PATTERN_FINDINGS_V3.md

music/exemplars/PATTERN_FINDINGS_V3.md

PATTERN FINDINGS V3 — the derivation re-run at N = 30

**Tier: MEASURED-SAMPLE FINDINGS. Four axes clear a re-calibrated multiplicity bar, down from

six, and the rejection envelope the generator is using has changed two of its three axes.**

Everything below is derived from the formula cards already on disk. Nothing was purchased,

nothing was generated, no audio was decoded by this pass. The cards are read as data; the

exemplar bytes never move.

Supersedes: PATTERN_FINDINGS_V2.md (N = 28, six axes clearing). V2 is not withdrawn — its

method is unchanged and it remains the correct answer for the pool it had, exactly as V2 says of

V1. What changed is the pool, not the apparatus.

Produced by: harness/music_gen/derive_patterns.py (26 controls, all green, mutation-proved).

Machine record: build/audio/exemplars/PATTERN_FINDINGS_V3.json.

Consumer: docs/proposals/music/MASTERPIECE_PROGRAM.md rung 2 (the archetype synthesis) and

rung 3 (the nostalgia predictor); and directly, harness/music_gen/slice_exemplars.py, whose

KEEP/REJECT filter reads the joint envelope this document re-derives.

Legal posture: analysis for understanding only, unchanged from the cards themselves.

1. The one-paragraph answer

Start with what got weaker, because that is the reason this re-run was owed. Two positives

were added — EX_014 and EX_094, whose quarantines were retired by inspection on 2026-08-07 —

and on the larger pool two of V2's six clearing axes fall back below the bar, bars and

n_drops; no axis clears the 99th percentile at all, where two did at N = 28; the leave-one-out

composite falls from 0.744 to 0.707; the length-independent composite falls from an already

sobering 0.497 to 0.426, which is now below chance rather than at it; and the joint rejection

envelope replaces two of its three axes, which means every KEEP/REJECT verdict in the round-1

generated slice was taken under a filter this document supersedes. One thousand one hundred and

eighty tracks are measured; thirty are corpus rows against twenty-eight in V2; nine hundred and

twenty-seven are their own album siblings. The bar itself fell from 0.195 to 0.184, so the

narrowing is not the bar rising — it is the observed effects shrinking faster than the bar fell.

The core V2 finding survives in shape: **structural size — how many sections a track has, how long

it runs, how much it repeats itself, how little it renews — still separates a corpus row from its

own album at 0.71 held-out, and still does so in both measured eras.** But it survives with less

room, and one 21.5-second attract sting is responsible for essentially all of the movement.

What moved, and by how much

Five rows changed a conclusion. The two that matter most are the last two, and they are the

subject of section 8.

2. The design, unchanged, and what N = 30 bought

Each corpus row is scored only against the tracks that shipped on its own record — composer,

scoring session, mixing chain, sample library, release year and audio codec held constant by

construction. Per-hit win rates are averaged with equal weight per hit, so a 208-track Pokémon

compilation cannot drown a four-track Bloodborne EP. This is the only framing the pool can answer

honestly, and it is V1's framing and V2's framing verbatim. The feature bank is the same 103 axes.

Nothing in the apparatus was touched for this pass.

The pool, all three passes

The card count falling by one while the pool grew is not an error and it reconciles exactly. Two

byte-identical UNMATCHED__ twins were retired by acquire_exemplar.py — one of the Tekken 2

attract cue, one of Rayman Legends' "Dark Creature Pursuit" — because each was a second card over

the same audio as a newly-attached hit, and scoring a hit against a copy of itself pulls every

within-album AUC toward the tie value. That is minus two. EX_094 gained a card it did not

previously have, which is plus one. EX_014's false card was repudiated and rewritten at net zero.

The control pool grew by exactly 76 — the Tekken 2 album's 31 siblings and Rayman Legends' 45 —

which moved out of the carried-but-unused pool the moment those albums carried a hit.

