SUMMIT_TESTS_VALIDATION.md

music/exemplars/instruments/SUMMIT_TESTS_VALIDATION.md

The two summit tests -- the piano reduction and the orchestra lint

Canon subordination

This is a measurement artifact and it rules on nothing. Canon is authority; this file is subordinate to it and to the ruling it serves. The ruling is Josh's, verbatim, in docs/spine/DECISIONS_PENDING_JOSH.md under ROUND-2 FLOOR ADDENDUM: what will make people still listen thirty to fifty years later, and what will play in orchestras. Nothing below changes a pillar, a number, a name or a hard line. The bands are proposal-tier craft judgements and are named as such wherever they appear.

Measured instrument

• Tier -- MEASURED INSTRUMENT. Test one is a NECESSARY CONDITION for the transmission route nostalgia travels on, not a prediction of nostalgia. Test two is a NOTATION AND PHYSIOLOGY lint, not a judgement of the music. No listening test stands behind either.
• Instrument -- summit_tests v1; law: A theme must hold as melody plus harmony on one instrument -- the only route nostalgia travels -- and its score must be playable by real players from real notation.
• Denominator -- 30 corpus hits / 927 within-album controls / 1180 cards, live at HEAD
• Measured at 2026-08-07T21:53:41 in 17.9 s

The answer, in one paragraph

YES ON THE CORPUS AND NO ON ROUND 2, with one large caveat that belongs in the same breath. THE TEST IS CALIBRATED ON WHAT IT CAN SEE. 22 of 22 judgeable corpus rows pass the piano test on bands set from theory rather than from their own distribution. 8 of 30 rows could not be judged at all: the predominant-melody tracker recovered under one note every two seconds on those dense mixes, and a failure to measure is reported as one rather than scored as a defect in the music. Round 2 fails: 3 of 3 cues fail the piano test on the exact score, two of them because the longest span of NOVEL melodic material in a four-minute cue is under four bars and the third because the whole cue contains ONE melodic idea -- and the reduction recovers each cue's own head cell exactly, so that is the music and not an estimate. THE CAVEAT: the four continuous axes do not separate anything. Round 1, which Josh turned down in its entirety, passes 23 of 24 on those same axes, and round 2 measured off its own RENDER passes too. Every bit of this test's discriminating power on our material sits in the two phrase-grammar clauses, which need a score and therefore cannot be calibrated against the corpus at all. Two of the director's predictions also did not survive measurement and are reported below rather than quietly dropped: round 2's self-sufficiency comes back HIGH, near 0.85, because the head cell is a D major triad outline sitting over a D chord -- it is a well-formed MOTIVE, and what the cue lacks is a theme built out of it; and the cues do carry a cadence and an antecedent-consequent pair or two, in the counter-lines rather than in the head. The orchestra lint is the kinder half and it is unanimous: 12 hard findings across the three scores, 12 of them the SAME defect -- sustained string writing with no bow change, up to seventy-three seconds on one contrabass line. Nothing is out of range anywhere. So the scores are close to writable and the music is not yet a theme, which makes the next pass a composition problem rather than an orchestration one; the melody is also simply ABSENT for most of every cue (melodic coverage 33%, 46%, 30%).

Test one -- the piano reduction

What is measured, and on what

• The reduction is melody plus harmony on one instrument, and for our own cues it is EXACT -- compiled from the plan through the same compile path the render uses, with no extraction and no error. For the corpus there is no score, so the melody comes off the rig's polyphony-safe predominant-melody extractor and the harmony off a chroma template match.
• The extraction error is measured rather than asserted. monophonic extraction from dense polyphonic audio is unreliable, and the chroma harmony reads the whole mix including the melody. Our three round-2 cues have BOTH a score and a render, so the SAME cue is measured down both paths and the delta is the audio tier's error bar
• Audio minus exact on singability -- mean +0.211, largest absolute 0.302, over 3 cue(s) measured both ways.
• Audio minus exact on self sufficiency -- mean +0.087, largest absolute 0.149, over 3 cue(s) measured both ways.
• Audio minus exact on transmissibility -- mean -0.072, largest absolute 0.166, over 3 cue(s) measured both ways.
• Audio minus exact on reduction harmonic interest -- mean -0.018, largest absolute 0.054, over 3 cue(s) measured both ways.
• Audio minus exact on piano melody recovered -- mean -0.129, largest absolute 0.267, over 3 cue(s) measured both ways.

