VARIETY_MATRIX.md

music/complete_tracks/VARIETY_MATRIX.md

THE VARIETY MATRIX — the tooth that refuses two tracks that sound the same

CANON SUBORDINATION — this document is PROPOSAL-TIER: it serves canon and never outranks it.

The canon: the CVD · the T1 foundation docs · the T0 registries (registries/) · the spine

(docs/spine/CH_*.md) · the region pages (_source/02_Tier_2_Region_Pages/). Authority order: docs/DOC_MAP.md § 0.

Canon served: T0_Theme_Registry · T0_Motif_Registry · T0_Chapter_Index · T0_Festival_Registry.

If this document disagrees with canon, CANON WINS and this document is the defect.

Emitted by harness/music_gen/emit_identity_cards.py. The cards themselves are

build/audio/complete_tracks/cards/HF_MT_*.json. The control arm is

build/audio/complete_tracks/SAMENESS_BASELINE.md.

0. The rule

No two cards may share tempo_band + instrumentation_family + energy_class.

Those three axes were chosen because they are what the sameness baseline measured as collapsed.

The shipped cohort carries 2 texture classes across 15 tracks, 2 length classes, and exactly one

value of distinct_composed_sections. Pitch-set distinctness — the axis the lane had been

optimising — is not on this list, because it is not audible in the first four seconds.

The exemplar model is the Super Mario 64 track list: a roster of roughly this size where every

entry separates from every other by SOUND before any melody is recognised.

1. The matrix

2. The proof

• 20 cards, 20 distinct triples, 0 collisions.
• Of 190 card pairs: 156 differ on all three axes, 34 on two, 0 on one. Zero differ on none — that is the tooth.
• Tempo bands used: 7 of 7 — T1_very_slow, T2_slow, T3_walking, T4_moderate, T5_brisk, T6_fast, T7_very_fast
• Instrumentation families used: 12 of 12 — choir_and_low_strings, full_orchestra_and_choir, full_orchestra_tutti, low_brass_and_organ_weight, percussion_forward, plucked_and_struck_period_ensemble, solo_instrument_led, strings_and_winds_chamber, voice_and_soundscape_minimal, voice_led_chant_ensemble, voices_and_body_percussion, wind_and_drum_period_band
• Energy classes used: 19 of 19

• Duration: every card 158–295 s (mean 258.9 s), all inside the 2.5–5 minute band. The shipped cohort's mean is 29.3 s and 15 of 15 are under a minute.
• Form: 5–12 named sections per card (mean 7.4), 56–176 bars. The shipped cohort measures 2 distinct composed sections on 15 of 15 tracks, one of which is a one-bar turnaround.
• Anti-tiling: every card declares ≥ 4 mutually-unlike motifs (20 of 20), which arms L2's target as a FLOOR — the RECOMPOSE_RUNG_TWO_FINDINGS § 1.3 finding, where the target was never enforced and the recomposition scored worse than the thing it replaced on the rubric's own headline mechanism.
• Loop-point cadence: weakened on every card — a half cadence or a bVII substitution, never a perfect authentic cadence (KONDO PHRASE-05, the anti-fatigue law).
• Reveal firewall: 13 held tokens scanned across 20 cards at every depth, 0 hits.
• Every tooth is mutation-proved, not asserted: --selftest forces each one to fire on a deliberate defect and to pass on the clean set.

3. The duration law, and the apparent conflict it resolves

SNES LOOP-01 prescribes loop lengths by cue class — hook-bearing theme loops 30–60 s, ambient

and exploration 1:30–2:30, combat 15–30 s. Josh asked for complete tracks of 2.5–5 minutes.

Those are not the same measurement and the cards carry both:

• target_duration_s is the composed track — the object with an intro, real sections, a

development and a coda. This is what a listener hears as a piece of music, and what the

exemplar soundtracks ship.

• in_game_loop_span_s is the span the runtime loops, chosen per cue class to LOOP-01.

The two are related by composition, never by tiling: the loop span is a designated region of the

composed track with a composed seam, and no section of any card is a verbatim repeat of another.

That is the NO TILING law from RECOMPOSE_RUNG_TWO_FINDINGS stated as a schema rather than

as an instruction — and it closes § 1.1's defect directly, because a card can now say *a loop

composed of N sections, each with its own bar count*, which is the thing the old schema could

not express.

4. What is HELD, and why — so a gap reads as a decision

HF_MT_XX_REGION_RIFT_VALLEY — East African Rift Valley

• Held because: The region page's Section 10 opens: documented instruments -- none attested in this corpus, and the section says so before it says anything else. No musicological source is in the corpus, so no rhythmic structure, scale, call pattern or instrument is described anywhere on the page. An identity card's central field is INSTRUMENTATION FAMILY; writing one here would invent the musicology the page deliberately refuses to invent.
• Not silent: CH_10 is EXPLORATION_derived with emotional_register = awe and environmental combat. It is scored in wave one by HF_MT_02_EXPLORATION_TRAVERSAL and HF_MT_09_WONDER_DISCOVERY, both of which serve it explicitly.
• Entry condition: A musicological source acquisition or a deep-research pass that attests an instrumentarium. Until then the chapter's ratified content (a healing-dance clapping and a women's song-line) is carried at reference register by the function cues, which is exactly what the page authorises.
• Source: _source/02_Tier_2_Region_Pages/rift_valley.md section 10

HF_MT_XX_FAIRY_REALM — Fairy Realm (Prologue, CH_01)

• Held because: Josh's slice ruling builds the Prologue and Ch 1 LAST, with the fairy realm designed once at slice-end on the matured concept-art pipeline. A music identity card written now would fix an instrumentation family for a realm whose visual and cultural design has not been made.
• Not silent: Nothing in the Prologue or Ch 1 is scheduled for scoring in this wave, so there is no gap to cover.
• Entry condition: The fairy-realm design pass at slice-end. The card is written from that design, not before it.
• Source: CLAUDE.md :: THE JOSH GATE (slice redefined 2026-07-27)

5. The stack gap, stated before anyone renders

The landed CPU stack is sfizz 1.2.3 over VSCO2-CE and VCSL — 23 instruments, all CC0. It

contains no percussion of any kind, no choir, no trumpet, no tuba, and no period instrument.

So the shipped cohort's single texture is not only a compositional choice; it is partly what the

stack could physically play. Every card declares its own stack_coverage, and the cards that

name a sound the stack cannot make are naming an acquisition, not claiming a render.

4 of 20 cards are renderable on the landed stack today. The rest are the

wave-two sourcing worklist: CC0 or public-domain percussion, choir and period sample sets, with

the same hash-and-licence discipline as the landed stack, and the standing rule intact — no

sample from any set is ever a conditioning, fine-tune or LoRA input to any generator.

6. How to check this rather than trust it

python harness/music_gen/sameness_baseline.py      # re-measure the control arm
python harness/music_gen/emit_identity_cards.py --selftest   # arm every tooth
python harness/music_gen/emit_identity_cards.py    # re-emit; refuses on any collision

Re-run the baseline against the next rung's renders. Every axis in its § 3 clustering table must

move, and the cards above are the specification of how far.

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