The two new positives, and why they are admissible

Both were quarantined in V2 and both quarantines were retired on evidence rather than assumption,

which this pass verified against CORPUS.json rather than taking on trust.

• EX_014 — Josh named a Tekken 2 "Opening Theme". The album carries no such track. It was already

on disk in full from the starter-ten sweep, and its only attract cue is track 01, "Tekken 2

Attract Movie: Sound Track". The prior attach, which measured a Pokémon X & Y opening movie, was

demoted with its reason logged before the new one landed, so the false claim is repudiated rather

than silently overwritten. The row now reads CARDED on a human's inspection.

• EX_094 — Josh named "Escape from Vector", which is on no Rayman Legends release. It was re-keyed

onto "Dark Creature Pursuit", the album's longest timed-chase cue and the TENSION reading of the

row, with four runners-up named on the row itself and override_open set.

Both are legitimate positives and this pass treats them as such. It is nevertheless a measured fact,

not an opinion, that EX_014 is a **21.5-second attract sting with 4 sections, 8 bars, 8 notes and

a 7-semitone melodic span** — and that the entire V2 finding is that corpus rows are structurally

large. Section 8 quantifies what that one row does.

The thirty, by class

EXPLORATION 8, BATTLE 5, TITLE_CHARACTER 4, CREDITS_TRIUMPH 4, MELANCHOLIC 4, TENSION 3, BOSS 2.

All seven purpose classes remain represented; TITLE_CHARACTER and TENSION each gained a row. Five

of seven are still below the five-row floor the MASTERPIECE_STANDARD's per-class bands need, and

BOSS at two is still the binding constraint. No per-class claim in this document is a test; the

per-class block in the JSON remains a descriptive read and is labelled as one.

3. The multiplicity price, re-derived for the new N

The bar is not guessed and it is not inherited. Within each album, re-draw at random which tracks

are "hits" (keeping each album's count), recompute every feature's within-album AUC, and keep the

best deviation from 0.5 across the whole bank. Four hundred draws under seed 20260807 — the

same draw count and the same seed discipline V2 used, stated here so the comparison is exact — and

re-run from scratch on the new pool rather than carried.

Features clearing the 95th-percentile bar: four of 103. Clearing the 99th: none. At N = 28 it

was six and two; at N = 11 it was zero and zero.

The bar moved down, from 0.195 to 0.184, and the reason is the one V1 §8 predicted and V2 §3

restated: the null distribution of a best-of-bank statistic narrows as the number of positives

grows, because each permuted draw averages over more hits. Two more positives is a small change and

it bought a small fall. What did not happen is the effects growing to meet it. The observed best

fell from 0.250 to 0.215 — more than the bar fell — and that is the whole arithmetic of losing two

axes.

4. The ranking

Do not read this table as "the set shrank from six to four." Read the margins. The six axes span

deviations from 0.2150 down to 0.1751 — a band 0.04 wide — and the bar at 0.1841 falls inside

it. repeat_sim_mean clears by 0.0009. n_drops misses by 0.0078. bars misses by 0.0090. These

are not four findings and two refutations; they are six axes smeared across a threshold, and at

this N the threshold is a coin toss for the middle four of them. The honest statement is that

section_count and total_length_s clear with room, and that the other four sit close enough to

the bar that their status will change again with the next two acquisitions in either direction.

novelty_mean still clears in the negative direction and is still the one non-size axis in the

group. Corpus rows sit below their siblings on mean structural novelty while sitting above them on

repeat_sim_mean; the pair says the same thing twice, and it is thematic return rather than

repetition — a track that keeps coming back to its own material instead of moving on.

Tempo is still not the confound, and this is the one control that got slightly stronger.

Rescoring against siblings within ±20% bpm (517 controls) moves section_count 0.715 → 0.717,

total_length_s 0.700 → 0.710, n_drops 0.676 → 0.693 and bars 0.675 → 0.691. Every size axis

holds or improves. "Battle themes are fast and battle themes are over-represented" has now failed

to explain the pattern at three separate values of N.