The bands, and why they are not the corpus's own percentiles

• set from music theory FIRST -- a comfortable vocal compass, a step-to-leap ratio, gap-fill after a leap, a breath-length phrase -- and only then checked against the corpus. A band derived from the corpus's own percentiles would have passed the corpus by construction and proved nothing
• Floors -- singability 0.45, self_sufficiency 0.45, transmissibility 0.45, reduction_harmonic_interest 0.2, piano_melody_recovered 0.3.
• The structural clause -- a tune is at least 4 bars. Two bars is a cell, which is the unit a phrase is built from.

The pools

What the pools say, read honestly

• THE FOUR CONTINUOUS AXES DO NOT SEPARATE GOOD MUSIC FROM REJECTED MUSIC. Round 1 was turned down in its entirety and 23 of its 24 judgeable tracks pass these axes; the album siblings, which are competent filler by construction, sit within a few hundredths of the tracks Josh named. Read the medians above as evidence that the axes are MEASURING SOMETHING REAL AND SHARED BY ALL COMPETENT MUSIC, and not as a ranking. This is the same shape of finding the structural-yield axis carries in floor_instruments.py, and it is stated here for the same reason.
• SO WHAT DOES THE WORK. On our own material it is the two phrase-grammar clauses, and only those. They need a score. They therefore have NO corpus calibration, and a reader must not treat the corpus pass rate as evidence for them.
• WHAT THE CORPUS PASS RATE ACTUALLY BUYS. It shows the four axes and their bands do not reject the music this program is trying to stand beside. That is a NECESSARY check and it was not guaranteed: on the first run these same bands failed twenty-one of thirty, all of it extractor artefact, and every repair is named in the module.

The corpus, row by row

The rows the extractor could not read, named

• Extractor coverage is 73%: 8 of 30 tracks recovered fewer than 0.7 notes per second across three minutes, which is feature_rig's own ideal floor for a melodic line. These are dense, loud mixes where no single voice dominates the salience.
• EX_007 Frog's Theme -- 0.1555 notes/s, salience 0.0
• EX_014 Tekken 2 Attract Movie: Sound Track -- 0.3721 notes/s, salience 0.4894
• EX_053 Cynthia Battle Theme -- 0.4158 notes/s, salience 0.3934
• EX_055 The Cyber Grind -- 0.3833 notes/s, salience 0.4181
• EX_064 Ludwig, the Holy Blade -- 0.4444 notes/s, salience 0.5684
• EX_094 Dark Creature Pursuit -- 0.3035 notes/s, salience 0.3547
• EX_142 Shinshu Field -- 0.1944 notes/s, salience 0.0
• EX_147 Fighting of the Spirit -- 0.3111 notes/s, salience 0.3403
• These are NOT failures. Scoring them would publish a verdict on a CQT and attach Frog's Theme's name to it. The honest reading is that a third of the corpus is out of reach of this extractor, and any future claim about the corpus as a whole has to carry that number.

