T2_IMPROVEMENT_BRIEF.md

music/benchmark/records/T2_IMPROVEMENT_BRIEF.md

T2 IMPROVEMENT BRIEF — where the melody edit is still missing, and the words most likely to move it

The row. CH02_SOUNDTRACK_T2 "Surging groove, variation two" — the v6-ambient pilot's

syncopated call-and-response rework, EXPLORATION, on the Chapter-2 soundtrack since 2026-08-09.

Wall record: build/audio/benchmark/records/BENCH_R1_ELEVENLABS_EDIT_V2_2.json. Wall card:

build/audio/benchmark/cards/BENCH_R1_ELEVENLABS_EDIT_V2_2.md.

Why this exists. Josh, 2026-08-09, on the melody-variety edit run against this track: *still not

right — grade and where should we improve?* This answers the second half. It cannot answer the

first, and says so in the next paragraph rather than implying otherwise.

THE GRADE IS NOT ISSUED HERE, BECAUSE THE THING TO GRADE IS NOT ON DISK. The melody-variety

edit variation lives in the ElevenLabs project HDKvkPwmmvlMvAwHEiE9 and has not been downloaded.

D:/audio/benchmark/elevenlabs/ holds nothing newer than the two caci-duel takes banked at 18:41,

and D:/audio/ch02_soundtrack/ nothing newer than CH02_T7_CaciDuel_GongEdit_V2.wav at 19:17.

THE DOWNLOAD IS OWED, and until the WAV is banked no number in this program describes the new

edit. Everything below is measured off the take that is on the registry NOW — which is the correct

baseline for aiming the next edit either way, because it is the thing the new edit was cut from.

When the WAV lands, the grade is one command:

D:/audio/rig/venv/Scripts/python.exe harness/music_gen/benchmark_judge.py \
  --audio D:/audio/ch02_soundtrack/<banked file>.wav \
  --engine elevenlabs --prompt-id EDIT-V2-2-MELODY --cue-class EXPLORATION \
  --title "CH02 T2 Surging Groove, melody-variety edit"

---

1. The one finding that sits under all five failures

The melody bands fail this track on five axes. They are not five problems. They are one problem

read five ways, and the number that names it is not one of the five:

repeat_share 0.9022 — 323 of the extracted line's 359 intervals are ZERO.

(BENCH_R1_ELEVENLABS_EDIT_V2_2.json L7085; the histogram at L7095-L7101 reads

{0: 323, 1: 4, 2: 4, 3: 3, 5: 19, 7: 5}.)

Thirty-five intervals in the whole three-and-a-half minutes actually move. Nineteen of them are

perfect fourths, five are fifths, and eleven are steps. **The lead is a repeated-note ostinato — a

pulse on a pitch — not a tune.** A line that repeats one note cannot turn, cannot arch, cannot use

a vocabulary of interval sizes, and gives a harmonising voice almost nothing to be consonant

against. That is exactly the five failures, in order.

This also explains why the previous edit did not land: an instruction that asks for *more melodic

variety* over a line that is 90% repeated notes is asking a still object to be more interesting

while leaving it still. The instruction has to say move, and say it in note-level terms.

Honest tier. Every number here is AUDIO-ESTIMATE: the line is extracted from dense polyphony by

predominant-melody tracking, with 46 octave errors repaired (12.81% of notes, L5822), and

harmonization_density is a LOWER BOUND by construction. A 90% repeat share on a 359-note

extraction is not a subtle reading, but it is an extraction, and the whole brief inherits that.

---

2. The five failed axes — what the corpus demands, what this measures, what to say

Bands are the EXPLORATION class's own, n=8, from

build/audio/exemplars/instruments/MELODY_BANDS.json. Axis definitions:

harness/music_gen/instr_melody_bands.py L781-L830.

2.1 interval_class_count — the vocabulary is six sizes and four of them are decorations

The instrument counts distinct absolute interval SIZES, capped at 12+, and it is deliberately a

count and never an entropy. Six sounds close to the p10 of 6.4 until the histogram is opened: the

six classes are 0, 1, 2, 3, 5 and 7, and outside the 323 repeats the entire line is 4 semitone-1s,

4 semitone-2s, 3 minor thirds, 19 fourths and 5 fifths. The corpus median is 8.5 sizes used

substantively.

Edit-instruction language. Name the sizes, do not ask for "variety":

Rewrite the lead melody so it uses a wider set of intervals — seconds, thirds, fourths, fifths

and at least one sixth — instead of repeating the same pitch. No pitch may repeat more than twice

in a row.

2.2 contour_turning_rate — the line changes direction once every fourteen notes

Direction changes per note of the line. The corpus turns roughly every 2.4 notes; this turns every

13.8. No exploration exemplar in the corpus reads below 0.3138, so this is not a near miss at

the edge of a distribution — it is outside the whole set.

