PASS3_VERDICT.md

music/pass3/PASS3_VERDICT.md

Pass 3 — the verdict

Honest tier: STRUCTURE. Nobody has heard these tracks — not Josh, not the director, not the lane

that made them. Every claim below is either a property of the written score or a number measured off

the rendered audio by the same rulers the corpus is measured with. There is no listening verdict

here and none should be read into it. Nothing from this lane goes on the review queue; publishing is

the director's act.

0. What round 2 failed on, and what pass 3 did about it

Josh's two rulings, verbatim, at the tail of docs/spine/DECISIONS_PENDING_JOSH.md:

• "the melodies are way too simple. There needs to be a lot more complexity added and to put some

intelligence into the melodies and how everything orchestrates."

• "The songs dont build and hold with proper peaks and valleys that still never lose the listeners,

I like more upbeat."

And two additions that landed mid-build and changed the brief:

• "please don't use the same exemplar melodies across all tracks, even if in the same region we

still need novelty."

• "there's a difference between a serene waterfall scene having a track that's gonna put you to

sleep and be peaceful versus having a track that's peaceful and still uplifting. we are building

an ARPG, and this is a fantasy video game that needs to have energy across the board in every

track. Scaled by theme as well."

The measured diagnosis was not "add notes". Round 2's entire melodic content across three cues,

32 layers and 580 seconds was ONE two-bar six-note cell restated twenty-one times, under seven

transform names of which six leave the interval spine bit-identical. In Caplin's taxonomy a two-bar

unit is a BASIC IDEA — the smallest constituent the grammar has — so round 2 contained no theme at

all. Pass 3 keeps the cells exactly as simple as they are and builds the level above them.

1. What was planned

THREE CUES, THREE THEMES. Each has its own basic idea, its own contour, its own rhythm and its own

cadence, and no two are transformations of each other. Region identity is carried by the mode, the

gong-waning instrumentation and its declared concert analogues, the colotomic cycle, the interlocking

elaboration and the care line — everything except the tune, which is where it always belonged.

Round 2's head cell measures a rest fraction of exactly 0.000 — it sounds continuously for all eight

of its beats — which is why both of its counterlines scored onset coincidence 1.00 and

exclusive-sounding fraction 0.000. Complementarity was impossible to write against it. Every pass-3

idea breathes, and the counter-melodies are written INTO those rests.

THE PARENT QUOTATION IS SPENT ONCE. T0_Theme_Registry :: REGION_FLORES_ISLAND.transformation_plan

requires the region cell to quote the JOURNEY_WORLD head at least once. Round 2 satisfied that by

making the parent head the shared skeleton of all three cues, which is the reuse Josh has now named

as a defect. Here it is a declared leitmotif reference: the parent's four notes, verbatim with the

licensed +9 leap, on trumpet, at cycle 23 of the road cue's arrival span — one moment, in one cue,

recorded in the plan with what earns it so a checker can tell design from convenience.

THE SIX-SPAN DEVELOPMENT SCHEDULE, implemented literally from the motivic lane's §6, with span

lengths deliberately unequal at 4, 4.5, 4, 4, 3, 4 cycles and a division arch of 0-1-2-2-0-1 that is

simplest at the two points of greatest weight and densest in the middle:

1. the theme, whole and plain, as a compound sentence — division 0, chain depth 1;

2. the answer as an expanded period whose cadence is EVADED at its expected bar and retried —

Schmalfeldt's one-more-time, and the reason this span is half a cycle longer than its neighbours;

3. the departure: the theme mutated by DEGREE MAP into the plagal face, on a new carrier, with the

contrasting idea entering as a real second voice — division 2, chain depth 3;

4. the displacement and the combination: the theme intervallically expanded by 1.5 and displaced

half a beat against an accompaniment that holds the grid — chain depth 4;

5. the subtraction: the presentation ONLY, in augmentation, deliberately incomplete, with the chain

RESET to the plant and the surface simplified back to division 0;

6. the arrival, near-plain and widest, then liquidation into the loop join with the cadence withheld.

2. What the score says — the primary acceptance evidence

build/audio/pass3/SCORE_SCORECARD.{json,md}, computed from the plans through the renderer's own

compiler before a sample was played, per the score-first ruling

(MUSIC_CRAFT_DOCTORATE.md §7.6). ROUND 2 WAS SCORED BY THE SAME CODE, which is what makes the

before-and-after a measurement rather than a memory.

TARGETS PASSED: 55 of 57 for pass 3; 7 of 57 for round 2, scored by the same code.

THE TWO MISSES ARE BOTH ON ONE AXIS OF ONE CUE and the mechanism is measured rather than guessed.

