REVIEW_SHEET.md

music/generated/slice_v1/REVIEW_SHEET.md

SLICE V1 -- the first generated soundtrack candidates

Honest tier: GENERATED CANDIDATES, not a score. Every track below was produced by Stable Audio 3 Medium from text alone, with no audio of any kind as a conditioning input. Each is a REGISTER AND ARRANGEMENT candidate for the theme it names. None of them is the leitmotif -- the 2026-08-06 ruling keeps canonical themes composer-in-loop whatever generates, and the melody A/B established that text conditioning cannot reliably realise a specified melodic shape. So grade these on *does this place sound like this, and does this cue hold for three and a half minutes*, and not on *is that the tune*.

No listening test stands behind any verdict on this page. The producing lane measured every candidate and could not hear one. The KEEP/REJECT column is the measured rejection filter of PATTERN_FINDINGS_V2.md section 14 and nothing else. Josh's ear is the first ear.

Emitted 2026-08-07T19:56:28+00:00 - 32 candidates, 18 inside the envelope, 14 rejected.

How to read the numbers

The rejection filter is the joint envelope of PATTERN_FINDINGS_V2.md section 8 -- the box over melodic span, mean section length and composed rest that contains 25 of 28 held-out corpus rows while admitting 44.8% of album filler. It is the only thing section 14 licenses as a filter, and it is necessary-not-sufficient: inside the box means *not unlike every track Josh named*, never *good*.

Two more columns rank the survivors and reject nothing:

• rules fired (of 6) -- the six held-out rules of section 7a, fitted on fifteen training hits and scored once on the untouched thirteen, where zero of six reversed sign. section_count >= 15 is the strongest at test AUC 0.783 and a 3.85x lift.
• size axes on the hit side (of 6) -- the six axes that clear the calibrated multiplicity bar, each read against its own sibling median. All six are measures of structural scale; structural novelty is the one that clears in the negative direction, so a corpus row sits BELOW its album on it.

The honest limit, stated on the derivation's own face: strip duration out of the composite and its held-out AUC falls from 0.744 to 0.497, which is exactly chance. These numbers say a candidate is built to the scale of the tracks that stayed loved. They say nothing about whether it is beautiful.

What the run found

18 of 32 candidates are inside the envelope.

Every single rejection is on the same axis: X_rest_grammar. Not one candidate fell out on melodic span or mean section length. Composed rest is the silence budget plus the drop rate, and the corpus band for it is narrow — 2.68 to 17.43 — while this run's median is 15.918. The reason is visible in the drop counts: the median candidate drops 48.5 times against a corpus median of 19.5 and an album-filler median of 6. The caption asked it to cut the texture away and it cut the texture away roughly two and a half times more often than the tracks Josh named ever do. That is one named, actionable defect in this generation config, and it is the only one the filter found.

Section count is the axis the generator does not reach. The strongest held-out rule in the whole derivation is section_count >= 15 (test AUC 0.783, 3.85x lift) and 9 of 32 candidates fire it; the run's median is 11.0 against a corpus median of 17. This is not a surprise and it was declared before the run: control_surface_sa3.json never probed section count, so it was never a demonstrated controllable dimension. It is now measured as one the model misses.

The controlled contrast: does the structure paragraph do anything?

Variant A asks in full — *at least fifteen clearly different sections, state the main tune plainly and early, cut the texture to near-silence many times*. Variant B holds the frame, the period, the cultural sounds and the care line byte-identical and asks only for a piece that circles its own material. Same seeds, same durations, same model.

The explicit ask survives the envelope twice as often — 12 against 6 — and it does not get there by raising section count. Variant B's sections are *higher*, and so are its drops, which is what puts it outside the box. Read plainly: asking for self-similar circling made the model chop more, and asking for it explicitly plus a plain tune statement made it chop less. This is one run with sixteen candidates per arm and no calibrated null of its own — a descriptive read, not a test. It is enough to pick variant A for the next pass and not enough to write a rule.

The second measurement: does it sound like released game music?

timbre_critic.py scores a card's SOUND -- centroid, rolloff, flatness, octave-band profile -- against 1182 measured cards of real released game music, and it is ast-proved blind to everything structural. It cannot hear whether the music is any good; it can hear whether the timbre is in the family.

Median distance 0.6586, range 0.1635-1.6916. Under the landed bands, 31 of 32 read INSIDE and all 32 are within MARGINAL (2.669).