Edit-instruction language.

Give the melody a rising-and-falling shape: it should change direction several times in every

phrase — up two or three notes, then down — rather than sitting on one note or moving in one

direction.

2.3 contour_depth — the shape has no hierarchy

Median Douglas-Peucker depth: not how OFTEN the line turns but whether its turns nest — a big arch

carrying smaller arches inside it. The corpus's floor and its minimum are the same number, 5.0:

every one of the eight exploration exemplars reads at least 5. This one reads 3.

Depth is the axis that separates a wiggle from a melody, and it is the one an instruction has to

ask for structurally rather than locally. Turning rate can be bought with ornament; depth cannot.

Edit-instruction language.

Build the melody in two layers of shape: an overall arch across each eight-bar phrase — rising to

a high point and settling back — with smaller rises and falls inside each two-bar group.

2.4 phrase_peak_position — every phrase's highest note is its first note

The median normalised position of the phrase's highest note. Zero means the peak IS the opening: the

line enters at its top and goes nowhere but sideways and down, fifty-one times. The corpus puts the

peak about a fifth of the way in (0.1965 median), which is what makes a phrase feel like it is going

somewhere before it arrives.

Note the class contrast, because it matters for the eventual battle work and would be the wrong

target here: the BATTLE band's own p50 is 0.0 — a battle riff DOES state its top note first. The

exploration band does not, and this is an exploration cue.

Edit-instruction language.

Do not start each phrase on its highest note. Let each phrase begin lower, climb to its peak

about a third of the way through, and fall away to a resting note at the end.

2.5 parallel_harmony_share — what harmony there is, is doubling and bare fifths

The share of LEAD notes carrying another melodic voice at a consonant interval (classes 3, 4, 5, 7,

8, 9 mod 12; octaves and unisons excluded as doubling) over at least half the note's length —

instr_melody_bands.parallel_harmony L470-L527. The detail block (L7036-L7048) is where the shape

of the failure is:

• doubled_share 0.3928 — 39% of lead notes carry another voice at the OCTAVE or in unison.

That is doubling, and the instrument refuses to count it as harmony on purpose.

• harmonising_interval_classes {3: 9, 4: 5, 5: 15, 7: 117, 9: 7} — of the 153 notes that ARE

harmonised, 117 are fifths. Thirds and sixths together account for 21.

So the cue is not unharmonised; it is harmonised the medieval way — octaves, unisons and open

fifths. harmonization_density 0.4994 sits INSIDE the band (p10 0.4049), which is the proof that

the defect is the KIND of second voice and not the amount of it.

Edit-instruction language.

Harmonise the main melody with a second instrument a third or a sixth away for most of its notes.

Do not double it at the octave or in unison, and do not move in parallel fifths.

---

3. The sixth number, which is the same defect from the other side

step_share 0.9246 against a p90 of 0.8515 (band L296-L304) — ABOVE the band, with

leap_share 0.014 against a corpus median of 0.0911 (L305-L313). Above-band is not a floor

failure and the instrument does not gate it, but read beside a 0.9022 repeat share it is the same

sentence: of the few intervals that move, almost all of them inch. The corpus's exploration

melodies place one or two real leaps of a fifth or more per line.

Edit-instruction language. Fold into the vocabulary ask above:

Include one or two wide upward leaps of a fifth or more per phrase as the melody's signature

gesture, and step back down from each one.

---

4. What NOT to ask for — the four instructions that would waste the edit

Each of these axes is already INSIDE the corpus band. Asking for them buys nothing and risks

pushing an in-band axis out:

• Do not ask for more notes or a faster melody. note_density_per_s 1.9944 (band 1.1978-2.7039)

and note_density_per_sounding_s 3.4999 (band 2.2247-8.1449) are both in band. The line is not

too sparse. It is too STILL.

• Do not ask for "more harmony". harmonization_density 0.4994 is in band. See 2.5 — the

quantity is fine, the interval is wrong.

• Do not ask for a wider range. range_semitones 12.67 is in band (7.282-31.1). A 12-semitone

compass with a 90% repeat share is a full octave that is barely used.

• Do not ask for more rhythmic variety. distinct_durations 11 sits at the p90 of 11.3. The

rhythm is already the most varied thing about the line.

• And do not say "more melodic variety" again. That phrasing produced the take being graded. It

is an adjective where the instrument reads note-level facts.

---

5. Two adjacent failures the same edit should be made to carry

Not melody-band axes, but the same edit is the only cheap chance to buy them, and both are

Josh's own named bars rather than harness inventions:

• hook_presence — "there is a statement and no return." The opening motif is never restated

above the track's own harmonic background; n_returns 0 over 230 s. Against the melody-first,

hummable-leitmotif direction this is the most consequential single line in the record. The edit

should state the reworked tune, leave, and BRING IT BACK — on a different instrument, which is

also the carrier-migration clause no instrument can measure.

