MELODY_BANDS_VALIDATION.md

music/exemplars/instruments/MELODY_BANDS_VALIDATION.md

MELODY BANDS -- validation

melody_bands melody-bands/1.0 -- generated 2026-08-08T19:53:22Z

MEASURED INSTRUMENT. Nothing here has been listened to. Every verdict below is a statistic about a pool, never a judgement about a tune.

The grade this file answers to

these main melodies really suck. They clash and have no rythm or complexity. Just a few notes slowly played back to back. Never any harmonies added on. These are more complex than round 2 but they're actually worse tracks.

-- Josh, on round 3, 2026-08-08. The battery that graded that round 57/57 measured none of it.

Pool

Multiplicity bar: a within-album |AUC-0.5| must beat 0.1881 (p95 of the best-of-bank deviation over 400 label permutations).

Per-axis verdict

s = score tier, a = audio tier -- whichever deviates further from chance is shown; both are in the JSON.

• SEPARATES -- within-album deviation beats the permutation multiplicity p95 bar -- the axis tells a beloved track from its own album siblings
• DEFECT-DETECTOR-ONLY -- does not clear that bar, but separates the corpus from at least one of OUR generated rounds by |AUC-0.5| >= 0.25. It detects machine-made, not good. Same label structural_yield carries in floor_instruments.py
• NULL -- neither. Reported plainly; five consecutive nulls have been published in this program and were the right answer each time
• density_proxy -- |Spearman| > 0.85 against voiced_fraction or note_count -- the axis is extraction density wearing a musical name (the phrase_count defect, rho 0.928)
• density_by_construction -- the axis IS a density and says so; the flag is declared, not excused

The grade, clause by clause

"no rythm or complexity / just a few notes slowly played back to back"

axes built for it: note_density_per_s, note_density_per_sounding_s, phrase_len_s_median, distinct_durations

• BELOW band: PASS5_FLORES_FALLS/score note_density_per_sounding_s 1.2997 (p10 2.5369)
• BELOW band: PASS5_FLORES_ROAD/score note_density_per_s 1.0293 (p10 1.1978)
• BELOW band: PASS5_FLORES_ROAD/score note_density_per_sounding_s 1.8036 (p10 2.2247)
• BELOW band: PASS5_FLORES_ROUNDS/score note_density_per_sounding_s 2.585 (p10 6.0917)
• BELOW band: PASS5_FLORES_ROUNDS/audio note_density_per_sounding_s 4.7117 (p10 6.0917)
• BELOW band: PASS5_FLORES_ROUNDS/audio phrase_len_s_median 2.009 (p10 3.0556)
• ABOVE band: PASS5_FLORES_FALLS/audio distinct_durations 13.0 (p90 12.0)
• ABOVE band: PASS5_FLORES_ROUNDS/score note_density_per_s 1.9694 (p90 0.8645)
• ABOVE band: PASS5_FLORES_ROUNDS/audio note_density_per_s 1.6833 (p90 0.8645)
• ABOVE band: PASS5_FLORES_ROUNDS/audio distinct_durations 10.0 (p90 8.2)

"they clash"

axes built for it: step_share, leap_share, interval_class_count, contour_turning_rate

• BELOW band: PASS5_FLORES_ROAD/score step_share 0.393 (p10 0.5048)
• BELOW band: PASS5_FLORES_ROUNDS/audio step_share 0.6623 (p10 0.7129)
• ABOVE band: PASS5_FLORES_FALLS/score leap_share 0.2083 (p90 0.1418)
• ABOVE band: PASS5_FLORES_FALLS/score interval_class_count 12.0 (p90 11.0)
• ABOVE band: PASS5_FLORES_FALLS/score contour_turning_rate 0.5979 (p90 0.5027)
• ABOVE band: PASS5_FLORES_ROAD/score interval_class_count 12.0 (p90 10.3)
• ABOVE band: PASS5_FLORES_ROAD/audio interval_class_count 11.0 (p90 10.3)
• ABOVE band: PASS5_FLORES_ROUNDS/audio interval_class_count 9.0 (p90 8.2)
• ABOVE band: PASS5_FLORES_ROUNDS/audio contour_turning_rate 0.4917 (p90 0.4876)

"never any harmonies added on"

axes built for it: harmonization_density, parallel_harmony_share

NOT CORROBORATED IN THE DIRECTION THE CLAUSE STATES. Round 3 is harmonised ABOVE the corpus band on both harmonisation axes at the score tier and at or above the corpus median at the audio tier. The harmony is in the score and it is in the render's band structure. Whatever the ear is missing, it is not concurrent melodic voices at consonant intervals -- which redirects the diagnosis rather than confirming it, and is the more useful outcome than an axis tuned until it agreed.