The critic's own controls are RED and the number is quoted anyway, for a stated reason. Its reference corpus has grown since its bands were landed, so its leave-one-out control reports the population's p90 has moved to 1.474 and its characterised blind spot has shifted. Re-landing those constants is a calibration decision for the critic's own lane, not something to do inside a generation pass. Under the drifted bands 28 of 32 are still INSIDE and all 32 are still within the outer band, so the reading does not depend on which calibration is right. That invariance is the whole reason it is quoted.

THE THIRD READ, AND IT IS NOW THE LIVE ONE (appended 2026-08-07 after the recalibration). The calibration decision this paragraph deferred has been taken in the critic's own lane. Its bands were re-derived from the corpus that now defines them and its blind spot re-characterised: INSIDE 1.658 -> 1.481, MARGINAL 2.669 -> 2.805, self-test back to 9/9. Two of the 1,182 reference cards were also retired in the same sitting -- byte-identical UNMATCHED__ twins of EX_014 and EX_094, which acquire_exemplar.py no longer mints -- so the reference population is 1,180. Re-scored against that population under those bands, this slice reads median 0.6605, range 0.1624-1.6941, 28 INSIDE / 4 MARGINAL / 0 OUTSIDE. The prediction the paragraph above made is therefore confirmed exactly: 28 inside, all 32 within the outer band, and the median moved by 0.0019 (0.6586 -> 0.6605) -- the whole effect of a 6x-grown, twin-cleaned reference on this reading is three thousandths of a robust sigma. The four MARGINAL rows are named rather than left to be inferred: EXPLORATION_FLORES_b_41022 1.6941, EXPLORATION_FLORES_a_41033 1.6282, BATTLE_CACI_b_42033 1.5557, BATTLE_CACI_b_42044 1.5211 -- all four sit between the new INSIDE and the old one, which is to say they did not move; the threshold did. Nothing here upgrades what the number means: a PASS is still necessary and never sufficient, and it still cannot hear whether the music is any good.

For scale: the dossier that promoted this generator measured the authored-plus-sampler route at a median distance of 1.932 with five of five MARGINAL. These sit at 0.6586. That is a statement about timbre and nothing else -- the authored route is the one that actually carries a composed tune, and this one does not.

Per theme

The slowest theme is the one that lands. The departure cue, asked for at 66 bpm at chamber scale, is the only theme whose median section count sits above the corpus median, and it is the only one that kept nearly everything. The two Flores themes, at 92 and 148, sit at eight and nine sections. Whether that is tempo, or register, or the fact that a sparse cue has fewer places to drop out of, this run cannot separate — it is four themes, each confounded with its own bpm, and it is recorded as an observation to test rather than a cause.

EXPLORATION - Flores Highlands -- the region at rest

• Theme registry row: REGION_FLORES_ISLAND - chapter CH_03 - region page flores_island - care tier standard
• Cue rows this would serve: MC_0048, MC_0053, MC_0050
• Time period given to the generator: pre-colonial Manggarai highlands of western Flores, around 1000 to 1500 CE
• Cultural sounds given as inspiration: Take inspiration from a tuned gong-and-drum ensemble of the eastern Indonesian islands: a small set of hanging and held bronze gongs of different pitches, hand drums, and bamboo flute, playing short interlocking cycles that repeat and hand the beat between players. Highland cloud forest, terraced fields, rain in the air. No instrument was excluded.
• Register asked for: warm, unhurried, at reference register -- the place at rest, never darkened; reference-composed and never a sampled performance

Care line honoured. flores_island.md Section 10 -- 'No reproduction of protected sacred music - the gong-waning register is reference-composed, never sampled from consecrated performance.' No audio is a conditioning input here (enforced in sa3_generate.py), so the register reaches the model as description only. T0_Theme_Registry :: REGION_FLORES_ISLAND.deny_register_list -- the region's own register is never the threat signal; this cue is the region at rest and carries no menace coding.