5. Duration: still the confound, and the hit pool is no longer uniformly long

V1's §4 verdict was blunt and correct for its pool: the top of the table is duration wearing a

costume, and the duration effect is curation — Josh named finished themes while game albums are

full of jingles, stingers and menu blips. That remains true. The control pool still carries 162

tracks under 30 seconds and 270 under a minute, against 927 in total.

What changed is on the other side of the comparison. V2 could say the hit pool held "exactly one row

under a minute". It now holds two: EX_102 (Magus Castle, 28.6 s) and EX_014 (21.5 s). The

corpus is no longer a set of uniformly finished-length themes, and the axis that most directly

encodes finished length is the axis that suffers most.

Read the last column. Raw duration is still the axis that dies — 0.700 down to 0.553, closer to

chance than V2's 0.581 — which is what "curation, not composition" looks like when you take the

stingers away. section_count, n_drops, repeat_sim_mean and novelty_mean again decay more

slowly and sit around 0.59–0.62 against siblings every bit as long as the hits, against V2's

0.62–0.65. The residual that is not length is real and it is smaller than it was.

This truncation is a robustness read, not a test. The permutation bar in §3 was calibrated on

the full control pool; a truncated pool has fewer and differently-distributed siblings, so its own

bar would sit somewhat higher and is not computed here. Nothing in this table is claimed to clear

anything.

6. The carried hypotheses, the pre-registered four, and two new axes offered this sitting

6a. The three carried hypotheses stay dead

recurrence_lag_norm is now the single worst-separating axis of 103, at 0.4997. Its descriptive

medians are hit 0.019 against sibling 0.084 — the same inversion V2 found, deeper. V1's best

surviving candidate is now the bank's floor. Nothing here argues for a fourth test.

A parallel lane reached the same verdict from the opposite direction on the same day, and the two

results should be read together. Its finding, recorded in the 2026-08-07 block appended to V2 §6 and

implemented in harness/music_gen/instr_hook_presence.py, is that this axis was never a clean test:

it runs on raw chroma and MFCC self-similarity, so a theme restated a fourth higher or at half speed

scored as new material rather than as a return, which is backwards for a leitmotif score. Rebuilt

with transposition and tempo invariance it fails again and harder — zero of fourteen metrics clear

that lane's own re-derived bar. **The axis is dead both as measured and as it should have been

measured**, which is a stronger retirement than either pass could issue alone.

6b. The four pre-registered hypotheses, still all failed

Their own four-feature null puts the bar at deviation 0.150 (AUC 0.650). None comes near it. The

four theories written into the extractor before the first run are, at thirty positives, measurably

not what distinguishes these tracks.

6c. Two new axes, priced at the same bar as everything else

A director derivation this sitting proposed structural yield — novelty-peak boundaries divided

by texture drops, over cards of 45 seconds or more — reporting AUC 0.911 separating the tracks Josh

named from eighteen rejected generated candidates, and 0.437 against those candidates' own siblings.

It was offered with the prediction that it would fail the within-album bar. It was added to the

feature bank and priced through permutation_null_best exactly like every other axis, alongside

boundaries_per_min. The prediction is correct and the axes fail, plainly.

Neither reaches a quarter of the bar. structural_yield is 0.038 against a required 0.184, and its

ungated form is 0.005 — the closest any axis in this document comes to exact chance. Both hold near

chance in both eras (structural_yield 0.488 sixteen-bit against 0.440 modern). Adding both to the

bank and re-running the four-hundred-draw null changes the bar not at all: median 0.141, 95th

percentile 0.184, maximum 0.238, identical to four decimal places, because neither axis ever wins a

permuted draw.

Two further facts belong on the record rather than in a footnote.

• structural_yield would not have been admitted to the bank at all under the tool's own rule.

Its 45-second gate leaves it undefined on 23% of the pool, and feature_bank requires 80%

coverage. It was priced here by explicit construction, not by admission.