• ABOVE band: PASS5_FLORES_FALLS/score harmonization_density 0.9774 (p90 0.9488)
• ABOVE band: PASS5_FLORES_FALLS/score parallel_harmony_share 0.8454 (p90 0.7056)
• ABOVE band: PASS5_FLORES_ROAD/score harmonization_density 0.9473 (p90 0.9346)
• ABOVE band: PASS5_FLORES_ROAD/score parallel_harmony_share 0.8348 (p90 0.6548)
• ABOVE band: PASS5_FLORES_ROUNDS/score harmonization_density 0.9665 (p90 0.9488)

The corpus bands

BATTLE (n = 5)

EXPLORATION (n = 8)

POOLED_SMALL_CLASSES (n = 17, POOLED)

each of these classes carries fewer than 5 corpus hits, so it gets no band of its own; they are POOLED and this is said out loud rather than one class silently borrowing another's percentiles

members: {'BOSS': 2, 'CREDITS_TRIUMPH': 4, 'MELANCHOLIC': 4, 'TENSION': 3, 'TITLE_CHARACTER': 4}

ALL (n = 30, POOLED)

every hit, regardless of class -- the fallback band for a cue whose purpose class the corpus does not carry

WHERE ROUND 3 FALLS OUTSIDE THE CORPUS BANDS

The same table for round 2 and round 1, for the comparison

Controls

21 / 21 passed at the run that produced this file.

• [PASS] K0 measure({}) returns available=False without raising
• [PASS] K0b the module exposes the battery's five names
• [PASS] K1 six notes over 30 s measure 0.20 notes/s
• [PASS] K1b ... and that is BELOW the corpus p10 density floor -- Josh's complaint FAILS
• [PASS] K2 a line of one repeated interval reads interval_class_count == 1
• [PASS] K2b ... and its step_share is 1.0 with leap_share 0.0
• [PASS] K3 two voices in parallel thirds read harmonization_density ~ 1.0 over TWO voices
• [PASS] K3b ... and parallel_harmony_share ~ 1.0
• [PASS] K3c a monophonic line reads harmonization_density 0.0
• [PASS] K3d an OCTAVE doubling folds to one voice and reads 0.0 -- doubling is not harmony
• [PASS] K4 a slow lead over a busy bed still reads a slow LINE density
• [PASS] K5 four rest-separated phrases segment into more than one phrase
• [PASS] K5b ... with a phrase peak position inside (0,1)
• [PASS] K6 an arch turns direction and a ladder barely does
• [PASS] M1 CONTROL: counting every voice's onsets breaks the slow-lead reading (mutation _MUT_DENSITY_FROM_ALL_VOICES)
• [PASS] M2 CONTROL: an interval-blind doubling fold destroys the parallel-thirds reading (mutation _MUT_FOLD_ANY_DOUBLING)
• [PASS] M3 CONTROL: refusing to split a line collapses the phrase axes (mutation _MUT_NO_PHRASE_SPLIT)
• [PASS] M4 CONTROL: dropping the consonance test lets a semitone clash read as harmony (mutation _MUT_IGNORE_CONSONANCE)
• [PASS] M5 CONTROL: unrounded intervals shatter the interval-class count (mutation _MUT_NO_INTERVAL_ROUNDING)
• [PASS] K7 the delegated line extractor still contains _octave_repair (read off its code object)
• [PASS] K7b the score tier's melodic-kind set is pass3_score.MELODIC_KINDS, which includes 'score' (instr_summit_tests.MELODY_KINDS :585 does NOT, and is blind to round 3)

Known blindness

• instr_summit_tests.MELODY_KINDS :585 omits 'score', so plan_reduction returns ZERO melody notes for every round-3 plan; that instrument's score tier cannot see round 3. This file uses pass3_score.MELODIC_KINDS :230 instead.
• audio-tier harmonization_density is a LOWER BOUND: two instruments a third apart in one register are one band to mi.band_lines :355.
• round-3 cue PASS5_FLORES_FALLS declares purpose class SANCTUARY, which the corpus does not carry; it is scored against the ALL band and that is said on its row.
• EXTRACTION DROPOUT contaminates note_density_per_s's floor. Six of the thirty corpus rows give the tracker a line over under a tenth of the cue (Frog's Theme: 13 notes), which drags the p10 to 0.31 notes/s. note_density_per_sounding_s divides by the time the line sounds and is immune to it -- read that one.
• BATTLE clears the class floor on rows (n=5) but carries only n=3 on the two harmonisation axes; _axis_band falls back to the ALL band FOR THOSE AXES and names the substitution on every row rather than judging against three tracks.

_elapsed 3.2 s_

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