Derived from:

• registries/T0_Theme_Registry [DRAFT v0.1]/T0_Theme_Registry.csv :: REGION_FLORES_ISLAND .instrumentation_substrate -- 'the gong-waning tuned gong-and-drum ensemble and its interlocking-cycle structure -- the region's Thread-21 vril-active instrument'
• _source/02_Tier_2_Region_Pages/flores_island.md Section 10 -- 'a highland-ceremonial register for Ch 3 (rising at the caci duel; the gong-waning ensemble at reference register)'; 'an ambient cloud-forest bed for Wae Rebo'
• registries/T0_Chapter_Index [ACTIVE v1.2] :: CH_03.era -- 'Composite - Pre-colonial Manggarai ~1000-1500 CE through Portuguese 1500s and Dutch colonial 1800s-1900s'
• build/audio/music_cue_table.csv :: MC_0048/MC_0050/MC_0053 -- CH_03 zone beds, theme_ref JOURNEY_WORLD, adaptive_class loop

Caption, structure variant A (the frame, period, cultural sounds and care line are identical across variants; only the structure paragraph moves):

We are creating an action-RPG soundtrack. This is the exploration theme for Flores Highlands -- the region at rest. Time period: pre-colonial Manggarai highlands of western Flores, around 1000 to 1500 CE. Take inspiration from a tuned gong-and-drum ensemble of the eastern Indonesian islands: a small set of hanging and held bronze gongs of different pitches, hand drums, and bamboo flute, playing short interlocking cycles that repeat and hand the beat between players. Highland cloud forest, terraced fields, rain in the air. The feel is warm, unhurried, at reference register -- the place at rest, never darkened; reference-composed and never a sampled performance. Build it as a long piece in many clearly different sections -- at least fifteen -- that keep returning to the same few ideas rather than moving on to new ones. State the main tune plainly and early, then bring it back again and again in different colours. Cut the whole texture away to near-silence many times, and let the gaps be heard. Instrumental only: no sung lyrics, no spoken word.

Caption, structure variant B (the frame, period, cultural sounds and care line are identical across variants; only the structure paragraph moves):

We are creating an action-RPG soundtrack. This is the exploration theme for Flores Highlands -- the region at rest. Time period: pre-colonial Manggarai highlands of western Flores, around 1000 to 1500 CE. Take inspiration from a tuned gong-and-drum ensemble of the eastern Indonesian islands: a small set of hanging and held bronze gongs of different pitches, hand drums, and bamboo flute, playing short interlocking cycles that repeat and hand the beat between players. Highland cloud forest, terraced fields, rain in the air. The feel is warm, unhurried, at reference register -- the place at rest, never darkened; reference-composed and never a sampled performance. Build it as a long piece that keeps circling its own material: a handful of ideas, stated plainly, repeated and re-coloured for the whole running time, with very little that is new after the opening minute. Instrumental only: no sung lyrics, no spoken word.

Hit medians for reference: sections 17, length 196.6 s, bars 114, drops 19.5, self-similarity 0.902, novelty 0.556 (corpus rows sit BELOW their album's 0.678). Joint box: melodic span 19.7-57.0 st, mean section 7.1-63.3 s, composed rest 2.68-17.43.

Rejected, and on what:

• SLICE_T1_EXPLORATION_FLORES_b_41044 - outside the box on X_rest_grammar
• SLICE_T1_EXPLORATION_FLORES_b_41022 - outside the box on X_rest_grammar
• SLICE_T1_EXPLORATION_FLORES_b_41011 - outside the box on X_rest_grammar
• SLICE_T1_EXPLORATION_FLORES_b_41033 - outside the box on X_rest_grammar

Held for Josh's grade:

• SLICE_T1_EXPLORATION_FLORES_a_41033 - 5/6 held-out rules (total_length_s, n_drops, bars, note_count, tempo_adjust_n), 5/6 size axes on the hit side
• SLICE_T1_EXPLORATION_FLORES_a_41022 - 5/6 held-out rules (total_length_s, n_drops, bars, note_count, tempo_adjust_n), 4/6 size axes on the hit side
• SLICE_T1_EXPLORATION_FLORES_a_41011 - 4/6 held-out rules (total_length_s, n_drops, note_count, tempo_adjust_n), 6/6 size axes on the hit side

BATTLE - The Caci Rounds -- the bought champion, and the broker behind him

• Theme registry row: REGION_FLORES_ISLAND - chapter CH_03 - region page flores_island - care tier standard
• Cue rows this would serve: MC_0058, MC_0059
• Time period given to the generator: pre-colonial Manggarai highlands of western Flores, around 1000 to 1500 CE
• Cultural sounds given as inspiration: Take inspiration from the ceremonial drumming that scores a village whip-duel in the eastern Indonesian islands: tuned bronze gongs of several pitches locking into a fast repeating cycle, hand drums driving the rounds, a ring of onlookers keeping time, bamboo flute over the top. No instrument was excluded.
• Register asked for: the contest's own music -- celebratory, proud and driving, the ensemble at full energy; a separate cold low string line arrives underneath as the outside pressure, and the ensemble itself is never darkened