• The gate also makes its AUC not comparable to the other 103. An axis defined only on cards of

45 seconds or more is scored against a length-truncated sibling pool, which is a different and

easier control group than every other axis faces. It fails anyway.

The director's reading of why is the correct one and this pass confirms it: a ratio of boundaries to

drops separates our generator from real music, not loved music from filler. Against album siblings —

real music by the same composer in the same session — there is nothing there. A director's finding

gets the same bar as anyone's, and this one does not clear it.

7. Held-out validation — three ways, all reported

7a. Frozen split, selection on train only

Sixteen hits train, fourteen test. The six best training axes were turned into thresholded rules by

Youden's J on the training data, then scored once on the untouched test side.

**Zero of six rules reverse sign, and one of six now fires on unseen corpus rows less often than on

filler.** Those two sentences are not in conflict; they are two different definitions of reversal,

and V2 used the second without naming it, so both are reported here and in the JSON.

• By AUC sign — train and test AUC on opposite sides of 0.5 — V1 reversed 6 of 6, V2 reversed

0 of 6, and V3 reverses 0 of 6. Unchanged and good.

• By held-out lift below 1 — the rule is worse than nothing on the test side — V1 failed 3 of 6,

V2 failed 0 of 6, and V3 fails 1 of 6. drop_depth_max_db ≤ −59.67 carries a held-out lift of

0.55×. This is V2's own "three of six reversed at N = 11" statistic, and it has gone from zero

back to one.

The selection also got noisier in a way worth naming: two of the six axes chosen on the training

half are now negative-direction axes, where V2's six were all positive-direction size axes. Four

rules still carry held-out lifts of 2.0–3.1×, against V2's 2.7–3.9×.

7b. Does section_count ≥ 15 survive at N = 30?

It is the strongest held-out rule in the program and everything downstream leans on it, so it was

tested as written rather than as refitted.

The rule survives, and it survives better than the threshold Youden's J now prefers. At N = 30

the training half has sixteen hits and ≥ 14 wins J there on recall (train TPR 0.81 against 0.75),

but on the untouched test side ≥ 15 is the better rule on every metric that matters — same recall,

lower false-positive rate, 3.15× lift against 2.62×. Across the whole pool it fires on 19 of 30

corpus rows and 21.3% of the 927 siblings, a 2.98× lift.

What did weaken is its recall and its axis. Test TPR fell from 0.62 to 0.50, and section_count's

held-out AUC fell from 0.783 to 0.656. The rule is still the best single statement this program has;

it now catches half the unseen corpus rows rather than five-eighths. **Keep the threshold at 15. Do

not refit it to 14 on this pool.**

7c. Leave-one-out, rule re-selected inside every fold

Essentially unchanged, and this is the most stable result in the document. The rule selected is

section_count ≥ 14 in 28 of the 30 folds.

7d. Leave-one-out composite — the honest headline

In each fold the top five axes and their directions are chosen on the other twenty-nine hits alone,

z-scored within album, summed, and the held-out track's percentile among its own siblings recorded.

The length-independent bank is defined with no reference to the hit labels — the correlation with

duration is measured on the control pool alone (|ρ| < 0.35), admitting 70 of 103 axes, the same

count and the same filter as V2. V1's leak stayed closed.

**0.707 ± 0.053 is the honest realized predictive power of the whole measured card on this pool, and

0.426 ± 0.046 is what is left of it once every axis correlated with duration is removed.** Both

numbers must be quoted together, and the second one is now worse than V2's honest headline rather

than merely as bad. At N = 28 the length-independent composite was exactly chance; at N = 30 it sits

below chance with nine of twenty-eight rows above their album median. It is not significantly

inverted — p = 0.087 — but the direction of travel across two re-runs is unmistakable, and the

axes those folds select (drop_depth_max_db, melody_span_semitones, dyn_arc_range_db,

hook_range_semitones) are exactly the ones that will not hold. **The composite works, the

composite is made of size, and stripping size does not leave a weaker predictor — it leaves

something that is beginning to point the wrong way.**

8. The rejection envelope, re-derived — and the box moved

Separation ("which side of a threshold") and containment ("inside which box") are different

questions, and the second is the one a generator can use even when the first is weak. Each envelope

is built from twenty-nine hits and validated by asking whether it contains the thirtieth.