Round 2 on the piano test, off the exact score

• PASS2_FLORES_CACI -- longest phrase of NOVEL material 3.50 bars < 4 -- a two-bar unit is a basic IDEA, so a cue built only of them contains no theme (MOTIVIC_DEVELOPMENT_AND_PHRASE_GRAMMAR.md:545)
• PASS2_FLORES_REST -- longest phrase of NOVEL material 3.75 bars < 4 -- a two-bar unit is a basic IDEA, so a cue built only of them contains no theme (MOTIVIC_DEVELOPMENT_AND_PHRASE_GRAMMAR.md:545)
• PASS2_FLORES_WATERFALL -- 1 distinct melodic idea(s) -- one idea cannot fill a period (MOTIVIC_DEVELOPMENT_AND_PHRASE_GRAMMAR.md:555)

The same three cues measured off their own renders, and why the gap matters

• Every cue that FAILS on its score PASSES on its render, and the whole of the difference is the phrase-grammar clauses, which an extracted line cannot support. That gap is the audio tier's error bar stated as a behaviour rather than as a number, and it is the single most important limitation in this file: measured the way the corpus has to be measured, round 2 looks fine.

Two director predictions that did not survive measurement

• SELF-SUFFICIENCY WAS PREDICTED TO FAIL AND IT PASSES, near 0.85 on all three cues. The reason is instructive rather than embarrassing: the head cell's first three notes are D, F sharp and A over a D chord, so the cell outlines its own harmony about as cleanly as a motive can. The cell is not the problem. The absence of anything built from it is, and that is what the phrase clauses catch.
• NO CADENCE AND NO ANTECEDENT-CONSEQUENT PAIR ANYWHERE was predicted; the measured counts are not zero. They are small and they sit in the counter-lines rather than in the head, which is consistent with the underlying claim -- the HEAD never cadences -- but the stronger phrasing does not hold as written, and the numbers are published as they came out.
• One more, in the other direction: the director measured no phrase longer than two bars, and this file measures 3.5 to 4.0 bars of novel material. Both describe one six-note cell; the difference is that the augmentation return states it at double note values, which is four bars of time carrying two bars of idea.

Test two -- the orchestra-playability lint

• SCORE-ONLY. It runs on our own cues and cannot run on the corpus, because a lint needs a part and a released stereo master has none. Its corpus calibration is limited to what the reference table itself can be checked against -- C12 (internal coherence) and C13 (every instrument our plans use has a row). No corpus lint number is reported, because there is no honest way to produce one
• no copy of Adler, Piston, Blatter or Sevsay is on disk in this repository; craft_research/ORCHESTRATION_TEXTURE_AND_DENSITY.md sections 2.3-2.5 cite all four externally. Each row NAMES the treatise whose published compass it follows and carries a confidence; NO PAGE NUMBER IS INVENTED, because a fabricated page would look checkable
• The reference carries 41 pitched and 10 unpitched rows, by confidence: consensus 8, outside 1, treatise 32.
• Inferred sample-map octave conventions, declared rather than absorbed -- glockenspiel +12 semitones; xylophone +12 semitones. These are claims about the SAMPLE SET, never about the instrument.

Round 2, per cue

The longest-span column counts only parts that must BREATHE or CHANGE BOW. A gong ringing for four minutes is not a finding, and while it owned this column it hid a seventy-three-second contrabass line underneath it.