Care line honoured. T0_Theme_Registry :: REGION_FLORES_ISLAND.deny_register_list -- 'no region's own register used as the threat signal, systematically darkened, inverted or menace-coded (CVD 17.1 collective protection expressed in music - the darkening operations belong to the antagonist material and travel WITH it, never with the place)'. THIS CAPTION OBEYS THAT LINE STRUCTURALLY: the gong-and-drum ensemble is the CONTEST's music and keeps its own bright, proud register throughout; the menace is carried by a separate cold low line that is not of the place, which is the broker's material. CH_03's own combat_framing routes the antagonist through the bought champion and the broker, 'never the community'.

Derived from:

• registries/T0_Boss_Encounter_Registry [DRAFT v0.1] :: CH_03 main -- 'Kaba Lomes, the bought Caci Champion (for broker Ngoeng Tana)'; phases '2 phases: ritual caci rounds; the broker's land pressure unmasked before the assembled community'; win_condition 'PACIFY;EXPOSE'; has_nonkill_path TRUE; care_tier standard
• registries/T0_Chapter_Index [ACTIVE v1.2] :: CH_03.combat_framing -- 'ritual caci whip-duel + broker exposure; routes through the bought champion, the deforestation crew, a corrupted dukun, and fauna, never the community'
• _source/02_Tier_2_Region_Pages/flores_island.md Section 10 -- 'the ceremonial drumming that scores the caci rounds'; 'rising at the caci duel'
• build/audio/music_cue_table.csv :: MC_0058/MC_0059 -- CH_03 combat_layer, boss_ref BE_0003.P1 / BE_0003.P2, layer_mode swap

Caption, structure variant A (the frame, period, cultural sounds and care line are identical across variants; only the structure paragraph moves):

We are creating an action-RPG soundtrack. This is the battle theme for The Caci Rounds -- the bought champion, and the broker behind him. Time period: pre-colonial Manggarai highlands of western Flores, around 1000 to 1500 CE. Take inspiration from the ceremonial drumming that scores a village whip-duel in the eastern Indonesian islands: tuned bronze gongs of several pitches locking into a fast repeating cycle, hand drums driving the rounds, a ring of onlookers keeping time, bamboo flute over the top. The feel is the contest's own music -- celebratory, proud and driving, the ensemble at full energy; a separate cold low string line arrives underneath as the outside pressure, and the ensemble itself is never darkened. Build it as a long piece in many clearly different sections -- at least fifteen -- that keep returning to the same few ideas rather than moving on to new ones. State the main tune plainly and early, then bring it back again and again in different colours. Cut the whole texture away to near-silence many times, and let the gaps be heard. Instrumental only: no sung lyrics, no spoken word.

Caption, structure variant B (the frame, period, cultural sounds and care line are identical across variants; only the structure paragraph moves):

We are creating an action-RPG soundtrack. This is the battle theme for The Caci Rounds -- the bought champion, and the broker behind him. Time period: pre-colonial Manggarai highlands of western Flores, around 1000 to 1500 CE. Take inspiration from the ceremonial drumming that scores a village whip-duel in the eastern Indonesian islands: tuned bronze gongs of several pitches locking into a fast repeating cycle, hand drums driving the rounds, a ring of onlookers keeping time, bamboo flute over the top. The feel is the contest's own music -- celebratory, proud and driving, the ensemble at full energy; a separate cold low string line arrives underneath as the outside pressure, and the ensemble itself is never darkened. Build it as a long piece that keeps circling its own material: a handful of ideas, stated plainly, repeated and re-coloured for the whole running time, with very little that is new after the opening minute. Instrumental only: no sung lyrics, no spoken word.

Hit medians for reference: sections 17, length 196.6 s, bars 114, drops 19.5, self-similarity 0.902, novelty 0.556 (corpus rows sit BELOW their album's 0.678). Joint box: melodic span 19.7-57.0 st, mean section 7.1-63.3 s, composed rest 2.68-17.43.

