SAMENESS_BASELINE.md

music/complete_tracks/SAMENESS_BASELINE.md

THE SAMENESS BASELINE — what the shipped orchestral cohort actually measures

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.

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

Emitted by harness/music_gen/sameness_baseline.py. Every number here is measured off the

artifact on disk or read off its own composition document. Nothing is estimated.

0. The ruling this answers

Josh, 2026-08-05: *"Why is all this music the fuxking same, all just 30 seconds each the exact

same track, whys it all just one type of sound on a 5 second loop, like wtf? Were supposed to be

creating at least 12 to 20 complete tracks with actual full orchestra complexity."*

The lane's previous defence was that the themes are DISTINCT — different pitch sets, different

keys, different head cells, each one measured and gated. That defence is true and it is

irrelevant, because pitch-set distinctness is not an axis a listener hears in the first four

seconds. The axes a listener hears first are duration, tempo, instrument set, texture and

dynamic shape. This table measures those. The verdict below is arithmetic, not opinion.

1. The cohort

15 orchestral rung-one renders on disk. The nostalgia census scored fourteen of them;

JOURNEY_WORLD was held out of that census as the recompose control and is included here

because a listener hears it like any other. HOUSE_PARENT_HELD is reveal-firewalled (R1,

served to nobody before Ch 38) and is measured but never shipped — it is marked HELD below.

2. The sameness table

3. The clustering — distinct values per axis

An axis whose distinct-value count is 1 is an axis on which every track in the cohort is

literally identical. An axis whose modal share is above 80% is an axis a listener will hear as

identical, because the exceptions are not enough to register as variety.

4. The scalars

• Duration: 21.6 s to 42.5 s, mean 29.3 s, total spread 20.9 s. 15 of 15 are under a minute. 0 of 15 reach the 2.5-minute floor a complete track needs.
• Tempo: 68 to 120 BPM, mean 93.7.
• Dynamic range (short-time RMS p95 − p10): 3.4 to 20.9 dB, mean 9.1 dB. A scored orchestral cue moves through 20 dB or more between its quietest sustained bar and its tutti; a cohort clustered in a narrow band has no dynamic shape at all — the machine-flat reading Josh heard.
• Crest factor: 10.4 to 27.9 dB, mean 14.2 dB.

5. The verdict, stated as the mechanism

The pinned mechanism reads, in the cohort's own numbers:

• One instrument set. 4 distinct instrument signatures across 15 tracks; the modal set covers 12/15. Every render is a solo carrier over the same two string-ensemble frame parts and the same contrabass.
• One texture. 2 texture classes; modal share 14/15. No counterpoint, no tutti, no ostinato-driven bed, no percussion of any kind anywhere in the cohort.
• One length class. 2 length classes; modal share 8/15. Josh said thirty seconds. The measured mean is 29.3 s.
• One form. 2 distinct declared forms across 15 tracks — and both of them are the same sentence, stated once or stated twice. No track in the cohort has a B section. distinct_composed_sections is 2 for every track, and one of those two is a one-bar turnaround. That is what "a 5 second loop" is measuring: an 8-bar sentence whose second half restates its first, heard twice.
• Machine-flat dynamics. The frame parts are sus_quiet on every track, the mix constants are shared, and the measured short-time spread confirms it. Nothing crescendos, nothing drops out, nothing arrives.

What WAS distinct: key (see the tally), head-cell pitch content, contour class and the

deviant. Those are the axes the lane optimised, and they are real. They are also, every one of

them, invisible to a listener who is hearing the same clarinet-over-quiet-strings sound at the

same dynamic for the same twenty-five seconds. Josh's ear is confirmed by the numbers.

6. What this baseline is FOR

This is the control arm. The identity cards in build/audio/complete_tracks/cards/ are the

treatment: 16-20 cue cards, each carrying a target duration in the 2.5-5 minute band, a

tempo band, an instrumentation family, an energy class and a real multi-section form — and a

variety matrix that REFUSES any two cards sharing tempo band + instrumentation family +

energy class. Re-run this script against the next rung's renders; every axis above must move.

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