrid orchestral-electronic boss battle theme for a big-budget
fantasy action-adventure RPG - a vast creature at the crater lakes of Flores island. 132 BPM, D. In
the spirit of Hopes and Dreams and Dancing Mad: a theme that restates rather than just intensifies.
Structure it in two phases. PHASE ONE: open with the melody ALONE in the first three seconds on a
solo horn, bare and slow. Then the beat enters - deep tuned gongs from the island and interlocking
hand drums over a modern cinematic sub bass. Then hand the melody between horn, strings and low
reeds, varying it each time, with quiet valleys where one instrument holds it and the low end cuts
out, building to a first full statement doubled in thirds. THE BREATH: everything drops to almost
nothing for about ten seconds. PHASE TWO: the same melody restated heavier from the bottom of the
ensemble with a new counter-line above it, processed percussion doubled, driving to a bigger peak
than phase one. Then settle at the end - the melody returned to the solo horn, resolved, handing off
to the next track. Real chord changes with a strong cadence. Enormous, modern, high-stakes, awe and
danger - never horror, never evil. Never heavy for long unbroken stretches. Big room, real players
over the electronics. No silence at the start or the end.

• AIVA: From a Style → Style Designer, "hybrid orchestral-electronic monumental". Key **D

minor. 4/4. Tempo 132. Duration 3:30. Compositions 3**. Two phases in the section

editor.

---

7. THE RUN PROTOCOL

1. Announce the browser session before driving Josh's Chrome — the standing GUI protocol. Never

drive it unannounced.

2. Per prompt, per engine: three variations. Reject only broken generations at the operator

tier; let the wall and Josh grade.

3. Download to D:/audio/benchmark/<engine>/<PROMPT-ID><variant>.<ext> — e.g.

D:/audio/benchmark/elevenlabs/TP-E4a.mp3. WAV where offered.

4. Judge each file:

D:/audio/rig/venv/Scripts/python.exe harness/music_gen/benchmark_judge.py \

--audio D:/audio/benchmark/elevenlabs/TP-E4a.mp3 \

--engine elevenlabs --prompt-id TP-E4 --cue-class EXPLORATION --variant a

Cue classes are EXPLORATION, SERENE (SANCTUARY is accepted as an alias), BATTLE, BOSS.

5. After the slate is populated, run --cohort once for cross-track distinctness.

6. Seed the review surface from build/audio/benchmark/SLATE_R1.json. Every row lands

ungraded and carries its engine's shipping posture and its blind spots.

7. Grade the arc grammar BY EAR against §0.3, movement by movement, because two of the four

movements have no instrument. Record the ear read on the row beside the wall's numbers.

8. Show Josh pre AND post iteration — his own extension of the ruling. Round 1 is the PRE.

What a grade attributes to

• Same intent, different palette (E1/E3/E4 · S1/S2 · B1/B2): the PALETTE. This is the "I hate

the instruments" question, answered.

• Same intent, different constraint (E1/E2): prompt specificity.
• Same intent, acoustic vs hybrid (B1/B3): the modern register.
• BS1 vs BS2: the PACKAGE only — two axes move together there, by design and by declaration.
• Same intent, different ENGINE: the engine as a package only. §5.2's confound stands.
• Across classes: whether the §0.4 class bends land. A serene cue that reads heavy, or a battle

cue with no valleys, is a bend failure and not a palette failure.

---

8. WHAT THIS DOCUMENT DOES NOT CLAIM

• That any prompt here will work. Twelve prompts built out of research is a hypothesis set. The

design's value is that a down-grade on any row still tells us which axis moved.

• That the judge wall can call a theme good. It cannot, and it says so on every row: its one

separating axis separates our generator from real released music, not good music from filler. And

two of the four arc-grammar movements it cannot see at all.

• That any of this is shippable. No engine on the slate is cleared to ship in this game today.

ElevenLabs' Model-Specific Terms carve Studio Games — commercialised video games on more than

one platform — out below Enterprise

(docs/licence_records/elevenlabs/READ_RECORD_2026-08-09.md), and AIVA's own legal page and

pricing page contradict each other (CURRENCY_SWEEP_2026-08.md §5.3). A graded-up theme is a

candidate for the development rung and an input to a licence decision, never a clearance.

• That the corpus bands are taste. They are percentiles of thirty carded tracks Josh named. A

theme outside them is not wrong; a theme BELOW the tenth percentile on note density and step share

is the thing he described in words, and that is the only claim the floor makes.

---

9. SOURCES AND READ STATUS

VERIFIED — fetched or read at HEAD 2026-08-09:

• docs/spine/DECISIONS_PENDING_JOSH.md — the 2026-08-09 rulings and extension; the round-3 and

round-4 grades verbatim; the energy floor and its per-class correction.

• Josh's live ElevenLabs v2 grade of 2026-08-09, relayed by the director — the source of §0 in full.
• _source/02_Tier_2_Region_Pages/flores_island.md §10 — the gong-waning, the interlocking-cycle

theory, the caci drumming, the prohibition.

• build/audio/exemplars/CORPUS.json — the 114-row original roster and its seven theme buckets

(182 rows total including the 68-row expansion; §3 uses the 114).

• build/audio/exemplars/instruments/MELODY_BANDS.json and build/audio/MIX_POLICY.json — every

percentile and loudness target in §2.4, re-resolved per class at HEAD including the SERENE and

BOSS fallbacks.