Rejected, and on what:

• SLICE_T2_BATTLE_CACI_b_42011 - outside the box on X_rest_grammar
• SLICE_T2_BATTLE_CACI_a_42022 - outside the box on X_rest_grammar
• SLICE_T2_BATTLE_CACI_a_42044 - outside the box on X_rest_grammar
• SLICE_T2_BATTLE_CACI_a_42033 - outside the box on X_rest_grammar
• SLICE_T2_BATTLE_CACI_b_42022 - outside the box on X_rest_grammar

Held for Josh's grade:

• SLICE_T2_BATTLE_CACI_b_42044 - 5/6 held-out rules (total_length_s, n_drops, bars, note_count, tempo_adjust_n), 6/6 size axes on the hit side
• SLICE_T2_BATTLE_CACI_a_42011 - 4/6 held-out rules (total_length_s, n_drops, note_count, tempo_adjust_n), 5/6 size axes on the hit side
• SLICE_T2_BATTLE_CACI_b_42033 - 3/6 held-out rules (total_length_s, n_drops, note_count), 6/6 size axes on the hit side

TITLE_CHARACTER - Protagonist Theme -- the main-title statement

• Theme registry row: PROTAGONIST_THEME - chapter CH_PROLOGUE - region page fairy_realm - care tier standard
• Time period given to the generator: outside human history -- a timeless otherworld at its height, with no forward weight and no clock; the same tune is re-voiced across the human world from about 1000 CE on
• Cultural sounds given as inspiration: Take inspiration from a full orchestra heard from inside an abundant, luminous place: one solo carrier out in front -- a single flute, a single cello, or one wordless humming voice -- with strings, harp and light bells behind it, phrased like old-world folk song rather than anything modern. No instrument was excluded.
• Register asked for: full, luminous and unhurried, heroic and intimate at once; never ominous, and nothing hints at anything to come

Care line honoured. docs/spine/CH_PROLOGUE.md L245 -- 'no reproduced living-tradition liturgy, and no prophetic or foreshadow cue anywhere'. The caption asks for no dread swell, no minor turn and no held warning note, and asks for a wordless hum rather than any sung text. T0_Theme_Registry :: PROTAGONIST_THEME.deny_register_list -- shares no audible material with GRAND_SAGE_REVEAL before the gate; a generated candidate carries no GRAND_SAGE material by construction, since no such material exists on disk.

Derived from:

• registries/T0_Theme_Registry [DRAFT v0.1] :: PROTAGONIST_THEME -- theme_type character; instrumentation_substrate 'orchestral with solo instrument leitmotif evolution by age stage'; carrier_voice 'one solo carrier, re-voiced per age stage'; mood_tags [heroic, contemplative, intimate, transcendent]; tier A
• registries/T0_Theme_Registry :: PROTAGONIST_THEME.head_cell_spec -- '4-note head, PURELY STEPWISE ..., singable inside a ninth; the ONE deviant is a borrowed degree on the third note. Never triad-outlining'
• docs/spine/CH_PROLOGUE.md L245 (music_mood) -- 'a full, luminous, unhurried register (the realm at its abundance, the birth) ... a note of timelessness (the fairy-time ambient, no forward weight) ... no prophetic or foreshadow cue anywhere'
• registries/T0_Chapter_Index [ACTIVE v1.2] :: CH_PROLOGUE.era -- 'Pre-game timeframe (dimensional)'; combat_framing 'the abundant unhurried ceiling register'

Caption, structure variant A (the frame, period, cultural sounds and care line are identical across variants; only the structure paragraph moves):

We are creating an action-RPG soundtrack. This is the title character theme for Protagonist Theme -- the main-title statement. Time period: outside human history -- a timeless otherworld at its height, with no forward weight and no clock; the same tune is re-voiced across the human world from about 1000 CE on. Take inspiration from a full orchestra heard from inside an abundant, luminous place: one solo carrier out in front -- a single flute, a single cello, or one wordless humming voice -- with strings, harp and light bells behind it, phrased like old-world folk song rather than anything modern. The feel is full, luminous and unhurried, heroic and intimate at once; never ominous, and nothing hints at anything to come. Build it as a long piece in many clearly different sections -- at least fifteen -- that keep returning to the same few ideas rather than moving on to new ones. State the main tune plainly and early, then bring it back again and again in different colours. Cut the whole texture away to near-silence many times, and let the gaps be heard. Instrumental only: no sung lyrics, no spoken word.