• dynamic_arc — plateau, hold share 0.728, longest hold 147.1 s, ZERO valleys (corpus hold

share 0.407). This is Josh's 2026-08-09 grade of the whole round in measured form — *too heavy

throughout, no valleys*. The four-movement arc grammar wants a real quiet valley before the last

build, and cycle_law names exactly where to cut it: the track holds unchanged from **104 s to

204 s**, 7.48 cycles, the worst span in the record.

---

6. What the wall should read if the edit works — the falsifiable prediction

Written down before the re-judge so the next pass cannot be graded by vibes:

If the melody axes move and Josh's ear still says no, the finding is that the melody bands are the

wrong proxy for this cue and that is worth knowing.

If they do not move, this track does not get another edit. The factory doctrine ruled the same

evening off the 3:30 Caci failure — docs/spine/DECISIONS_PENDING_JOSH.md :: *FACTORY DOCTRINE

LEARNED 2026-08-09* — says it in one line: edits repaint rooms, they cannot add walls, and

**two failed edits on the same defect means the track is DEAD; the concept regenerates fresh with

the requirement written as a GENERATION MANDATE, never a third edit.** The melody-variety pass was

edit one against this defect. The instruction in section 7 is edit two. If it fails, T2 is retired

with its scores ledgered and the surging-groove concept refires from the locked EXPLORATION

template with the melody requirement in the prompt's opening sentence:

The main melody must be a real tune with a rising-and-falling shape — no pitch repeated more than

twice in a row — harmonised in thirds and sixths, stated near the start and brought back at least

twice on different instruments.

The doctrine's own boundary applies in T2's favour and is why the edit gets a second try rather

than none: its structural bones are sound (structural yield 0.8333, which reads like real released

music), so this is a LOCAL defect on a track with good bones — exactly the case the ruling leaves

to Edit Song.

---

7. THE RECOMMENDED COMBINED INSTRUCTION — ready to paste

One instruction, multiple changes, per the multi-change law for Edit Song at measured seams. It

names note-level behaviour rather than adjectives, targets every failed axis at once, asks for

nothing already in band, and carries the return and the valley:

Rewrite the main melody line so it actually moves. No pitch may repeat more than twice in a row —

at the moment the tune sits on one note and pulses. Give it a clear rising-and-falling shape: each

eight-bar phrase should start lower, climb to its high point about a third of the way through, and

fall away to a resting note, with smaller rises and falls inside each two-bar group. Use a wider

set of intervals — seconds, thirds, fourths, fifths, and one or two wide upward leaps of a fifth

or more per phrase as the melody's signature gesture, stepping back down after each leap.

Harmonise the melody with a second instrument a third or a sixth away for most of its notes; do

not double it at the octave or in unison and do not move in parallel fifths. From 1:44 to 3:24 the

track currently holds unchanged — open a real quiet valley in there where the percussion drops

away and the melody is nearly alone, then build back. State the melody near the start, leave it,

and bring it back at least twice on a different instrument each time, and let the last phrase

settle and resolve rather than fading out.

Keep the fixed frame with it (unchanged from the locked EXPLORATION template, restated because

an edit prompt that drops them invites the engine to drop them too): instrumental only, fantasy

action-adventure RPG, the Flores highlands — bamboo flutes, sasando, gongs and drums — a caci

contest is a CONTEST and never a menace, and no vocal or lyric content.

If it must be split into two passes (the engine dilutes long instructions): pass 1 is the

melody rewrite — sentences one through four, the note-level ones, which carry all five failed axes.

Pass 2 is the valley, the returns and the settle. Do not reverse the order: a valley cut around a

line that does not move buys a quieter still track.

---

What this brief does not claim

• No listening test stands behind any of it. The wall is a structural instrument and Josh's ear is

the only bar that says good; his one-line rejection of the previous edit outranks every number

here.

• The tier is AUDIO-ESTIMATE throughout, on a line extracted from dense polyphony with 12.81% of

its notes octave-repaired. harmonization_density is a LOWER BOUND, so 2.5's reading is the

softest of the five.

• harmony_fit — the instrument that would settle the harmonisation question at score grain — is

NEEDS-SCORE and unrecoverable for a text-prompt engine. It is a named blind spot on this row.

• Whether the melody is passed BETWEEN instruments, and whether the ending settles, are UNMEASURED

BY ANY INSTRUMENT. Section 5 asks for both; only an ear can confirm either.

• The EXPLORATION band is n=8. It is the cue class's own band rather than a borrowed one, which is

the strong case, but eight exemplars is a small population and a p10 off eight rows is a soft

edge.

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