• build/audio/pass7/plans/PASS7_FLORES_{ROAD,FALLS,ROUNDS}.json — tempos, metre, pitch set,

durations, purpose classes, care blocks.

• docs/review_candidates.json :: PASS7_FLORES_ROAD — the round-4 grade verbatim.
• https://elevenlabs.io/docs/overview/capabilities/music and /music/best-practices — the input

roster, prompting guidance, output formats and duration bounds.

• https://aiva.crisp.help/en/article/general-user-manual-44klp4/ — the four modes, the settings

roster, the influence minimums, the editor controls.

LANE-REPORTED or NOT FETCHED — named because a number from an unread source is not evidence:

• The ElevenLabs composition-plan JSON schema: the documented path 404s. Section plans are

written as timed prose inside the prompt and no field names are invented.

• The craft-research documents in docs/proposals/music/craft_research/ are read through

MUSIC_PRACTICE_GAP_AUDIT.md at the read-status that audit records. No number in §2 comes from

them; §2's numbers are all from our own measured artifacts.

• AIVA preset style NAMES: the manual documents the modes and settings but the live preset roster

is a UI listing. Each AIVA row names a style FAMILY; the operator picks the nearest actual preset

and records which one on the slate row.

• The exemplar picks in §3.2 are editorial, chosen from the correct bucket on the quality each

lends. They are not a measured similarity ranking, and no instrument was used to select them.

Consumers: harness/music_gen/benchmark_judge.py (scores these prompt ids and cue classes),

build/audio/benchmark/SLATE_R1.json (holds the rows), docs/DOC_MAP.md (index row).

THE LOCKED BASE TEMPLATES (Josh rulings 2026-08-09 -- the factory generates from BOTH)

Two templates, complementary by measurement; the factory alternates them per theme for

built-in variety, and AIVA repairs each one's known gap at the development stage.

TEMPLATE A -- "v4", THE GROOVE-ALIVE BASE

Measured profile: melody floor PASS, syncopation 0.267 (the aliveness Josh loves),

yield 0.44 (back half repeats -- AIVA'S REPAIR: second-half development at the named

seam -- new hook, instrument-set change, tempo change).

PROMPT VERBATIM: Video game exploration music for an epic fantasy action-adventure RPG

set among the highland villages of Flores, Indonesia. A singable main melody with real

emotional pull: wonder, longing, quiet joy. Short intro, reach the main melody within

ten seconds; the rhythm enters soon after and the full ensemble within fifteen more.

The groove evolves through the track: drum patterns change, tempo lifts and relaxes,

the melody passes between bamboo flute, warm strings, and bright bronze bells,

sometimes bold, sometimes tender, with quiet valleys and surging peaks. Weave in the

feeling of village community and celebration: communal hand-drum circles,

call-and-response phrases between instruments, festive interlocking bells. Short

ending, settle briefly onto the melody and close cleanly so the next track can follow.

Vary the structure and instrumentation freely and surprise me, as long as the main

theme stays whistleable and returns changed. No vocals.

TEMPLATE B -- "v6", THE STRUCTURE-COMPLETE BASE

Measured profile: melody floor PASS, yield 1.0 (real-released-music register -- the

second half genuinely develops), syncopation 0.0 (AIVA'S REPAIR: restore the groove

aliveness -- syncopated interlocking percussion back in, Template A's feel).

PROMPT VERBATIM: Template A's text with ONE sentence inserted before "Short ending":

The second half must not repeat the first: develop the main theme into something new

there, a key change or a fresh counter-melody or a bolder transformed return, before

the final settle.

SLOTS (per Josh's automation ruling): the region/place identity, the cue-class

register (exploration/battle/boss/serene/main-story/side-quest/trade/home), the

instrument-inspiration palette, and the emotion words swap per theme; the structural

spine of each template never changes. Exemplar-title influences attach per class from

the 114-roster buckets. Every winner records its SECTION MAP so AIVA edit points are

named seams.

THE CH-2 SLICE SOUNDTRACK SLATE (Josh directive 2026-08-09: "how many more of each theme?")

Founding four (graded up, all EXPLORATION-class): the Variation-2 edits. The chapter

slate to ~22-30 unique tracks, classes per canon (arrival, village, healer, caci ring,

festival) + game practice:

• EXPLORATION: +2 day variants + 1 NIGHT variant (have 4) -> 7
• BATTLE: 3-4 (standard combat, caci-rhythm flavored per the region)
• BOSS: 2 (chapter boss + miniboss/phase variant)
• SERENE / REST: 2-3 (sanctuary, campfire recharge, healer's home)
• VILLAGE / HUB: 2 (community daily-life theme, market bustle)
• FESTIVAL / CELEBRATION: 1-2 (the caci festival, communal dance)
• EMOTION / STORY: 2-3 (loss, warmth, the healer's child moments -- quiet underscore)
• TENSION / DANGER: 1-2 (pre-combat dread, night danger)
• TRAVEL / TRANSITION: 1 (the compressed travel seams)
• STINGERS (short cues, 5-15s): 4-6 (victory, discovery, quest complete, rest-wake)

Template assignment: exploration/village/festival from TEMPLATE A+B alternation; battle/

boss bend the arc per the class table (short valleys, phased escalation); serene extends

settles; stingers are single-gesture prompts. All classes carry the region identity,

exemplar-title influences per bucket, and the no-noise guard. Full volume runs after

the CD-recalibrated rubric; the founding four are the reference sound for every class.

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