The box the generator is using is superseded

The joint box changed two of its three axes.

Only X_rest_grammar survives from V2's box. melody_span_semitones fell from the top of the

envelope ranking to 17th, and section_mean_s from 2nd to 9th. As a filter the box got

worse in both directions at once: it now admits 54.9% of album filler instead of 44.8%, for

slightly *lower* recall, and its enrichment fell from 1.99× to 1.58×.

Why: one 21.5-second sting, and the arithmetic is exact

A leave-one-hit-out sensitivity attributes the change with no ambiguity. This is a robustness read

with no null of its own; the N = 30 bar of 0.184 is quoted only as a reference line.

The last row is a positive control and it lands: dropping both new positives reproduces V2's numbers

to four decimal places — section_count 0.7501, total_length_s 0.7323, repeat_sim_mean 0.7138,

n_drops 0.7099, bars 0.7177 — which confirms the derivation is stable and that the pool is the

only thing that changed. The second row is the finding: **remove EX_014 alone and V2's six axes

and V2's exact box both come back.**

EX_014 measures melody_span_semitones 7.0 against a previous hit minimum of 19.67, and

section_mean_s 5.375 against a previous minimum of 7.145. Widening those two lower bounds pushed

melody_span_semitones's filler admission from 69.5% to 92.5% and section_mean_s's from 70.0% to

85.2%, which collapsed both lifts and demoted both out of the top three. A single 21.5-second attract

sting, correctly attached, is the entire cause.

How many round-1 verdicts flip

build/audio/generated/slice_v1/verdicts.json names V2's box as its _envelope_source, so every

verdict in the round-1 slice was taken under a filter this document supersedes. All 32 candidate

cards were re-judged under both boxes with the same rule the filter uses — reject if any box axis is

outside or unmeasurable.

• Re-judging under the V2 box reproduces the shipped verdicts exactly, 18 KEEP and 14 REJECT.

That is the positive control for this count.

• Under the V3 box the split becomes 16 KEEP and 16 REJECT. Two verdicts flip, both

KEEP → REJECT, both rejected on drop_depth_max_db, whose deepest drop must reach −10.73 dB and

which measures −9.45 and −7.14 on the two candidates.

• Both flips are SLICE_T1_EXPLORATION_FLORES_a_41022 and SLICE_T1_EXPLORATION_FLORES_a_41033,

and this is the part that matters: **they are ranks 1 and 2 of that theme and both carry

kept_for_review = true.** The EXPLORATION theme's three-candidate review shortlist loses its top

two picks and drops to two candidates. No other theme is affected.

Two flips of thirty-two is a small number. Two flips out of the twelve candidates actually shortlisted

for review, both at the top of their theme, is not. **The round-1 EXPLORATION shortlist should be

re-cut against the V3 box before anyone listens to it.**

X_rest_grammar earning a place in the joint box while failing badly as a separator remains the

separation-versus-containment distinction, not a contradiction: corpus rows occupy a narrow band of

composed rest without sitting high or low on it. The box remains a **necessary-not-sufficient

constraint**, and at 54.9% filler admission it is now a weak one.

9. The era mix — the size axes hold, and the modern half is where the damage is

Four sixteen-bit-era scores carry corpus rows — Chrono Trigger, FINAL FANTASY VI, Tales of Phantasia

and Streets of Rage 2 — thirteen of the thirty hits and 228 of the 927 siblings. Both new positives

are modern-era, so the split is now 13 sixteen-bit against 17 modern.

Both V2 readings reproduce.