Caption, structure variant B (the frame, period, cultural sounds and care line are identical across variants; only the structure paragraph moves):

We are creating an action-RPG soundtrack. This is the title character theme for Protagonist Theme -- the main-title statement. Time period: outside human history -- a timeless otherworld at its height, with no forward weight and no clock; the same tune is re-voiced across the human world from about 1000 CE on. Take inspiration from a full orchestra heard from inside an abundant, luminous place: one solo carrier out in front -- a single flute, a single cello, or one wordless humming voice -- with strings, harp and light bells behind it, phrased like old-world folk song rather than anything modern. The feel is full, luminous and unhurried, heroic and intimate at once; never ominous, and nothing hints at anything to come. Build it as a long piece that keeps circling its own material: a handful of ideas, stated plainly, repeated and re-coloured for the whole running time, with very little that is new after the opening minute. Instrumental only: no sung lyrics, no spoken word.

Hit medians for reference: sections 17, length 196.6 s, bars 114, drops 19.5, self-similarity 0.902, novelty 0.556 (corpus rows sit BELOW their album's 0.678). Joint box: melodic span 19.7-57.0 st, mean section 7.1-63.3 s, composed rest 2.68-17.43.

Rejected, and on what:

• SLICE_T3_TITLE_PROTAGONIST_b_43044 - outside the box on X_rest_grammar
• SLICE_T3_TITLE_PROTAGONIST_a_43044 - outside the box on X_rest_grammar
• SLICE_T3_TITLE_PROTAGONIST_b_43011 - outside the box on X_rest_grammar
• SLICE_T3_TITLE_PROTAGONIST_b_43022 - outside the box on X_rest_grammar

Held for Josh's grade:

• SLICE_T3_TITLE_PROTAGONIST_a_43011 - 5/6 held-out rules (total_length_s, n_drops, bars, note_count, tempo_adjust_n), 5/6 size axes on the hit side
• SLICE_T3_TITLE_PROTAGONIST_a_43022 - 5/6 held-out rules (total_length_s, n_drops, bars, note_count, tempo_adjust_n), 4/6 size axes on the hit side
• SLICE_T3_TITLE_PROTAGONIST_a_43033 - 4/6 held-out rules (total_length_s, n_drops, bars, tempo_adjust_n), 4/6 size axes on the hit side

MELANCHOLIC - Home and Loss -- the departure at the slice's close

• Theme registry row: HOME_AND_LOSS - chapter CH_13 - region page south_africa - care tier elevated
• Cue rows this would serve: MC_0216
• Time period given to the generator: southern Africa around 1100 to 1450 CE, at a dry-stone sacred enclosure, with a far older desert tradition alive around it
• Cultural sounds given as inspiration: Take inspiration from a sparse, voice-first southern African sound world: one unhurried wordless voice, ankle rattles, struck and shaken idiophones, a single hand drum, long open air between sounds and almost no chords underneath. One solo cello or flute from a small chamber group carries the tune. No instrument was excluded.
• Register asked for: intimate, warm and mournful at chamber scale; it aches without ever swelling, and returns at the end to the same warm opening sound, unbroken and unregretted

Care line honoured. south_africa.md Section 10 -- 'no sacred ceremonial content is reproduced. The trance-dance and rain-rite music is referenced and never recorded, performed or synthesized, and no melody attributed to a living ceremonial tradition enters the score.' The caption asks for TEXTURE families (rattle, struck idiophone, wordless voice) and names no song, no ceremony and no tradition-owned melody. T0_Theme_Registry :: HOME_AND_LOSS.allow_register_list -- the exit returns to the abundance register 'INTACT AND UNIRONISED', which the caption states in plain words; and its deny list forbids any prophetic or foreshadow cue.

Derived from:

• registries/T0_Theme_Registry [DRAFT v0.1] :: HOME_AND_LOSS -- instrumentation_substrate 'orchestral at chamber scale with a single solo carrier voice ... reduced to near-silence at the mother's scene, and returning to the abundance register intact and unironised at the exit'; head_cell_spec '5-note head, purely stepwise, chamber-scaled; the ONE deviant is a metric oddity - the phrase begins on the second beat and never resolves onto a downbeat'; tier A
• registries/T0_Theme_Registry :: HOME_AND_LOSS.recurrence_anchors -- 'beats CHPRO_B07, CHPRO_B16, CH01_B13, CH01_B08, CH02_B11 and CH13_B13'; CH13_B13 is the departure this cue states
• build/audio/music_cue_table.csv :: MC_0216 -- 'ch13 home_and_loss statement', region_page south_africa, terrain sacred_site, care_tier elevated, theme_ref HOME_AND_LOSS
• _source/02_Tier_2_Region_Pages/south_africa.md Section 10 -- 'sparse, vocal-first, rattle and struck-idiophone percussion, no orchestral pad, and a deliberate absence of harmony under the melody'; 'no melody attributed to a living ceremonial tradition enters the score'
• registries/T0_Chapter_Index [ACTIVE v1.2] :: CH_13.era -- 'Composite - Pre-civilizational continuous /Xam tradition plus Great Zimbabwe dry-stone florescence ~1100-1450 CE and Mutapa successor'