Every finding, with its location

PASS2_FLORES_CACI

• SOFT divisi_overflow -- part B01 (gong.stroke), bar 1 beat 1.0 (0.0 s): 2 simultaneous notes (C4, D4) on gong / kempul family, which one player can sound 1 of; this part needs a second player on the stand
• SOFT ledger_extreme -- part B09 (organ.quiet), bar 30 beat 1.0 (47.027 s): about 8 ledger lines beyond the grand staff (nearest: treble), written pitch; this part wants an 8va or a clef change to be readable
• HARD no_bow_change -- part B10 (cello_ens.sus_quiet), bar 73 beat 1.0 (116.757 s): 25.9 s of unbroken sound with no gap of 0.40 s to change bow; the stated human limit is 12 s (a whole bow at the slowest usable speed)
• SOFT ledger_extreme -- part B10 (cello_ens.sus_quiet), bar 39 beat 1.0 (61.622 s): about 8 ledger lines beyond the bass staff (nearest: bass), written pitch; this part wants an 8va or a clef change to be readable
• HARD no_bow_change -- part B11 (violin_ens.sus_quiet), bar 29 beat 1.0 (45.405 s): 25.9 s of unbroken sound with no gap of 0.40 s to change bow; the stated human limit is 12 s (a whole bow at the slowest usable speed)
• SOFT outside_good_compass -- part B12 (viola_ens.sus_quiet), bar 33 beat 1.0 (51.892 s): 15 of 60 notes sit outside viola section's GOOD compass C3-A5; playable, but the tone thins [Walter Piston, Orchestration (Norton, 1955)]
• HARD no_bow_change -- part B12 (viola_ens.sus_quiet), bar 73 beat 1.0 (116.757 s): 25.9 s of unbroken sound with no gap of 0.40 s to change bow; the stated human limit is 12 s (a whole bow at the slowest usable speed)
• SOFT ledger_extreme -- part B12 (viola_ens.sus_quiet), bar 33 beat 1.0 (51.892 s): about 6 ledger lines beyond the alto staff (nearest: alto), written pitch; this part wants an 8va or a clef change to be readable
• SOFT ledger_extreme -- part M02 (vibraphone.hard), bar 26 beat 1.0 (40.541 s): about 5 ledger lines beyond the treble staff (nearest: treble), written pitch; this part wants an 8va or a clef change to be readable
• SOFT outside_good_compass -- part M05 (glockenspiel), bar 33 beat 1.0 (51.892 s): 26 of 104 notes sit outside glockenspiel (sounding)'s GOOD compass G5-A7; playable, but the tone thins [Samuel Adler, The Study of Orchestration (Norton, 4th ed. 2016)]
• HARD no_bow_change -- part M08 (violin_ens.spic), bar 73 beat 1.0 (116.757 s): 25.7 s of unbroken sound with no gap of 0.40 s to change bow; the stated human limit is 12 s (a whole bow at the slowest usable speed)
• SOFT ledger_extreme -- part M10 (harp), bar 67 beat 1.0 (107.027 s): about 11 ledger lines beyond the grand staff (nearest: treble), written pitch; this part wants an 8va or a clef change to be readable
• SOFT outside_good_compass -- part M12 (folk_harp), bar 81 beat 1.0 (129.73 s): 23 of 112 notes sit outside lever harp's GOOD compass C2-G6; playable, but the tone thins [Alfred Blatter, Instrumentation and Orchestration (2nd ed.)]
• SOFT ledger_extreme -- part M12 (folk_harp), bar 81 beat 3.0 (130.541 s): about 8 ledger lines beyond the grand staff (nearest: treble), written pitch; this part wants an 8va or a clef change to be readable
• HARD no_bow_change -- part X01 (contrabass.trem), bar 61 beat 1.0 (97.297 s): 45.4 s of unbroken sound with no gap of 0.40 s to change bow; the stated human limit is 12 s (a whole bow at the slowest usable speed)
• SOFT ledger_extreme -- part X04 (horn.mute), bar 85 beat 1.0 (136.216 s): about 8 ledger lines beyond the treble staff (nearest: treble), written pitch; this part wants an 8va or a clef change to be readable
• SOFT ledger_extreme -- part F06 (flute.sus_vib), bar 78 beat 1.0 (124.865 s): about 6 ledger lines beyond the treble staff (nearest: treble), written pitch; this part wants an 8va or a clef change to be readable
• SOFT breathing_tight -- part F07 (clarinet.sus), bar 93 beat 1.0 (149.189 s): 13.0 s unbroken; playable but past the comfortable 10 s and a player will want a mark here
• SOFT ledger_extreme -- part F08 (flute.sus_vib), bar 110 beat 1.0 (176.757 s): about 6 ledger lines beyond the treble staff (nearest: treble), written pitch; this part wants an 8va or a clef change to be readable