• The clearing axes still do not shift with the era mix. Every one holds its sign and rough

magnitude in both halves. A structural-size effect that reproduces independently in 1994 Super

Famicom sample-ROM scores and in 2016 live-recorded soundtracks is not an artifact of production

era, budget or codec. This remains the strongest single result in the program, and V2's numbers

for the sixteen-bit half are essentially untouched (section_count 0.762 both times) because

neither new positive is sixteen-bit.

• The carried hypotheses still fail because of the era mix. All three sit above 0.5 in the modern

half and below 0.5 in the sixteen-bit half, and cancel to chance pooled. Unchanged from V2.

The leave-one-out composite folds split by era give a mean held-out percentile of 0.724 for the

sixteen-bit rows against 0.692 for the modern rows on the all-axes composite, against V2's 0.726

and 0.763. **The sixteen-bit half is unchanged to three decimals and the entire drop in the headline

composite is in the modern half**, which is where both new positives landed. On the

length-independent composite it is 0.479 sixteen-bit against 0.380 modern, both down from V2's 0.599

and 0.395, and both now below chance.

The lineage is still only half covered. Zero N64, zero Zelda, zero Kondo, zero Nintendo

first-party of any kind. It is the same gap V1 named, V2 named, and this pass names again.

10. What the numbers say at descriptive tier

Not discriminative findings unless marked as clearing the bar in §4. These are the measured profile

of the thirty, usable as sanity bands for authored candidates.

Four readings worth carrying into rung 2.

• **V1's most important sentence survives a third time, and it is the most stable result in the

program.** Every axis the folk theory of hummability is built from — stepwise fraction, gap-fill

after a leap, hook compass, contour turn rate, interval 3-gram repetition, time to hook, melodic

salience — sits between AUC 0.46 and 0.59 at eleven positives, at twenty-eight and now at thirty.

Not one has moved meaningfully in three passes. The most economical explanation is unchanged:

the album siblings already have them. These are well-written game scores throughout, and the

corpus row is not the only competent track on its record. What the card can see that separates

them is size and return, not melodic craft.

• A corpus row is still structurally larger and more self-returning than its own album, by

roughly the same margins — nearly double the section count, double the length, double the bars,

nearly triple the drop count. Part of it is curation, and about 0.59–0.62 of it survives against

equally long siblings.

• The tempo effect is now gone. AUC 0.344 at N = 11, 0.441 at N = 28, 0.416 at N = 30. It never

cleared anything and it should not be carried.

• **max_lanes is 9 for every one of the thirty hits, at eleven positives, at twenty-eight and at

thirty**, and takes only the values 8 or 9 across all 927 controls — which is why it fails the

bank's own distinctness admission and is not one of the 103. **Three consecutive derivations have

now confirmed this is a measurement ceiling in the spectral_band_proxy method, not a finding.**

Any arrangement-width claim from these cards is void until a real instrument roster exists. The

cards' own declaration that the lane method is not a roster is overdue and should be written.

11. What more corpus would still sharpen — the power model, re-fitted

K_eff, the number of effectively independent axes this bank behaves like, is chosen so the

simulated bar at N = 30 reproduces the observed permutation 95th percentile, and only then

extrapolated. Real sibling counts are used.

Rows needed for 80% power: 16 for an effect of 0.80, 30 for 0.75, 40 for 0.70, 75 for 0.65. The

pool sits exactly at the 0.75 threshold, which is why the clearing set is unstable — four axes above

and two just below is what a study sitting on its own detection boundary looks like.

One caution the model earns on itself: **the fitted K_eff moved from 103 at N = 28 to 50 at

N = 30**, with a better calibration error (0.0013 against 0.007). That is a large swing from two

positives, and it means the extrapolated columns are softer than their decimals suggest. Read the

whole table as lower bounds, as V2 did, and read the fitted constant as unstable.

What the next acquisitions are worth, in priority order.

• The Nintendo lineage is still entirely absent and it is still the highest-value buy. Zero N64,

zero Zelda, zero Kondo, zero Nintendo first-party. §9 shows the era split is measurable and that

it already overturned three hypotheses. A third era is the strongest available test of whether the

surviving axes are era-invariant or merely invariant across the two eras now held.