The sanctuary cue's obbligato scores onset coincidence 0.357 against a 0.35 ceiling and an

exclusive-sounding fraction of 0.055 against a 0.10 floor. Its carrier is the densest per sounding

bar of the three (4.42 notes) on the shortest phrases (standard scale packs presentation,

continuation and cadence into eight bars instead of sixteen), so the gaps a counter-melody can

answer in are the narrowest available anywhere in the set. The scorecard now reports the ceiling

beside the score: the carrier's own rest fraction over the shared window is 0.599, so there is room

in principle and the writing is not using it. THE NAMED REPAIR for the next rung is to hand the

division layer to the ornament part and leave the carrier nearer its skeleton, which widens the

gaps without thinning the cue; that is a composition change of real size and is not attempted here.

3. What the render measures — the fidelity check, run second

The audio battery is the RENDER-FIDELITY check under the score-first ruling and is not evidence of

quality. Five of its instruments have returned honest nulls against Josh's grades — the 3-to-5-cycle

law reads within-album AUC 0.501, sensory roughness 0.4963, counter-melody presence from audio

0.195, which is on the WRONG side of chance — so a pass from it bounds where the defect is NOT and

says nothing about whether the music is good.

All three cues rendered end to end with no manual step, all parts present (29 of 29, 17 of 17,

31 of 31), and every measurement below is emitted rather than transcribed

(build/audio/pass3/RENDER_SUMMARY.json, measurements/*_PLANNED_VS_MEASURED.json,

DROP_PROVENANCE.json).

TWO ROUND-2 DEFECTS ARE FIXED AND MEASURED. Both published pass-2 cues carried silence fraction

exactly 0.0 — a property 0 of 62 reference tracks share — and a deepest drop of about -6.3 dB against

a corpus floor of -10.73. All three pass-3 cues now go genuinely quiet, and two of the three clear

the drop-depth floor with room. The composed silence windows were verified against the COMPILED

NOTES before the render and land where they were written.

THE BUILD-AND-HOLD AXIS IMPROVED ON THE INSTRUMENT THAT OWNS IT. The battery's dynamic_arc

instrument FAILED all three round-2 cues; it passes ROAD and FALLS. That is the axis Josh's addendum

named in words, measured by the instrument built for it.

STRUCTURAL YIELD WENT THE WRONG WAY AND THE DENOMINATOR IS PART OF WHY. DROP_PROVENANCE.json

walks every measured drop against the plan's own removals:

Thirty-eight of ROAD's forty-one drops remove fewer than two spectral lanes, which is the sampler

artefact the pass-2 prototype identified and timestamped (its section 4.4: 28 of 31, near-regular, on

the cycle period). So the true figure is bracketed rather than known — [0.34, 1.56] for ROAD and

[0.48, 4.00] for FALLS — and neither end is comparable to a corpus number measured on mastered

commercial audio whose denominator carries no artefacts at all.

ONE HYPOTHESIS WAS TESTED AND REFUTED, and it is recorded so the next lane does not spend a render on

it. Moving the kotekan from pass 2's generator branch into a written-out score dropped the

dur_scale of 5.0 that let struck metal ring — the exact repair pass 2 had already paid for. It was

restored and everything was re-rendered. It changed almost nothing: yield 0.486 to 0.485, 0.357 to

0.342, 0.071 to 0.067. The ringing-length regression was real and is fixed; it was not the cause.

4. My own sighted-structure read, including where plan and render disagree

THE ROUNDS CUE IS THE ONE REAL FAILURE AND ITS CAUSE IS STRUCTURAL, NOT A BUG. Twenty-two planned

sections measured as THREE. Round 2's caci measured eight from a near-identical density profile

(planned mean 15.7 active parts against pass 3's 15.6), so the battle cue got WORSE on the boundary

counter while getting better on every score-side axis. Reading the two plans against each other, the

difference is what replaced what: round 2's caci carried six single-cycle emptyings scattered through

the cue; pass 3's carries ONE seven-cycle collapse, because that is what WQ_0078's named repair and

the build-and-hold doctrine both asked for. Six short holes produce twelve boundary events; one long

structural collapse produces two.

SO TWO OF OUR OWN INSTRUMENTS ARE PULLING IN OPPOSITE DIRECTIONS, AND THAT IS THE FINDING OF THE PASS.

Structural yield and boundaries-per-minute reward frequent texture change. The dynamic-arc instrument

and Josh's own words reward long plateaus. A chroma-and-MFCC self-similarity segmenter reads a

sustained hold as ONE section by construction — that is what a hold IS to it. The corpus does both at

once (hold share 0.407, boundaries 5.09 per minute), which means real music holds its ENERGY while

continuously changing its TREATMENT — which is exactly the doctorate's own section 2 finding: the

loved tracks repeat their material and change their treatment, continuously and in small increments,

under a sparse ridge of about three real events. OUR HOLDS HOLD BOTH. A pass-3 hold keeps the energy

flat AND keeps roughly the same layer set for four to six cycles, so the segmenter sees a block.