PASS2_FLORES_REST

• SOFT divisi_overflow -- part B01 (gong.stroke), bar 1 beat 1.0 (0.0 s): 2 simultaneous notes (C4, D4) on gong / kempul family, which one player can sound 1 of; this part needs a second player on the stand
• SOFT ledger_extreme -- part B08 (organ.quiet), bar 22 beat 1.0 (54.783 s): about 8 ledger lines beyond the grand staff (nearest: treble), written pitch; this part wants an 8va or a clef change to be readable
• HARD no_bow_change -- part B09 (contrabass.sus_quiet), bar 29 beat 1.0 (73.043 s): 73.0 s of unbroken sound with no gap of 0.40 s to change bow; the stated human limit is 12 s (a whole bow at the slowest usable speed)
• HARD no_bow_change -- part B10 (cello_ens.sus_quiet), bar 65 beat 1.0 (166.957 s): 52.2 s of unbroken sound with no gap of 0.40 s to change bow; the stated human limit is 12 s (a whole bow at the slowest usable speed)
• SOFT ledger_extreme -- part B10 (cello_ens.sus_quiet), bar 42 beat 1.0 (106.957 s): about 8 ledger lines beyond the bass staff (nearest: bass), written pitch; this part wants an 8va or a clef change to be readable
• SOFT outside_good_compass -- part B11 (viola_ens.sus_quiet), bar 49 beat 1.0 (125.217 s): 8 of 32 notes sit outside viola section's GOOD compass C3-A5; playable, but the tone thins [Walter Piston, Orchestration (Norton, 1955)]
• HARD no_bow_change -- part B11 (viola_ens.sus_quiet), bar 65 beat 1.0 (166.957 s): 52.2 s of unbroken sound with no gap of 0.40 s to change bow; the stated human limit is 12 s (a whole bow at the slowest usable speed)
• SOFT ledger_extreme -- part B11 (viola_ens.sus_quiet), bar 51 beat 1.0 (130.435 s): about 6 ledger lines beyond the alto staff (nearest: alto), written pitch; this part wants an 8va or a clef change to be readable
• SOFT ledger_extreme -- part M02 (vibraphone.soft), bar 22 beat 1.0 (54.783 s): about 5 ledger lines beyond the treble staff (nearest: treble), written pitch; this part wants an 8va or a clef change to be readable
• SOFT ledger_extreme -- part M04 (harp), bar 22 beat 1.0 (54.783 s): about 11 ledger lines beyond the grand staff (nearest: treble), written pitch; this part wants an 8va or a clef change to be readable
• SOFT outside_good_compass -- part M05 (folk_harp), bar 41 beat 1.0 (104.348 s): 30 of 160 notes sit outside lever harp's GOOD compass C2-G6; playable, but the tone thins [Alfred Blatter, Instrumentation and Orchestration (2nd ed.)]
• SOFT ledger_extreme -- part M05 (folk_harp), bar 42 beat 3.0 (108.261 s): about 8 ledger lines beyond the grand staff (nearest: treble), written pitch; this part wants an 8va or a clef change to be readable
• SOFT outside_good_compass -- part M07 (glockenspiel), bar 45 beat 1.0 (114.783 s): 8 of 32 notes sit outside glockenspiel (sounding)'s GOOD compass G5-A7; playable, but the tone thins [Samuel Adler, The Study of Orchestration (Norton, 4th ed. 2016)]
• DECLARED range_declined -- part M10 (strumstick), bar 0 beat 0.0 (0.0 s): strumstick (three-string diatonic folk dulcimer): not an orchestration-reference instrument. RANGE LINT DECLINED: there is no published compass to lint against and inventing one would be worse than the gap
• SOFT ledger_extreme -- part F03 (flute.sus_vib), bar 13 beat 5.0 (33.913 s): about 6 ledger lines beyond the treble staff (nearest: treble), written pitch; this part wants an 8va or a clef change to be readable
• HARD no_bow_change -- part F04 (violin_solo.vib), bar 49 beat 1.0 (125.217 s): 20.9 s of unbroken sound with no gap of 0.40 s to change bow; the stated human limit is 12 s (a whole bow at the slowest usable speed)