• The clearing set is one or two acquisitions wide. Six axes sit inside a 0.04-wide band around

the bar. The next two positives will move bars and n_drops back across it or push

repeat_sim_mean and novelty_mean out. Nothing built on the four-versus-six distinction should

be treated as settled; build on the ordering, which is stable across all three passes.

• Five of seven purpose classes are still below five rows — BOSS 2, TENSION 3, CREDITS_TRIUMPH 4,

MELANCHOLIC 4, TITLE_CHARACTER 4. No per-class band can be stated until each carries five-plus

rows. BOSS is the binding constraint.

• Two hits still have no siblings at all. EX_082 and EX_092 sit on a four-track vocal

compilation and are excluded from every within-album statistic here. The Metal Gear Solid score

albums would recover two positives already paid for.

• A leave-one-album-out pass is now overdue. Chrono Trigger and FINAL FANTASY VI contribute 11

of the 30 hits between them; two Pokémon compilations carry 353 of the 927 controls. §8's

leave-one-hit-out sensitivity shows how much a single row can move this pool, and albums are

heavier than rows.

12. Honest register — what this is and is not

• N = 30 positives, 28 usable for within-album statistics. Every number above carries that. This

is a measured-sample derivation, not a validated theory, and nothing here is yet a

MASTERPIECE_STANDARD band.

• Four axes clear a calibrated bar, down from six, and none clears the 99th percentile. The

correct reading is not that two findings were refuted but that six axes are smeared across a

threshold in a 0.04-wide band, with the bar inside it. repeat_sim_mean clears by 0.0009 and

n_drops misses by 0.0078. Treat the ordering as the finding and the membership as provisional.

• **The joint rejection envelope changed two of its three axes, and the round-1 slice was filtered

with the old one.** Two of thirty-two verdicts flip, both KEEP → REJECT, and both were shortlisted

top-two picks in the EXPLORATION theme. This is the most consequential change in the document.

• One row explains almost all of it. Removing EX_014 — a correctly attached 21.5-second attract

sting — restores V2's six clearing axes and V2's exact joint box. Removing both new positives

reproduces V2 to four decimals. That is a clean attribution and also a warning: at N = 30 a single

atypical positive can rewrite the generator's filter.

• Strip duration and the composite is now below chance. 0.707 with all axes, 0.426 with the 70

length-independent ones, against V2's 0.744 and 0.497. The single most quotable number in this

document is only honest when quoted with the second one, and the second one got worse.

• One held-out rule is now worse than nothing. drop_depth_max_db ≤ −59.67 carries a held-out

lift of 0.55×. Zero of six reverse by AUC sign; one of six reverses by the lift definition V2's

own "three of six at N = 11" used.

• section_count ≥ 15 survives and should not be refitted to 14. Its held-out recall fell from

0.62 to 0.50 and its axis AUC from 0.783 to 0.656, but as written it still carries a 3.15×

held-out lift, better than the refitted threshold on every test-side metric.

• The two new axes fail cleanly. structural_yield scores 0.462 within-album against a required

0.184 deviation, boundaries_per_min 0.467, and the ungated ratio 0.495. Neither would move the

bar if adopted. The 0.911 that motivated the proposal was measured against generated candidates,

which is a different question from the one this pool answers.

• The 30-year property is still not measured. It is not in these cards and cannot be. What is

measured is whether the tracks that have it share anything the cards can see; the answer is yes,

on structural size, and on nothing else the card can see.

• The pool is still skewed, and slightly more so. Eighteen albums, two Pokémon compilations

carrying 353 of 927 controls, two records carrying 11 of 30 positives.

• **The §5 truncation table, the §8 sensitivity and the §9 era split have no calibrated nulls of

their own.** They are robustness reads on a finding whose bar was set on the full pool, and the

0.184 line is quoted in them as a reference only.

• Per-class blocks are descriptive reads, never tests. Five of seven classes are below five rows.