THE NAMED REPAIR, and it is a composition change of real size: inside every hold, rotate the carrier,

the instrument family and the irama level every cycle or two while the declared energy stays flat. It

is not attempted here, because one hypothesis has already been refuted this pass and diagnosing

before regenerating is the discipline WQ_0078 itself sets.

WHERE PLAN AND RENDER DISAGREE, itemised:

• SECTION COUNT. ROAD 21 planned against 15 measured; FALLS 18 against 17; ROUNDS 22 against 3. The

plan's sections are where the COMPOSER intends a boundary and the card's are where the SIGNAL

changes; only some of the first become the second, and the property that decides is density swing.

FALLS is nearly exact because it is the sparsest cue, with a planned density floor of 0 and a mean

of 10.0 active parts.

• est_layers_mean. Planned 16.3 / 10.0 / 15.6 against measured 4.0 / 5.2 / 5.4. Different units —

planned simultaneous PARTS against measured spectral BANDS — and the battery's own layer census

says it saturates near 8 streams where a real cue runs 24 to 40 parts, so it cannot adjudicate the

layer clause at all. The stem manifest is the only honest answer to how many layers there are, and

it says 29, 17 and 31, all rendered.

• ENTRANCE SPREAD. FALLS planned 0.964, measured 0.450. The measured proxy counts parts with an entry

rather than activation transitions, so a part that rests and returns is counted once against the

plan's two. The planned figure answers the question; the measured one answers a different one.

• THE JOINT BOX. ROAD and FALLS sit inside; ROUNDS falls out on section_mean_s at 64.9 s, which is

the same three-boundary failure expressed a second way rather than an independent finding.

• MELODIC COMPLEXITY, and this one is uncomfortable. The battery's audio-side melodic instrument reads

ROAD at 0.208 against round 2's REST at 0.266 — LOWER on the axis this whole pass was about. Its

interval entropy is higher (3.43 against 2.29) and its contour depth deeper (8 against 5), and the

instrument returns NO VERDICT because zero of its five axes reach AUC 0.75 and its counter-melody

axis is on the wrong side of chance. It is a predominant-pitch extraction from a dense mix, and the

score-first ruling exists precisely because that measurement is worse than reading the notes we

already hold. It is reported because hiding it would be the dishonest move.

THREE THINGS I FOUND BY READING THE CONSTRUCTION THAT NO PREDICATE CAUGHT:

• THE ORNAMENT LAYER WAS A UNISON DOUBLING. My first implementation gave a second instrument the

carrier's own division notes — the everything-doubled wall, a named failure mode where parallel

reinforcement reduces the effective voice count and the texture sounds thick and empty at once. It

is now a heterophonic variant on a small timing offset: same identity, different detail. No gate

would have found it; the layer count, the crowding index and the slot separation were all fine.

• THE THREE CUES SHARE ONE FORM. Different themes, different tempi, different instrumentation,

different section maps, different span-5 lengths and scales — and an IDENTICAL operation sequence

(plant, sequence, modal mutation, displacement, augmentation, profound reharmonisation,

liquidation), an identical 35 / 28 operation split and an identical chain shape of 2.286 / 4 / 1.

That is the level-A schedule being an invariant engine, which is what section 6 prescribes; it is

also the exact shape of the complaint Josh has just made about melodies, one level up. Expect

"all three have the same shape" as the next note, and the fix is a per-cue span schedule rather

than one schedule instantiated three times.

• THE OPERATION BUDGET AND THE LEVEL-B ENGINE CONTRADICT EACH OTHER. Section 6.3 caps sequence,

fragmentation and liquidation at twice per cue; section 6.2's engine sequences inside every

continuation and liquidates into every cadence, and is declared an invariant that every span

instantiates. Six spans therefore spend ten sequences, five fragmentations and five liquidations

before a composer chooses anything. Resolved in favour of the engine — the cap guards against

noodling on a cell and the engine's uses are structural — and the overage is REPORTED in the

scorecard rather than hidden, so the next revision of section 6.3 can be made against numbers.

ONE SCORE THAT LOOKS TOO GOOD, NAMED AS SUCH. The derived counter-melodies read onset coincidence

0.000 on ROAD and ROUNDS. That is partly the measure checking its own generator: the five-step rule

places every onset inside a carrier rest BY CONSTRUCTION, so a zero there is close to a tautology.