PASS2_FLORES_WATERFALL

• HARD no_bow_change -- part B02 (contrabass.sus_quiet), bar 23 beat 1.0 (80.0 s): 65.5 s of unbroken sound with no gap of 0.40 s to change bow; the stated human limit is 12 s (a whole bow at the slowest usable speed)
• SOFT ledger_extreme -- part M01 (hand_chimes), bar 6 beat 3.0 (20.0 s): about 9 ledger lines beyond the treble staff (nearest: treble), written pitch; this part wants an 8va or a clef change to be readable
• SOFT ledger_extreme -- part M03 (harp), bar 14 beat 2.5 (48.636 s): about 11 ledger lines beyond the grand staff (nearest: treble), written pitch; this part wants an 8va or a clef change to be readable
• HARD no_bow_change -- part M04 (cello_ens.sus_quiet), bar 31 beat 1.0 (109.091 s): 36.4 s of unbroken sound with no gap of 0.40 s to change bow; the stated human limit is 12 s (a whole bow at the slowest usable speed)
• SOFT ledger_extreme -- part M04 (cello_ens.sus_quiet), bar 11 beat 1.0 (36.364 s): about 8 ledger lines beyond the bass staff (nearest: bass), written pitch; this part wants an 8va or a clef change to be readable
• SOFT breathing_tight -- part F01 (recorder.tenor), bar 1 beat 1.0 (0.0 s): 14.5 s unbroken; playable but past the comfortable 10 s and a player will want a mark here
• SOFT outside_good_compass -- part F02 (recorder.alto), bar 15 beat 1.0 (50.909 s): 1 of 6 notes sit outside alto recorder in F's GOOD compass F4-E6; playable, but the tone thins [Alfred Blatter, Instrumentation and Orchestration (2nd ed.)]
• SOFT ledger_extreme -- part F02 (recorder.alto), bar 16 beat 1.0 (54.545 s): about 6 ledger lines beyond the treble staff (nearest: treble), written pitch; this part wants an 8va or a clef change to be readable
• HARD no_bow_change -- part M05 (violin_solo.vib), bar 25 beat 1.0 (87.273 s): 14.5 s of unbroken sound with no gap of 0.40 s to change bow; the stated human limit is 12 s (a whole bow at the slowest usable speed)

The listenable reduction

• there is no piano in the rung-one palette (sfz_palette.PALETTE, 23 instruments, one keyboard); organ.quiet is the only keyboard voice and is used for the listenable reduction. Its manual compass 36-96 covers both hands of these reductions with nothing folded
• PASS2_FLORES_CACI -- D:/audio/pass2/PASS2_FLORES_CACI/PASS2_FLORES_CACI.reduction.wav (107 melody notes, 240 harmony notes, 0 octave folds)
• PASS2_FLORES_REST -- D:/audio/pass2/PASS2_FLORES_REST/PASS2_FLORES_REST.reduction.wav (81 melody notes, 184 harmony notes, 0 octave folds)
• PASS2_FLORES_WATERFALL -- D:/audio/pass2/PASS2_FLORES_WATERFALL/PASS2_FLORES_WATERFALL.reduction.wav (30 melody notes, 80 harmony notes, 0 octave folds)

What this does not say

• Nothing here predicts nostalgia. Test one measures a NECESSARY CONDITION for the route nostalgia travels on -- a tune that survives being reduced can be hummed, played on a school piano and arranged by a fan; one that cannot, cannot. A track can pass this test and still be forgettable.
• Nothing here is a listening test. No human has heard a single number in this file.
• Test two says a part is WRITABLE, not that it is well orchestrated. A lint with no findings on a boring part is a clean lint on a boring part.
• The corpus numbers carry an extraction error that our own dual-path measurement puts on the record; the round-2 numbers do not, because those are the score itself.

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