13. Reproduction

• python harness/music_gen/derive_patterns.py --self-test — 26 controls, all green. The count

has grown from V2's 22: C12b asserts no quarantined id is an attached corpus row, C18b proves

no hit leaked into the controls by BYTES rather than by name, C18c proves that dedup armed by

construction with a synthetic twin and a one-byte-mutated sibling as its mutation arm, and C18d

proves the two historical twins left the live pool and were retired rather than deleted. C18c

is worth noting as a pattern: it previously censused the disk for real twins and went red when

acquire_exemplar.py retired them, which is a control expiring because the defect it proved was

fixed. It now makes its claim by construction and cannot expire.

• python harness/music_gen/derive_patterns.py --perm 400 --json build/audio/exemplars/PATTERN_FINDINGS_V3.json

— the full derivation, deterministic under seed 20260807, the same draw count and seed discipline

V2 used.

• The JSON carries every table above plus the full 103-row ranking, per-fold detail for both held-out

passes, the complete power curve and the per-class descriptive block. Its auxiliary key carries

the §5 truncation table, the §6c new-axis pricing, the §8 sensitivity and slice re-judgement, the

§9 era split, the reversal census under both definitions, and the lane-ceiling re-verification.

Everything under auxiliary is derived from the same public functions — load_cards → flatten

→ label_rows → within_album_auc / permutation_null_best / matched_ctrl / rule_scores —

and is labelled there as a robustness read where it has no null of its own. Everything outside

auxiliary is derive()'s own unmodified output.

• The two new axes are computed as len(duration_and_structure.novelty_peak_strength) divided by

len(event_grammar.drops) for structural_yield, gated to cards of 45 s or more, and by minutes

of runtime for boundaries_per_min. Card drops lists and counts.drops agree on all 1180 cards,

which was checked rather than assumed. Folding both into derive_patterns.flatten is a one-line

change and is left to whoever next edits that file; they are priced here and they fail, so nothing

downstream depends on it.

14. What rung 2 should do with this

• Re-cut the round-1 slice against the V3 box before anyone listens to it. Two of thirty-two

verdicts flip and both were shortlisted top-two EXPLORATION picks. Re-point

slice_exemplars.py's FINDINGS constant at PATTERN_FINDINGS_V3.json, re-run filter and

sheet, and let _envelope_source name V3. Until that is done the review sheet is quoting a box

that no longer exists.

• Build the archetype synthesis on the ordering, not on the membership. section_count,

total_length_s, repeat_sim_mean and novelty_mean clear at N = 30; n_drops and bars sit

within 0.009 of the bar and cleared at N = 28. All six are the same story and their rank order is

stable across three passes. A band on any of them is defensible at the honest tier "clears or sits

at a calibrated multiplicity bar at N = 30, era-invariant across the two eras measured, partially

confounded with curation." A band that depends on the four-versus-six distinction is not.

• Keep section_count ≥ 15 as written. It is still the program's best single statement and it

beats the threshold Youden's J now prefers on every held-out metric.

• Do not build a band on melodic craft from these cards. Every hummability axis is flat at 30 as

at 28 as at 11. If the difference lives there, this card cannot see it, and the correct action is

a better feature — a real instrument roster, a real hook extractor — not a band.

• Retire the three carried hypotheses permanently. recurrence_lag_norm is now the worst axis of

103. They have been tested at the N their own power model asked for.

• Do not adopt structural_yield or boundaries_per_min as corpus axes. They separate the

generator from real music, which is a useful thing and a different thing. If a generator-detection

filter is wanted, build it as one and name it as one; do not let it enter a corpus derivation

where it measures 0.462.

• Rung 3's predictor floor moves down, and its caveat moves up. 0.707 ± 0.053 held-out, from a

composite of size axes, with 0.426 length-independent beside it. Any predictor built on this must

carry the second number, which is now below chance rather than at it.

• Buy the Nintendo lineage next, and expect the clearing set to move again when it lands. §8 has

just demonstrated that one row can rewrite the filter; an era can do more.

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