The axis that is NOT guaranteed by construction is the exclusive-sounding fraction — 0.303 and 0.350

against round 2's 0.000 — and that is the number to read.

5. Does pass 3 beat round 2 on the score?

YES, DECISIVELY, ON THE SCORE-SIDE AXES — and the comparison is a measurement rather than a memory,

because round 2's own plans were scored by the same code in the same run.

• TARGETS PASSED: pass 3 55 of 57; round 2 7 of 57.
• Round 2 had NO phrase object at all: longest theme 0 bars, 0 declared operations, 0 cadences, chain

depth 0, and its three cues scored pairwise theme similarity 1.000 — three arrangements of one

tune. Pass 3 carries complete themes of 18, 10 and 18 bars, 35 named operations of which 28 are

spine-changing, seven cadences including a real evaded-and-retried one, mean chain depth 2.286

reaching 4 with one reset, and pairwise similarity 0.50 across all three pairs.

• The theme-carrying layers went from sounding in 28-30 percent of bars to 83-90 percent, and from

2.1-4.6 percent of composed notes to 8.7-11.3 percent.

• The counter-melodies went from onset coincidence 1.000 and exclusive-sounding 0.000 — lines written

to FUSE with the tune — to 0.000 / 0.303 and 0.000 / 0.350 on two of the three cues.

• The plans went from an intended hold share of 0.000 / 0.000 / 0.061 to 0.444 / 0.586 / 0.467, with

longest planned holds of 41.4, 40.0 and 32.4 seconds where round 2's measured 0, 13.4 and 23.8.

• Composed silence went from exactly zero to 0.0117-0.0250 measured, and the deepest drop from about

-6.3 dB to -12.6 and -16.8 on two of three.

NO, NOT ON THE RENDER'S STRUCTURAL-YIELD AXIS, and one cue fails it badly. ROAD and FALLS fall from

round 2's published 1.100 and 1.571 to 0.342 and 0.485, with the denominator dominated by sampler

artefacts and the composed-drops-only bracket at 1.56 and 4.00. ROUNDS cannot be scored on it at all —

three measured sections against twenty-two planned — and it is the one genuine failure of the pass.

Its diagnosis is above, its repair is named, and it is not fixed here.

WHAT THIS DOES NOT SAY. That pass 3 sounds better. Nobody has heard it. The score-side scorecard says

the WRITING changed in exactly the ways the diagnosis called for, and the diagnosis was Josh's own

words turned into numbers. Whether the writing turned into music is the one question none of this

answers, and Josh is the only instrument that can.

6. Honest limits

• NOTHING HERE HAS BEEN HEARD. Every number is a property of the writing or of a measured render.

The score says what was WRITTEN; only an ear knows whether it works, and Josh is the first ear.

• THE TARGETS ARE PROPOSED, NOT CALIBRATED. Every floor in §2 comes from the motivic lane's §3.6,

where each is labelled REPORTED or PROPOSED, and none has a multiplicity null. They cannot be run

on the corpus at all, because we hold no scores for the corpus. Clearing them is evidence that the

WRITING changed. It is not evidence that the music improved.

• THE PHRASE GRAMMAR IS CLASSICAL-PERIOD EUROPEAN PRACTICE. It is the frame around the region's own

form and never a replacement for it: the colotomic cycle governs the clock in all three cues and

the sentence-and-period apparatus fills the melodic level inside it. Every phrase is laid ON the

cycle grid; the cycle is never cut to fit a phrase.

• THE KOTEKAN FIGURE-SELECTION RULE IS A DECLARED IMPORT. The ornament lane's own care line warns

against filling a documented Manggarai gap with Balinese kotekan because Balinese kotekan is what

a composer knows. What is taken is the general interlocking-cycle principle the region page itself

attests for the gong-waning, realised by a figure-selection rule borrowed from a neighbour and

named as borrowed. The page's specific gap is left open.

• THE SANCTUARY ENERGY FLOOR IS BORROWED. SANCTUARY has no corpus rows, so its class floor is

MELANCHOLIC's — the nearest measured class, n=4. That is the weakest number in the whole scorecard

and it is declared at its site.

• THE PLANNED ENERGY CURVE AND THE MEASURED FLOW CURVE ARE DIFFERENT QUANTITIES. The class floor's

minimum is a flow tenth-percentile measured off a render; the plan's curve is a declared

intention. The comparison is a sanity band, not a test. HOLD SHARE is unit-free on both sides and

is the axis that compares directly.

• NO PITCHED GONG EXISTS IN EITHER CC0 SET, the tuning is 12-TET and a gong waning is not, and the

bamboo flute is a baroque recorder. Those limits are unchanged from pass 2 and are carried in

every plan's care.honest_limits.

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