Caption, structure variant A (the frame, period, cultural sounds and care line are identical across variants; only the structure paragraph moves):

We are creating an action-RPG soundtrack. This is the melancholic theme for Home and Loss -- the departure at the slice's close. Time period: southern Africa around 1100 to 1450 CE, at a dry-stone sacred enclosure, with a far older desert tradition alive around it. Take inspiration from a sparse, voice-first southern African sound world: one unhurried wordless voice, ankle rattles, struck and shaken idiophones, a single hand drum, long open air between sounds and almost no chords underneath. One solo cello or flute from a small chamber group carries the tune. The feel is intimate, warm and mournful at chamber scale; it aches without ever swelling, and returns at the end to the same warm opening sound, unbroken and unregretted. Build it as a long piece in many clearly different sections -- at least fifteen -- that keep returning to the same few ideas rather than moving on to new ones. State the main tune plainly and early, then bring it back again and again in different colours. Cut the whole texture away to near-silence many times, and let the gaps be heard. Instrumental only: no sung lyrics, no spoken word.

Caption, structure variant B (the frame, period, cultural sounds and care line are identical across variants; only the structure paragraph moves):

We are creating an action-RPG soundtrack. This is the melancholic theme for Home and Loss -- the departure at the slice's close. Time period: southern Africa around 1100 to 1450 CE, at a dry-stone sacred enclosure, with a far older desert tradition alive around it. Take inspiration from a sparse, voice-first southern African sound world: one unhurried wordless voice, ankle rattles, struck and shaken idiophones, a single hand drum, long open air between sounds and almost no chords underneath. One solo cello or flute from a small chamber group carries the tune. The feel is intimate, warm and mournful at chamber scale; it aches without ever swelling, and returns at the end to the same warm opening sound, unbroken and unregretted. Build it as a long piece that keeps circling its own material: a handful of ideas, stated plainly, repeated and re-coloured for the whole running time, with very little that is new after the opening minute. Instrumental only: no sung lyrics, no spoken word.

Hit medians for reference: sections 17, length 196.6 s, bars 114, drops 19.5, self-similarity 0.902, novelty 0.556 (corpus rows sit BELOW their album's 0.678). Joint box: melodic span 19.7-57.0 st, mean section 7.1-63.3 s, composed rest 2.68-17.43.

Rejected, and on what:

• SLICE_T4_MELANCHOLIC_HOME_AND_LOSS_b_44044 - outside the box on X_rest_grammar

Held for Josh's grade:

• SLICE_T4_MELANCHOLIC_HOME_AND_LOSS_a_44033 - 6/6 held-out rules (section_count, total_length_s, n_drops, bars, note_count, tempo_adjust_n), 6/6 size axes on the hit side
• SLICE_T4_MELANCHOLIC_HOME_AND_LOSS_a_44022 - 6/6 held-out rules (section_count, total_length_s, n_drops, bars, note_count, tempo_adjust_n), 6/6 size axes on the hit side
• SLICE_T4_MELANCHOLIC_HOME_AND_LOSS_b_44033 - 6/6 held-out rules (section_count, total_length_s, n_drops, bars, note_count, tempo_adjust_n), 6/6 size axes on the hit side

What to do with a grade

• A track Josh likes becomes the register reference for its theme: the authored leitmotif (author_head_cell.py) is what will carry the tune, and this candidate is what the arrangement around it should sound like.
• A track Josh rejects is kept on disk. A losing candidate calibrates the filter, and the whole reason the six pre-registered hypotheses of V1 died honestly is that nothing was thrown away.
• A theme where everything was rejected routes to the generator question, not to a re-prompt: control_surface_sa3.json already shows this model does not respond to structural instruction on four of eight probed dimensions.

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